LocalAI/docs/content/features/addons.md

title

title
Addon contracts

Status and trust

The version-one contracts in pkg/addons/contract define managed-process addons. The owner installs, verifies, starts and supervises them, and a deployment with several frontends forwards every addon request to that owner over a pinned mutual TLS carrier. The routes are described in Public API, the Apps and Add-ons pages in Apps, and Tables, the first first-party addon, in Tables. The first-party package is described in First-party package and release channels; it is not published to a registry by any test or by acceptance.

Trust has three parts, and they are different:

• Package identity. A package is installed only after its digest matches the catalog entry and, when the library requires it, its signature matches the identity policy of that library. A library, a channel and a trusted signer never change by themselves, and an upgrade is refused when the trusted identity of the new package differs from the installed one.
• Executable code. An installed executable is trusted code, not sandboxed code. It runs as the owner's operating system user, with the owner's file system and network. The host gives it a loopback port, a per-start credential and its own data directory; it does not confine it. Install only packages you trust.
• Authority. An addon never decides who a caller is or whether a change was approved. The host supplies the actor from the authenticated session or key, decides every role, and applies a mutation from an API client, an MCP session or the OpenAI-compatible chat route only after the proposer's own browser session confirms it. The one exception is the chats on the server (/api/chats), their agents and their triggers: turning an add-on on for a chat or an agent is the consent, and the change is applied at once, with the credential of the chat turn, for the user who turned the add-on on, and with an undo where the add-on keeps one (see Proposals and confirmation). Prompt text, a model and an addon cannot supply that approval.

Every addon requires an executable; a UI bundle alone is not an addon.

Singleton owner and its limits

Addon state lives in one place. Exactly one LocalAI instance is the owner of an addon root (its storage identifier and Root directory): it holds the journal of installed packages, the resource policy store, the proposals, the artifact store, the backups and the data of every addon, and it supervises the addon processes. The other instances are frontends: they authenticate the caller against the shared account store and forward the typed operation to the owner. The consequences are by design:

• No second owner. A second process that opens the same root is refused while the first holds it; two owners are never active. There is no election and no automatic failover. Moving the owner to another machine is an operator action: stop the owner, move or mount the root, configure the new owner and start it.
• An owner outage is an addon outage. While the owner is down, every frontend answers 503 for every response type (JSON, progress, status, file, asset and proposal calls) with the same typed error, and executes nothing locally. Nothing is queued or applied later. A mutation whose reply was lost is reported as outcome_unknown with its original operation ID and is resolved by receipt lookup after the owner is back, never by sending it again under a new ID.
• No state on frontends. A frontend keeps no addon data, no addon process and no sticky session. Any frontend can serve any request, and a request may use another frontend than the previous one.
• One writer per database. A Tables database is one SQLite file on the owner's disk with a single writer. Throughput is bounded by that machine, and the data is only as available as that disk and its backups.
• The addon process is local to the owner. It listens on loopback only. The carrier between a frontend and the owner is the only network path, and it is authenticated in both directions by pinned certificates and signed requests.

A deployment with one instance is the owner and has no frontend; everything above still holds.

The --addons-role flag chooses how a process takes part. combined is the default and is both the owner and the only frontend. owner and frontend split them, as described in Running the owner as its own workload.

Distributed state. In [distributed mode]({{% relref "features/distributed-mode" %}}), the owner is a single runner by design: the operator runs one owner replica, and the lock file in its root is the guard against a mistake, not a way to scale. The frontends are stateless for add-ons. The one exception is the add-on pins of a frontend that has no account database, which are per-instance state (see Frontend limits).

Turning the owner on or off

A normal local-ai run starts the addon owner, so Apps, Add-ons and Tables work without extra setup. The node is its own owner: it keeps addon state under the addons folder of the data path (--data-path) and reaches the owner over a loopback carrier with keys that are made at each start and never written to disk.

Flag	Environment variable	Default	What it does
--addons-enabled	LOCALAI_ADDONS_ENABLED	true	Runs the addon owner. --addons-enabled=false turns it off completely: no owner starts, no addon runs, the UI hides the add-on entries, and every add-on, fine-tuning and quantization route answers 503. See The addons feature gate.
--addons-ownership-mode	LOCALAI_ADDONS_OWNERSHIP_MODE	local-owner when authentication is off, named-user when it is on	named-user gives each account its own access. local-owner makes every caller share one owner. A value you set always wins over the default. Any other value stops startup with an error.
--addons-official-library	LOCALAI_ADDONS_OFFICIAL_LIBRARY	true	Registers the official library on the first start. --addons-official-library=false does not register it.
--addons-backend-library	LOCALAI_ADDONS_BACKEND_LIBRARY	true	Registers the backend library on the first start, like the official library. A build from master reads it as the backend source; a tagged build reads the backend gallery until the release cutover. --addons-backend-library=false starts without it and reads the backend gallery on every build. A library you remove stays removed.
--official-libraries	LOCALAI_OFFICIAL_LIBRARIES	true	Use the default library sources of add-ons, backends and models. --official-libraries=false leaves them listed but unread. It is also the runtime setting official_libraries. See Library sources.
--addons-backend-max-bytes	LOCALAI_ADDONS_BACKEND_MAX_BYTES	8589934592 (8 GiB)	Largest compressed size of a backend image installed from a library. Add-ons keep their own limit of 512 MiB.
--addons-role	LOCALAI_ADDONS_ROLE	combined	combined is the single node described above. owner runs only the owner and serves the carrier on the network. frontend runs no owner and forwards every add-on request to it. Any other value stops startup with an error.
--addons-owner-address	LOCALAI_ADDONS_OWNER_ADDRESS	empty	https://host:port of the owner carrier, with no path. Use a stable service name, never a pod address. Required for owner and frontend; an error with combined.
--addons-owner-listen	LOCALAI_ADDONS_OWNER_LISTEN	empty	owner only. The address the carrier listens on. Empty listens on all interfaces, on the port of --addons-owner-address.
--addons-carrier-dir	LOCALAI_ADDONS_CARRIER_DIR	empty	Directory with the carrier key set made by local-ai util addons-carrier-keys. Required for owner and frontend; an error with combined.

named-user needs authentication. With authentication on and no mode set, LocalAI uses named-user. With authentication off there are no accounts, so LocalAI uses local-owner by default: every caller that can reach the server shares one owner and can install, run and remove add-ons. See Explicit local-owner mode for what that means. To keep add-ons off on a server without authentication, set --addons-enabled=false, or set --addons-ownership-mode=named-user; with authentication off, named-user leaves the owner off and LocalAI logs a warning.

The owner and frontend roles split the owner from the frontends. The frontend role does not create or open the addons folder of the data path. The key set for both roles comes from local-ai util addons-carrier-keys; see Running the owner as its own workload.

The addons feature gate

Fine-tuning and quantization are add-ons, so --addons-enabled=false turns off all three. GET /api/features reports one key, addons, and the keys fine_tuning and quantization carry the same value. The key is true when add-ons are available on this instance:

• In the combined and owner roles it is true when add-ons are enabled.
• In the frontend role it is true when add-ons are enabled and the owner answered a reachability call within 2 seconds. The answer is kept for 10 seconds, so /api/features does not contact the owner on every request.

While the key is false:

• The UI shows no Add-ons entry, no pinned apps, and no fine-tuning or quantization entry. /app/addons, /app/fine-tune and /app/quantize lead back to /app. The UI does not call /api/addons/pins or /api/addons/builtin. A server that sends no addons key counts as true in the UI.
• Every path under /api/addons, /api/fine-tuning and /api/quantization answers 503 with the same body, also for a path that does not exist. The answer does not show which routes exist. The body is the LocalAI error envelope ({"error": {"message", "type", "code"}}, code unavailable). The message is "add-ons are turned off on this instance" when the feature is off and "the add-on owner is not reachable" when a frontend cannot reach the owner.
• With --addons-enabled=false the handlers of these routes are not registered at all. A frontend with --addons-enabled=false needs no --addons-owner-address and no --addons-carrier-dir. It loads no keys and builds no carrier. If you set --addons-role, --addons-owner-address, --addons-owner-listen or --addons-carrier-dir, LocalAI ignores them and logs one line.

This gate is for the whole instance. It does not replace the permission of each user: the addons, fine-tuning and quantization permissions of a user are checked after the gate.

The reachability answer of a frontend is a per-instance observation and is not shared. Two frontends can disagree for up to 10 seconds after the owner goes up or down. A request through a frontend that still reports true gets the 503 of the carrier while the owner is down.

Running the owner as its own workload

A normal local-ai run is both the owner and the only frontend (the combined role). In a deployment with several frontend replicas, run the owner as one separate workload and let the frontends forward to it. The owner keeps all its state in files under one root, as before. Frontends keep none and scale freely.

• One owner, one volume. Run exactly one owner replica (--addons-role=owner) with its own volume for the data path. Use a recreate-style rollout, so two owner processes never overlap: the old one must stop before the new one starts. A second owner on the same root stops at once with "addon storage already owned" and the process exits with an error. It does not wait.
• Fixed mount path. The journal stores absolute paths. Mount the volume at the same path on every start. If the data path or the mount point changes, every installed add-on points at a directory that does not exist.
• A local volume. The root needs a local file system with working flock, atomic rename and directory fsync. A network block volume that one node mounts at a time is fine. A shared network file system is not supported.
• A stable owner address. Frontends reach the owner at --addons-owner-address. The owner does not store it, so you can change it between starts. Use a stable service name, never a pod address.
• Frontends have no add-on volume. A frontend (--addons-role=frontend) does not create or open the addons folder. Give it no volume for it.
• Existing add-ons. A root written by a single node loads under the owner role when the owner uses the same owner and storage identity. The stored address does not matter. The journal keeps absolute paths, so the owner must see the same data path and mount point as the node did. Otherwise the installed add-ons point at missing directories. For an empty volume, the owner starts fresh.

{{% notice warning %}} The owner is a node-bound singleton. If the owner is down, add-ons are down: the frontends stay up and answer 503 for every add-on request. Failover is manual. There is no election and no standby owner. {{% /notice %}}

Make the carrier keys

The frontends and the owner trust each other through a static key set. Make it once, anywhere you can run local-ai:

local-ai util addons-carrier-keys --out ./carrier-keys --owner-host owner.example

--owner-host is the name the frontends dial, and it must match the host in --addons-owner-address. The command prints which files go where:

Role	Mounts
Owner	ca.pem, owner.pem, owner.key, frontend.pem, frontend-sign.pub
Frontend	ca.pem, owner.pem, frontend.pem, frontend.key, frontend-sign.key

Store the set as a secret and mount only the files of each role. Point --addons-carrier-dir at the mount. LocalAI refuses key files that group or other users can read, so give them mode 0600 or tighter.

All frontend replicas share one identity and one signing key, so a new or rescheduled frontend is accepted at once and the owner needs no list of replicas. Anyone who holds the frontend files can act as a frontend. A forwarded request still carries the user, and the owner checks that user against the account database, so the key set does not give a user any authority on its own. The command does not keep the CA key, so nobody can make a third certificate from the set.

{{% notice note %}} Rotation is a coordinated restart. The frontends pin the exact certificate of the owner and the owner pins the one of the frontends. To change the key set, make a new one, update the secret, and restart the owner and every frontend. There is no overlap window and no automatic renewal. {{% /notice %}}

Example

This Docker Compose file shows the add-on flags of one owner and two frontends. Add the other flags of your deployment (the database, NATS and the rest of [distributed mode]({{% relref "features/distributed-mode" %}})) to each service as usual.

services:
  owner:
    image: localai/localai:latest
    command: run
    hostname: owner.example
    environment:
      LOCALAI_ADDONS_ROLE: owner
      LOCALAI_ADDONS_OWNER_ADDRESS: https://owner.example:8443
      LOCALAI_ADDONS_CARRIER_DIR: /run/addon-keys
      LOCALAI_ADDONS_OWNERSHIP_MODE: named-user
      LOCALAI_DATA_PATH: /data
      LOCALAI_AUTH: "true"
      LOCALAI_AUTH_DATABASE_URL: postgres://localai:secret@db.example:5432/localai
    volumes:
      - owner-data:/data
      - ./carrier-keys-owner:/run/addon-keys:ro
    deploy:
      replicas: 1
      update_config:
        order: stop-first

  frontend:
    image: localai/localai:latest
    command: run
    environment:
      LOCALAI_ADDONS_ROLE: frontend
      LOCALAI_ADDONS_OWNER_ADDRESS: https://owner.example:8443
      LOCALAI_ADDONS_CARRIER_DIR: /run/addon-keys
      LOCALAI_ADDONS_OWNERSHIP_MODE: named-user
      LOCALAI_AUTH: "true"
      LOCALAI_AUTH_DATABASE_URL: postgres://localai:secret@db.example:5432/localai
    volumes:
      - ./carrier-keys-frontend:/run/addon-keys:ro
    deploy:
      replicas: 2

volumes:
  owner-data:

./carrier-keys-owner holds the owner files and ./carrier-keys-frontend holds the frontend files from the table above.

Settings every role shares

• The same ownership mode. Every role of a deployment must use the same --addons-ownership-mode.
• The same account database. In named-user mode all roles must use the same account database. The owner checks the user of each forwarded request against it.
• --addons-enabled=false keeps its meaning in every role.

What a restart looks like

When the owner stops, the kernel drops its lock and its add-on processes end with it. The new owner takes the lock, checks its identity file, and reads the journal: operations that had only started are marked failed, and operations that were already prepared or applied are finished. Recovery journals and pending proposals finish or wait for an administrator. Then the add-ons that were enabled start again.

While this happens, the frontends answer 503. A frontend that starts while the owner is down starts normally and answers 503 until the owner is up. Some state is in memory only and is lost: an upgrade that was being canceled becomes failed, and the memory of failed automatic upgrades is cleared. A mutation whose reply was lost is reported as outcome_unknown with its operation ID, and relays that were in progress (progress, file and asset) are cut. The owner picks a new audience at each start. A request signed for the old owner is refused, and the frontend gets the new audience on its next call without any action.

The models path

Fine-tuning and quantization export models into the models path of the owner. The owner and every frontend must mount the same models storage at the same absolute --models-path. LocalAI checks this with a probe file:

• At start, the owner writes .localai-addons-probe.json into its models path. It holds a random nonce that changes at every start. Do not delete it, and do not exclude it from the shared volume. LocalAI does not list it as a model.
• Each frontend checks at start, and then every 60 seconds, that it sees the same file with the same nonce at the same absolute path. If it does not, the frontend logs an error, reports model_exports: false in /api/features, and answers the export and import routes with 503.
• The owner refuses an export or import while its own probe file is missing, also for the add-on host operations. If the owner cannot write the probe at start, it keeps running and refuses exports.
• If the owner is unreachable, or is an older version without the probe, the state is unknown. LocalAI logs a warning and leaves exports on. The state clears within 60 seconds after the owner is back.

The /api/fine-tuning and /api/quantization routes of a frontend still write to the models path of that frontend. The probe is what blocks them when that path is not the shared storage.

Frontend limits

A process with --addons-role=frontend holds no add-on state. This has these consequences:

• Library sources live on the owner and are changed through it. The frontend never opens, creates or writes sources.json or anything else under the addons folder of the data path. Every /api/library/sources route is forwarded to the owner, which is the only writer. The caller must be a named account and an administrator: the frontend checks the account, and the owner checks the administrator right again, as it does for its own routes. A legacy API key has no named user and gets 401 on a frontend, in local-owner mode too. The routes answer the same statuses as on a single node (404 for an unknown source, 409 for a duplicate, 400 for an invalid request). With add-ons off a frontend has no owner to reach, and these routes answer 503 with the message "library sources are managed on the owner". When the owner is down they answer 503 with "the add-on owner is not reachable". A change whose reply was lost answers 503 with the operation ID in the message; read the list again before you repeat it.
• The gallery lists follow the owner. A frontend asks the owner for its source list every 60 seconds (a read over the pinned carrier that needs no signed-in user) and, only when the list or the official_libraries setting changed, rebuilds its model and backend gallery lists from the galleries named by --galleries and --backend-galleries plus the owner's sources. A change made on the owner shows up on every frontend after at most 60 seconds, and a change made through a frontend shows up on that frontend at once. Until the owner has answered once, a frontend reads only its own flags. While the owner does not answer, the frontend keeps its last lists and logs that they may be stale. This is a per-instance observation: two frontends can show different lists for up to 60 seconds.
• official_libraries is set on the owner. The owner's value is the value of the deployment, and a frontend takes it at each refresh. POST /api/settings with a different official_libraries value answers 409 on a frontend. Set it on the owner.
• Backend installs run on the frontend node. POST /backends/library/install and GET /backends/library/updates work on a frontend. The owner resolves the entry (including the choice of a meta backend for the capability of the frontend) and pins the image by digest. The frontend then installs it into its own backends path, verifies the signature there, and applies its own size limit and --require-backend-integrity. The owner's backends path stays unchanged. The update check compares the local backends with the releases the owner lists. The background update check has no user, so it works in named-user mode through one read-only system call (POST /addons/library/system, capability library-system). The owner accepts it only from a pinned frontend, only for the release lookup, and without an actor. Every call a user starts still needs the administrator check on the owner.
• The owner and the frontends must run on the same OS and architecture. The owner resolves entries for its own platform, so a frontend checks the platform the owner reports (GET /addons/capabilities) and refuses every library call to an owner on another one with a 503. A frontend also refuses library calls to an owner that does not serve them, with "the add-on owner does not support library operations; upgrade the owner". Upgrade the owner before the frontends.
• Pins are per frontend without an account database. With an account database, the pins of the Apps entry are shared by all frontends. Without one, each frontend keeps its own pins in a file in its own data path, so two frontends can show different pins. A frontend logs a warning at start when it has no account database.

The official library

On the first start, when the owner has no sources.json in the addons folder yet, LocalAI registers one library with the ID official. It is a YAML index of the first-party add-ons (Tables, Voice, Face, Fine-tuning, Quantization, Studio, Talk, Simple Office, Sandbox and Web). The index has one entry for each enabled platform of each add-on, so an install on Linux (amd64 or arm64) and on macOS (Apple silicon or Intel) finds the add-ons for its own platform. The Sandbox add-on is Linux only. No add-on is built for Windows yet, so the official library lists nothing there.

Which index the library reads depends on the build:

Build	Index	Entries	Signer the library requires
A release (the version is a v<semver> tag)	addons-index.yaml, an asset of that GitHub release	Pinned by digest to the packages the release built	The package workflow run of that tag (refs/tags/v<semver>)
Any other build (master, a source build, a commit after a tag)	gallery/addons/index.yaml on the master branch	The mutable development builds	The package workflow run on master (refs/heads/master)

https://raw.githubusercontent.com/mudler/LocalAI/master/gallery/addons/index.yaml
https://github.com/mudler/LocalAI/releases/download/v<version>/addons-index.yaml

So a release installs the immutable, tag-pinned packages that the same release built and signed, and never a development build signed by master. The first start writes the index and signer of the running build into sources.json. After an upgrade, the next start moves the official library to the index and signer of the new build; a library that you pointed at another URL is left as it is.

The index has the same format as the other libraries (see Catalog document and limits). A package that the named workflow run did not build and sign is refused at install.

gallery/addons/index.yaml is generated, never edited by hand:

sh scripts/addons/catalog.sh --channel development \
  --repository quay.io/go-skynet/local-ai-addons --out gallery/addons/index.yaml

CI runs the same command with --check and fails when the committed file is stale, for example when a platform is enabled in addons/<id>/package/platforms.yaml or a manifest changes. The release index is rendered by the add-on package workflow when a v<semver> tag is pushed and attached to the release when it is published (see Preparing a release).

Add-ons that are not built in

Studio, Talk, Voice, Face, Fine-tuning, Quantization, Simple Office and Web are not part of the LocalAI binary or image. A fresh install has none of them. The old links to the pages they replaced (see Old links) open the add-on page, which says the add-on is not available until it is installed.

To get them on a new install:

1. Sign in as an administrator (with authentication off, you already are one) and open Add-ons, then Find add-ons. The list is the official library for this machine's operating system and architecture.
2. Pick the add-on and choose Install and turn on. The install checks the signature against the library policy before anything runs. Studio is the usual first one: it holds the image, video, 3D, speech, sound and audio tools.
3. The add-on then appears in My add-ons. Use Show in sidebar to pin it.

An install needs the network to reach the library and the package registry. A machine without it can run a private library (see Private catalog service) that mirrors the packages. --addons-official-library=false starts without the official library, and --addons-enabled=false turns the add-on owner off, so nothing can be installed. Installed add-ons live in the addons folder under the data path, so keep that path on a volume in a container, or the add-ons must be installed again after the container is replaced.

The first start writes sources.json, and from then on that file is the record of your choices. If you remove the official library (in the Add-ons page or through the source routes), it stays removed, and a sources.json you already have is never changed. To start without it, set --addons-official-library=false before the first start. To get it back, add a source with the URL above, or enable it again (see Library sources).

Upgrading: Studio and Talk are add-ons now

Studio (image, video, 3D, speech, sound and audio) and Talk (realtime voice) used to be pages of LocalAI. They are add-ons now, and two things change at the upgrade:

• Nothing is installed for you. No add-on is preinstalled. Studio and Talk appear only after an administrator installs them from the library (see Add-ons that are not built in). Until then the old links open a page that says the add-on is not available.
• Users need the addons feature. With authentication on, every add-on page is behind the addons feature, which is off by default for a user. A user who had the image, speech or realtime features but not addons loses Studio and Talk at the upgrade. The upgrade does not grant addons to anyone. An administrator always has it.

To restore access:

1. Install Studio and Talk: Add-ons, Find add-ons, then Install and turn on for each. A machine with no access to the network installs them from a private library that mirrors the packages (see Private catalog service and Library sources).
2. Grant the addons feature to each user who should keep them, in Settings → Users and keys (see Chats permissions for the same screen). The calls the pages make still need the user's own features: images for an image, audio_speech for speech, and so on.

With authentication off there are no user features: the one person is an administrator and sees every installed add-on.

Release notes. The paragraph below is ready to copy into the release notes:

Studio and Talk are add-ons. Studio (image, video, 3D, speech, sound and audio) and Talk (realtime voice) moved out of LocalAI into add-ons. Nothing is preinstalled: after the upgrade an administrator installs them from Add-ons → Find add-ons (an air-gapped host uses a private library that mirrors the packages). With authentication on, the add-on pages need the addons feature, which is off by default and is not granted at the upgrade, so users who used Studio or Talk lose access until an administrator grants addons to them in Settings → Users and keys.

Library sources

A library source is a place LocalAI reads entries from. There are three kinds:

Kind	What it provides	Read by
addon	Add-ons, as a library index	The Add-ons catalog
backend	Backends, as a library index or a backend gallery index	The Backends page, and the add-on catalog for a library
model	Models, as a library index or a model gallery index	The model gallery

All sources of all kinds are in one store, addons/sources.json under the data path. local-ai library (see the CLI reference), the HTTP routes below and the Sources tab work on the same store with the same code, so a change made in one place shows in the others.

A source can be a default: one that LocalAI ships (the official add-on library, the model library and gallery index, and the backend library and gallery index). Of the two model sources and the two backend sources only one is on by default, the one the build reads (see [Which source a build reads]({{% relref "features/backends#which-source-a-build-reads" %}})). A default is marked default. If you remove a default, LocalAI keeps a record of the removal (a tombstone, removed: true), so the default does not come back at the next start. Turn the source on again to bring it back. A source you added is deleted when you remove it.

The Sources tab

The Sources tab of the Add-ons page lists the sources of all three kinds. Use the kind filter to show only add-ons, backends or models. Each source shows its kind, a Default badge for a shipped default, and its state: enabled, disabled, removed or, while the official libraries are off, off. Administrators can add a source (kind, format, name and address), turn a source on or off, and remove it. A removed default stays in the list with a Restore button. An add-on source can also be edited, to set its signature policy and credential reference, and refreshed. Backend and model sources have no signature policy of their own.

Conflicts

When two libraries list the same entry, LocalAI does not choose silently:

• A source you added wins over a default. The default is ignored for that entry.
• Two sources you added that list the same entry are a conflict. LocalAI refuses to install or resolve that entry until you turn one of the sources off or remove it.

The list shows the message next to each source that clashes. Only sources that LocalAI fetches as a library index can clash, so gallery-format sources never show a conflict.

Use official libraries

The setting official_libraries (default on) switches all default sources of all three kinds at once. In the Settings page it is the toggle Use official libraries, in the group Backends and galleries. On the command line it is --official-libraries (LOCALAI_OFFICIAL_LIBRARIES).

When it is off, LocalAI does not read the default sources. The default add-on library, model gallery index and backend gallery index are not used for search, install or update checks. They stay in the store, still marked default, and the Sources tab shows them as off (suspended: true in the routes). Each source keeps its own enabled state, so turning the setting back on restores exactly what was there. Sources you added are not affected. The change applies at once, with no restart. On a frontend the switch is read-only: the owner's value applies, and a change there reaches the frontends within 60 seconds.

Source routes

The routes are for administrators. A caller who is not an administrator gets 403, and a caller who is not signed in gets 401 when authentication is on. They are not add-on routes, so they do not need the add-ons owner to run. On a frontend they go to the owner (see Frontend limits).

Method and path	What it does
GET /api/library/sources	List every source of all kinds, tombstones included
POST /api/library/sources	Add a source: {"kind": "addon|backend|model", "url": "...", "id": "...", "name": "...", "format": "library|gallery"}. id, name and format are optional. The same kind and URL twice, or a taken id, is 409. An add-on library at an http address of this machine is 400: the owner refuses packages that do not come over https, so such a library would never list anything
POST /api/library/sources/{id}/enable	Turn a source on. This also brings back a removed default
POST /api/library/sources/{id}/disable	Turn a source off without removing it
DELETE /api/library/sources/{id}	Remove a source. A default leaves a tombstone

The list answers {"sources": [...], "officialLibraries": true}. Each source has kind, id, name, url, format, enabled (what LocalAI does with it), default, removed, suspended, conflict (the message, when there is one), policy and, for an add-on source, refresh (the last fetch). A source that does not exist is 404. Changes to model and backend sources apply to the running server. No MCP tool maps to these routes.

Backends in the library

A library lists more than add-ons. Every entry has one schema and an optional kind field. The kinds are:

Kind	What the entry installs	State
addon	A supervised add-on package. It is the default: an entry without kind is an add-on, so every index written before kinds existed stays valid.	Supported
backend	A backend image for the backends directory.	Supported. The backend library is the source of the backends.
model	A model.	Reserved. No reader installs it yet.

The backend library (gallery/backends/index.yaml) is the source of truth of the backends, the way the model library is for the models. The backend gallery (backend/index.yaml) is a generated copy of it. A tagged build before the cutover reads the gallery by default; a build from master reads the library. See The backend library. A library entry of kind backend can also be installed by digest from the Find add-ons tab, as described below.

How a reader treats an unknown kind

A reader keeps the entries of the kinds it knows and skips every other entry. It does not read the fields of a skipped entry, so a newer kind can carry fields this version has never seen. The index is not refused. A skipped entry is reported in two places:

• the refresh result and the state of the library have skipped (a count) and skippedKinds. The Sources tab shows them;
• the refresh writes a warning to the log.

A kind that is not a plain lowercase name, or that is not text, is a malformed entry and the index is refused. An unknown field on an entry of a known kind is still refused. Search lists add-ons unless the caller asks for a kind: catalog-search takes an optional kind argument (addon, backend or all).

A backend entry

- kind: backend
  id: cpu-local-store           # the backend name, also its directory name
  alias: local-store            # optional: the name a model uses
  name: cpu-local-store
  description: Local Store is a local-first vector database.
  license: MIT
  icon: https://example.com/local-store.png
  urls:
    - https://github.com/mudler/LocalAI
  tags: [vector-database, CPU]
  version: 0.0.0-latest.0a1056fd8633
  channel: latest
  schemaVersion: 1
  artifact: oci://quay.io/go-skynet/local-ai-backends:latest-cpu-local-store
  digest: sha256:0a1056fd863328adf9d1097808aa2b52f68c4ca2e5385100a754755b3ca6218f
  mirrors:
    - oci://localai/localai-backends:latest-cpu-local-store
  capabilities:
    default: cpu-local-store
    metal: metal-local-store
  os: linux
  arch: amd64

One entry is one image for one platform, so a backend built for several platforms has several entries with the same id. The reader is strict: an unknown field on an entry of kind backend is refused. These fields are for a backend only and an add-on entry that sets one is refused:

Field	What it holds
schemaVersion	The version of the shared entry header. Absent means 1. A host refuses a higher one.
license, urls, tags	Display text, as in the gallery. urls are http or https.
icon	The URL of the icon, https only. The add-on ui.icon stays a path in the package. LocalAI never fetches the icon: the browser loads it.
mirrors	Fallback images of the same backend, tried in order. Each is an oci:// image. The install pins a mirror with the digest of the artifact.
capabilities	A map from a host capability (default, nvidia, metal, ...) to the id of the concrete backend that serves it.
mutable	The artifact is a moving tag, such as a development image. The entry has no digest.

The digest is the digest of the image for the platform. The install uses the image by digest and never by tag. A backend entry must use an oci:// artifact. It has no main page or ui.icon, and only a backend entry can set alias. The backend name is a directory name: it has no / and no ...

Meta backends. An entry with capabilities and no artifact is a meta backend, such as llama-cpp. It has no digest, os, arch or mirrors, and it is listed for every platform. Installing it installs the backend its capability names for this host, in the same library and the same channel, pinned and verified like any other. A meta backend that names nothing for the host, or names a backend the library does not have, is refused.

Development images. A master- image is rebuilt all the time, so the library cannot pin it. Its entry keeps the tag, has channel: development and mutable: true, and has no digest. The install resolves the tag to a digest, verifies that image, and records the digest it installed in metadata.json.

Signer identity pattern

The signature policy belongs to the library, as for add-ons. A backend library can name the signer with identityRegex instead of an exact identity, so images that the merge workflow signed on master and on a tag both verify:

{
  "requireSignature": true,
  "issuer": "https://token.actions.githubusercontent.com",
  "identityRegex": "^https://github\\.com/mudler/LocalAI/\\.github/workflows/backend_merge\\.yml@refs/(heads/master|tags/.+)$"
}

Set identity or identityRegex, not both. The pattern must be one expression that starts with ^ and ends with $. A pattern with a top-level alternation, such as ^a|b$, is refused, because one branch would match a part of the identity only. An add-on install refuses a library policy that has a pattern: add-on libraries name the signer exactly.

Install a backend from a library

An administrator installs a backend from the Find add-ons tab of the Add-ons page. The entry shows its kind. The page sends the install to the backend route, not to the add-on lifecycle. The page installs an add-on and turns it on in one action (Install and turn on); the install itself leaves an add-on disabled. A backend has no on or off: it is ready to use when the install ends, and its button says Install.

The route is POST /backends/library/install. It is for administrators and exists where the libraries are: on the instance that runs the add-on owner, and on a frontend, which asks the owner to resolve the entry and installs the backend on its own node:

curl -X POST http://localhost:8080/backends/library/install \
  -H 'Content-Type: application/json' \
  -d '{"sourceId":"official-backends","name":"cpu-local-store","version":"latest"}'

The request names the entry and nothing else. The library resolves the image, the digest and the policy, so a request cannot change them. The install:

1. resolves the entry of kind backend for this platform;
2. refuses it when --require-backend-integrity is on and the library does not require a signature;
3. pins the image by digest, verifies the signature before it fetches any layer and refuses an image over the size limit (the layers are summed from the manifest);
4. installs with the same code as the gallery, so the backend lands in the same directory with the same metadata.json, plus library, the ID of the library it came from.

The request answers at once with a job, the same kind of job the backend gallery uses: {"id": "...", "status_url": ".../backends/jobs/<id>"}. Read GET /backends/jobs/<id> to follow it. The phase of the job is the step it is in, in this order: resolving, verifying (only when the library requires a signature), size, fetching (with progress and the sizes of the fetch) and writing. A job that ends has processed set. It ends with the phase done and message completed, or with error set and the phase of the step that failed. The Find add-ons tab shows these steps. Refusals that depend on the entry arrive this way: an unknown entry, a name installed from another origin (a gallery, another library or a manual install; nothing changes), an image over the size limit, a library that strict mode refuses. The request itself answers 400 when it does not name an entry and 409 when the same backend is already being installed. Installing again from the same library is a reinstall.

Library backends on the Backends page

The Backends page marks a backend that came from a library with the library ID, and the detail view shows the digest that metadata.json pinned. A backend of a gallery has no mark, and the gallery upgrade check and the review of updates are unchanged.

When the cached index of the library lists a different build than the one installed, the backend shows "Update available" and an Update button. The check is GET /backends/library/updates. It reads the index the last refresh cached, so it does not use the network, and it installs nothing. The build to compare is the one in the channel of the installed version. If the library no longer lists that version and has more than one channel, the check says nothing instead of guessing. The Update button starts the same install as above, with the same steps, for the version the check names. There is no automatic upgrade.

The LocalAI Assistant and local-ai mcp-server expose this check as the read-only list_backend_library_updates tool. Taking an update stays a confirmed install_library_backend call.

A library install touches only the node that serves the request. In a distributed deployment the backend is installed on that node only; the per-node routes of the backend gallery are unchanged.

Size limits

The limit is on the compressed size of the package or image, by kind:

Kind	Default	Configure
addon	512 MiB	No
backend	8 GiB	--addons-backend-max-bytes

The backend library

gallery/backends/index.yaml is the backend library. It lists every backend of the backend gallery: 1088 backends (934 concrete backends and 154 meta backends) in 1240 entries, because a backend built for several platforms has one entry for each. LocalAI registers it on the first start as the source official-backends, next to the backend gallery localai-backends, and a source you remove stays removed. To add it again later, use the Sources tab with this URL and the policy above:

https://raw.githubusercontent.com/mudler/LocalAI/master/gallery/backends/index.yaml

The library is the source of truth. To add or change a backend, change the library; see [Backends]({{% relref "features/backends#the-backend-library-and-the-generated-gallery" %}}). The digests are never edited by hand. Pin them again after the backend images are rebuilt; this is the only command that reads the registry:

go run ./scripts/backends/cmd/backendcatalog \
  --library gallery/backends/index.yaml --repin

A release image that the registry has no image for when the library is rendered is marked mutable and has no digest, and the command lists it. It installs by its tag until it is pinned.

Manifest

ParseManifest accepts a single JSON object, at most 64 KiB, and rejects unknown fields, duplicate keys, trailing values, and nesting beyond 64 levels. ValidateManifest additionally checks the supplied host compatibility tuple. See tests/addons/fixtures/contracts for service-only, headless service/tool, page/navigation, chat-action-only, and incompatible-host fixtures.

The required fields are schemaVersion (1), id, version, executable, os, arch, host, api, protocol, dataFormat, and capabilities. Compatibility ranges are inclusive positive integer {min,max} pairs, not semver expressions. The package version uses immutable major.minor.patch, optionally a lowercase prerelease starting with a letter; moving aliases and build metadata are not package versions. Registry channels are separate from package identity.

Namespace IDs match [a-z][a-z0-9-]{0,63}. IDs are never paths. An add-on ID must not be one of the first path segments the host serves under /api/addons/: apps, proposals, artifacts, resources, conversations, dispatch, relay, catalog, icon, assets, installed and sources. Such a manifest is refused at parse, at install and when a catalog library is refreshed, because the asset URL /api/addons/<id>/assets/... would be answered by the host route of the same name. A test derives the registered segments from the real routes and fails when one is missing from this list. Executable and UI asset paths are normalized package-relative ASCII paths: no absolute paths, parent traversal, backslashes, percent escapes or URL schemes. These checks do not substitute for safe extraction, symlink checks, executable mode validation, or checking declared files exist, which belong to package installation.

Capabilities are independent: service is a boolean, tools is a unique list of namespace names, and optional ui contains required entry and sdk range, optional css, pages, navigation, and chatActions lists. Navigation names must refer to declared pages. UI assets must not duplicate each other. Absent UI requires neither assets nor SDK compatibility. Declared UI must be consistent; an empty object does not count as a valid declaration. A chat action is either a bare identifier (inert: it has no placement and is never drawn) or a full declaration, and ui.slots and ui.renderers place components in closed host slots (see Extension points and slots). They are host-controlled entries, not DOM hooks or permission to read chat history. Platform tuple equality is compatibility only, not evidence that any release platform has passed native runtime tests.

Four optional fields describe how an addon appears as an app (see Apps). Every one is optional, so earlier manifests stay valid and a headless manifest needs none of them.

Field	Meaning and bounds
launchMode (top level)	Absent or always: the owner runs the addon while it is enabled. on-demand is reserved and, like any other value, is rejected at install.
ui.mainPage	The page the Open action navigates to. It must be one of ui.pages. Without it the addon has no Open action. ui.entry keeps its meaning: the packaged ES module.
ui.icon	A packaged image: .png, .webp, .jpg or .jpeg, a normalized package-relative path of at most 256 bytes, at most 256 KiB. A remote URL, a data: URL, an absolute path and a traversal are rejected. It cannot also be the module or a stylesheet.
ui.title, ui.description	Plain text: one string, or a map of BCP 47 language tag to string (at most 16 tags). At most 80 and 280 characters. Markup characters (<, >), control characters and direction overrides are rejected.

Skills

An add-on can ship skills, the procedures a chat model reads before a task (see [Skills]({{%relref "features/chats#skills" %}})). The model loads one by its name with the chat's load_skill tool, which also names the files of the skill's folder, such as references/, for read_file. The manifest declares them in capabilities.skillDefinitions and the package holds the files:

"skillDefinitions": [
  { "name": "office" },
  { "name": "office-quickstart" }
]

skills/<name>/SKILL.md
skills/<name>/references/<file>      (up to 3 folders below the skill folder)

The description is not in the manifest: the SKILL.md frontmatter is its only source. The manifest is checked at parse, at install and when a library is refreshed:

Rule	Limit
Skills per add-on	16
Name	[a-z][a-z0-9-]{0,63}, unique in the manifest
A plain name	the add-on ID, or the ID and - and more (office-quickstart), so two add-ons never ship one name
A built-in skill name	refused for a plain skill (writing-skills, localai-workspace, localai-apps and the office skills)
shadows	optional; must repeat name and be a replaceable built-in skill: xlsx, docx, pptx or pdf (contract.ReplaceableSkills)

Staging checks the files against the manifest. The folders under skills/ must be the declared names, and a package that declares no skills must have no skills/ folder. Every entry is a plain file or a folder: links, special files and executable files are refused. A path segment uses lowercase ASCII letters, digits, ., _ and -, does not start with a dot, and a path is at most 128 bytes. A file is a .md, .txt, .json, .csv, .yaml or .yml file of UTF-8 text with no NUL byte. There are no scripts: .py, .sh, .js and other files are refused. A skill that needs to run code is a built-in skill with requires: [exec]. SKILL.md is at most 64 KiB and must parse with a name equal to its folder and a description; a reference file is at most 256 KiB; a skill has at most 64 files and 1 MiB, and a package at most 4 MiB of skill files. The receipt records each file with its size and SHA-256 digest.

The add-on owner copies the skills into the chats' add-on skill layer, in the workspace at /addon-skills/<add-on>/, when it installs or upgrades the add-on and every time it starts, and deletes them when the add-on is uninstalled. Each copied file is checked against the digest in the receipt, so a package folder changed after install never reaches a chat. A copy that fails is logged; the add-on itself still installs and runs. The files stay while the add-on is stopped. A chat offers them only while the add-on is installed, running, usable by the user and on for that chat, and for a chat with a saved agent only when the agent grants the add-on.

Skills that replace a built-in skill

An add-on skill can replace a built-in skill of the same name, as a tool replaces a built-in tool (see Tools that replace a built-in tool). Only the office skills xlsx, docx, pptx and pdf can be replaced; writing-skills, localai-workspace and localai-apps cannot. The manifest says so with shadows, which repeats the name:

"skillDefinitions": [
  { "name": "xlsx", "shadows": "xlsx" },
  { "name": "docx", "shadows": "docx" }
]

• In a chat. While the add-on can serve (installed, running, usable by the user and on for the chat), the model sees one xlsx skill, the add-on's, and the built-in one is hidden: load_skill xlsx returns the add-on's. The index grows by nothing: one line replaces another. A skill a person writes (folder, personal or shared) still wins over the add-on's.
• The built-in comes back. When the add-on is stopped, off for the chat or removed, or the chat switches the add-on's skill off, the built-in skill serves again. The turn tells the model in one line, and the Skills menu says so on the built-in skill.
• Checked twice. The install refuses a shadows that is not on the list, and the chats refuse it again at each turn: an add-on skill with the name of a built-in skill that cannot be replaced is skipped as invalid.
• Several add-ons. Only one add-on serves a name. An administrator picks it in Settings, Tools, in the same runtime setting as the tools (chats_tool_providers, for example {"xlsx": "office"}; tool and skill names never overlap). With no choice, the add-on installed first serves, by the same rule as for the tools (see Install time). A choice of an add-on that is not running leaves the name to the built-in skill. The Skills page marks the skill of an add-on that is not chosen "Not active: provides it", and the built-in skill "Replaced by ". GET /api/chats/admin/tools lists the replaced skills in skills, with who serves each one, the candidates oldest install first, default (the one that serves with no choice) and installed_at (the install time of each).
• Before install. The library entry and the installed view list the skills an add-on replaces in skillShadows, and the install dialog names each one: "This add-on replaces the built-in skill xlsx." When another installed add-on already replaces it, the dialog says that only one can be active and that the choice is in Settings > Tools. The install is still allowed.
• On the Add-ons page. The details of an add-on list its chat skills and the built-in skills it replaces. For an administrator, a skill that several add-ons replace has a switch, Use for xlsx in the chats: on, this add-on serves it; off, the next add-on that replaces it does. The switch writes the same choice as Settings > Tools.

In distributed mode the state is shared: the files are in the workspace store that every frontend reads, and only the owner writes them, because only the owner holds the verified package. Which add-ons are usable is read from the broker at each turn, as for the tools. With the owner down the add-ons are down, their skills are not offered and the built-in skills serve.

SDK versions

This host provides SDK version 2. An add-on that has a UI declares the SDK versions it works with in ui.sdk as a range, for example "sdk": {"min": 1, "max": 2}. The host installs and loads the add-on only when its own SDK version is inside that range. If it is not, the install is refused with a message that names the range the add-on needs and the SDK version the host provides:

the add-on needs SDK 3 to 3 and this host provides SDK 2

The dispatch response for that refusal has status 403 and error type sdk_incompatible, with the message above. A package that does not match its library entry, or is not a valid add-on package, is 403 with error type package_unverified and a message that says nothing was installed. A failed install or upgrade can also name its step in the error type: library_unavailable (the library could not be read), library_conflict (two enabled libraries list the add-on), release_not_found (the library has no such add-on or version for this platform), package_fetch_failed (the package could not be downloaded) and install_storage_failed (the package could not be written to the add-on storage). The owner writes the underlying error to its log as Addon operation failed, with the action, the add-on ID and the operation ID. Every other failed install keeps the generic Addon request denied message. The owner reads the library once on demand when its startup refresh left nothing in the cache, so an install does not depend on that refresh having finished. If the administrator sends the same operation ID again, the owner returns the stored failure, so the repeat gets the generic message and not the refusal text. Use a new operation ID to see the refusal again.

The SDK 2 surface

An add-on page gets these from the SDK object. The page component also receives the query string of its address as search, next to page and path. Props pass through unchanged, and the components are the ones the core pages use, so an add-on page and a core page draw the same thing.

• sdk.ui: Button, Text, Stack, Badge, Spinner, EmptyState, and the shared PageHeader, ConfirmDialog, ErrorWithTraceLink, ModelSelector, MediaInput, WaveformPlayer, DistributionBars, EmbeddingInspector, TabSwitch, LoadingSpinner, Modal, Lightbox and HomeModelPicker. A ref reaches the host component, so HomeModelPicker keeps its handle.
• sdk.ui also has the settings kit, for a settings section or any form: see The settings kit.
• sdk.theme.tokens: the host's design tokens as CSS values (var(--dk-...)), for the add-on's own scoped style sheet: text, muted, surface, inset, edge, hover, accent, accentText, accentWash, onAccent, okText, okWash, warnText, warnWash, errorText, errorWash, space1 to space6, radius, radiusSmall, radiusMedium, textMeta, textSmall, textBody, fontBody and fontMono. Every value is a custom property the host theme defines, in both themes. The same custom properties can be written in the style sheet directly.
• sdk.hooks: useMediaCapture, useObjectUrl, useDelayedAction, audioBufferToWavBlob, useModels, useResources, useVoiceProfiles, useOperations and useMCPClient. They run in the host React, so call them like any hook. A host without them gives an empty object.
• sdk.navigation.openApp(address, { replace, state }): opens a page of the web app, a core page or a page of another add-on. Only addresses under /app pass; any other address, scheme or API route is refused. The only state a page may pass is the way back, { from, fromLabel }, which the Model Editor reads.
• sdk.portal(children, node): draws children into a node outside the add-on's own tree, for a full-screen sheet.
• sdk.viewer: isAdmin() and hasFeature(name). They answer for the signed-in user and are read-only. The server still decides every call.
• sdk.iconNames: the icon names sdk.Icon can draw. Any other name draws a neutral glyph. SDK 2 adds arrow-left, arrow-right, chart-bar, circle-dot, clipboard, code, cube, filter-off, gender, globe, headphones, help-circle, mic, play, plug-off, save, shield, smile, waveform, camera, stop, circle, image, spinner, user-shield and mask. The Studio and Talk add-ons add boxes, compass, gauge, ghost, git-branch, images, key, layers, maximize, memory, music, pencil, repeat, server, sliders, square-dashed, store, undo, video, volume, palette, phone, columns and chevron-left. sdk.Icon passes a style object.

SDK 2 is a superset of SDK 1, but a range is checked as written. A manifest that declares only "max": 1 is refused by a version 2 host. Before you raise max to 2, check the add-on against the SDK 2 surface. A range that is malformed (for example min above max) keeps the generic invalid or incompatible UI error.

Runtime wire version 1

JSON field names and ordering follow the exported DTO declarations and encoding/json; envelope.json fixes the canonical envelope representation. Opaque arguments, payload, and preview summary are JSON, not executable SQL or authority. This is not a cross-language cryptographic canonicalization scheme; forwarding must bind the actual transmitted bytes. Unknown/duplicate fields are rejected at manifest and envelope/result/preview/receipt boundaries.

An envelope consists of request and host-issued authority. Request carries addonId, optional resourceId (empty for addon-level actions), operationId, nonzero lifecycle generation, action, and object arguments. Authority carries actor (userId, credentialId, mode), scope, and RFC3339 expiresAt. Modes are named-user and explicit local-owner; scopes are read, preview, apply, output and act. The scope act and the action act go together: an envelope has one only with the other. Principal IDs are opaque nonempty strings up to 256 bytes. Syntactic validity is not authentication: the owner must create authority, verify current identity/credential/ACL and mode, and bind scope to the actual method. Never accept browser/model-provided approval. Call Validate(now, limits) at admission; unmarshalling checks structure and nonzero expiry but intentionally does not use the wall clock.

Runtime provides Read, Preview, Apply, Receipt and OpenOutput. A runtime that serves action tools also implements ActRuntime (Act), described below. Preview is {operationId,revision:{schema,data},expiresAt,summary}. Schema revision is nonzero; data revision can start at zero. Apply receives that exact revision separately and must reject stale schema or data. Receipt is {operationId,state,revision:{schema,data},result}. States are pending, committed, rejected, unknown; unknown is not success or rejection and must be reconciled before another mutation. Committed receipts require a nonzero schema revision. Result has optional payload and outputs; each output has opaque handle, safe basename name, MIME mediaType, and nonnegative size. Handles are not URLs or filesystem paths. OpenOutput returns an authenticated stream; callers close it. DTOs do not create artifacts or authorize downloads.

The settings kit

sdk.ui carries the host components a settings section, or any form of an add-on, is made of. They use only the host's classes and tokens, so an add-on draws its settings the way the host draws its own, in the light and the dark theme and at phone width, without a style sheet of its own. Use them instead of plain <input>, <select> and class names of your own: a class the host does not define has no style.

Component	Props	What it draws
Section	title, description, actions, testId	A titled block with a line under the title. Sections that follow each other are separated by a rule.
Field	label, help, unit, error, testId	A label, the control (the children), its unit after it and a help line under it. The label names the control.
Select	value, onChange(value), options: [{value, label, disabled}], label, disabled, testId	The host select.
TextInput	value, onChange(value), type (text, password, number, url, search), placeholder, min, max, mono, label, disabled, testId	The host text input.
Switch	checked, onChange(next), label, disabled, testId	The host switch.
Choice	legend, hideLegend, value, onChange(value), options: [{value, label, description, chip, disabled}], disabled, testId	One choice of a few, each option with a line that says what it does.
StatusChip	tone (ok, warn, error, idle, accent), title, testId	A state in colour, icon and words.
Callout	tone (info, ok, warn, error), title, actions, role (alert), testId	A notice inside the page.
List, ListRow	List: label, testId. ListRow: title, meta, status, details, detailsLabel, testId	Rows in one bordered list: a name, a line under it, a chip on the right, and details collapsed under a toggle.
Code	testId	A monospace block for a raw error or a command, wrapped and scrolled.
Actions	none	A row of buttons that wraps.
Table	caption, columns: [{key, label}], rows, testId	The host table.
ExternalLink	href (https only), testId	A link out, in a new tab.

The Sandbox and Web sections are made of these components only (addons/sandbox/ui/src/Settings.jsx, addons/web/ui/src/Settings.jsx).

Operation IDs and errors

Generate an operation ID once, before submission. Bind it to actor, addon, resource, generation, action, arguments, and expected revision. Repeated IDs with identical binding retrieve the same durable outcome; changed binding is a conflict. Never retry a mutation under a new ID after timeout, disconnect or unknown receipt. Cancellation does not prove rollback. Runtime implementations must preserve these invariants transactionally; these contracts contain no storage or deduplication implementation.

HTTP adapters should map malformed input to 400, missing authentication to 401, current permission denial to 403, absent permitted resources to 404, stale revision/generation or conflicting operation binding to 409, expired proposals or artifacts to 410, oversized requests to 413, admission/quota limits to 429, and unavailable owner/runtime to 503. A timeout (504) is an uncertain mutation outcome requiring Receipt lookup, not implicit retry. Errors must be redacted and must not leak paths, credentials or unauthorized resource existence. No HTTP error DTO or endpoint is implemented by this package.

Finite implementation defaults

Start with DefaultLimits(), change explicit host configuration fields, then call Validate(). All values must remain positive; zero never means unlimited. Durations are time.Duration (nanoseconds if serialized as JSON). These are implementation defaults, not immutable product requirements.

Limit	Default
Manifest / envelope / payload bytes	64 KiB / 2 MiB / 1 MiB
Read / preview / apply / shutdown timeout	30 s / 30 s / 60 s / 10 s
Initial / maximum backoff; attempts	250 ms / 30 s; 5
Batch size / query depth / query rows	100 / 8 / 1000
Query result bytes	1 MiB
Proposal expiry / artifact expiry	5 min / 1 hour
Single artifact / aggregate artifact quota	100 MiB / 1 GiB (a stored artifact is at most 64 MiB, the relay bound)
Artifact count / concurrency	100 / 4

Initial backoff cannot exceed maximum; single artifact cannot exceed quota; query bytes cannot exceed payload bytes, which cannot exceed envelope bytes; query depth cannot exceed the parser's 64-level ceiling. Runtime/owner consumers must enforce timeouts, bounded backoff (without new mutation IDs), query/batch limits, proposal lifetime, artifact expiry/quotas and concurrency. DTO validation alone does not run timers or jobs. Default JSON unmarshalling and ParseManifest use fixed default ingress bounds; a transport needing different bounds must use a bounded, strict wire decoder and validate the resulting DTO with its configured limits (not rely on default unmarshalling to enlarge ingress limits).

Verification

Run sh scripts/addons/verify.sh contracts. The inventory requires every named Ginkgo case to pass in its JSON report: no empty selections, skips, stale reports, or successful commands without reports. sh scripts/addons/verify.sh prerequisites regression-tests those gates and checks real native toolchain/browser readiness.

Private catalog service

core/services/addons/catalog provides owner-side source configuration, refresh, search and release resolution. It does not expose routes, UI, installation or runtime execution yet. Model and backend managers remain separate.

Construct Sources with a protected owner configuration path, an authoritative AdminCheck, and a CredentialResolver. Save creates or replaces a source; Source{ID: id, Removed: true} explicitly removes one. Source records contain ID, display name, HTTP(S) URL, enabled state, authentication reference and verification policy. Files are replaced atomically with mode 0600. The operator must provide a private parent directory and only one writer. Loading rejects symlinks and group/world-readable files. Source edits invalidate cached results. Removing a source never opens or changes installed packages, receipts or data. Installed receipts must retain their pinned source; changing that association requires a separate authorized lifecycle operation, not a catalog refresh.

Authority and credentials

The existing core/http/auth.RequireAdmin checks an authenticated Echo user's role. contract.Actor has no role and is not an authenticated Echo context. This service therefore injects AdminCheck(context.Context, contract.Actor); a missing checker denies all edits and configuration reads. The future owner adapter must revalidate current user, credential and administrator status. It must not infer privileges from user IDs, mode strings, model arguments or browser values. Refresh is an internal owner operation; a future public refresh/test route must enforce admin authority before calling it. No new secret endpoint is provided.

Configure reference bindings only in trusted owner code. Each binding pairs an exact scheme/host/port origin with a pkg/credentials.Store. That store supports bearer/basic secrets, including existing environment/file references. Secrets resolve at request time. Source JSON contains only the reference; resolver JSON omits bindings. Catalog credentials are not installed into the OCI downloader's store and releases do not carry authentication references. Configure OCI pull credentials separately. Plain HTTP credentials require the store's explicit allow_insecure setting; use HTTPS in production. URLs with embedded credentials, query parameters or fragments are rejected.

All catalog requests use pkg/httpclient, a 30-second timeout, and no redirects. Redirect, credential, HTTP-status and parse failures return redacted errors, not remote bodies or URLs. A failed refresh records its check time and last successful refresh time and prevents stale cached releases from resolving successfully.

Catalog document and limits

A catalog (the library index) is one YAML document: a block list of releases with id (an index written before this name may still use addonId, which is read for one release), name, optional description, version, artifact (oci://host/repository), digest (sha256: plus 64 lowercase hex characters), os and arch. It uses the same format as the model and backend gallery indexes:

- id: tables
  name: Tables
  description: Tables for LocalAI
  version: 1.0.0
  channel: latest
  artifact: oci://quay.io/go-skynet/local-ai-addons:1.0.0
  digest: sha256:0000000000000000000000000000000000000000000000000000000000000000
  os: linux
  arch: amd64
  ui:
    title: Tables
    mainPage: databases

An entry may carry a kind (see Backends in the library); without it the entry is an add-on. Unknown fields, a second YAML document and duplicate keys are rejected. A JSON array is refused too, even though YAML can read it: write the index as a block list. An empty library is []. Package manifests (manifest.json) stay JSON; only the library index is YAML. Source identity and verification policy always come from protected source configuration, not from the remote release. The policy describes requirements for later package verification; refresh does not claim to verify executable signatures.

A release may carry an optional ui object of display hints: title, description (default-language plain text), mainPage, icon (a path inside the package, never a URL) and launchMode. They follow the manifest rules for bounds, markup and launch mode, and a catalog that breaks them fails refresh as a whole. They only help browsing before install: the manifest inside the verified package is what an installed addon shows, and the host serves an icon only after install.

A release may also carry toolCount and tools, the same names the installed view uses: the chat tools the package declares in capabilities.tools (admin tools are not chat tools and are not counted). tools holds at most 32 names and toolCount is the full number. The add-on cards, the Find add-ons tab and the details panel show "Adds N tools" from them before install. addonpkg catalog writes both fields, and addonpkg catalog-tools adds them to an existing index from the manifests in addons/<id>/manifest.json. They are a hint like ui, and only an add-on entry may carry them.

Refresh accepts at most 1 MiB and 1,000 releases. Duplicate release identities within a source fail. Duplicate names across sources remain distinct. Search matches names, IDs and descriptions, sorts by source/release identity and uses a default page size of 50 (maximum 100). Disabled sources are neither searched nor fetched. Resolve requires an explicit source, addon ID and exact version; it selects only the exact owner target OS/architecture. Configure Catalog.TargetOS and Catalog.TargetArch in trusted owner code before use; each empty field defaults to this process's runtime.GOOS/runtime.GOARCH. Forwarded installs must use the owner's target, never the forwarding frontend's platform. An absent target returns ErrNotFound; ambiguous exact-target entries fail closed (duplicate identities are also rejected during refresh). Search still shows all available platforms. Optional channel labels (for example latest, development, or branch/name) resolve to a single exact-platform release and its pinned digest. version remains the immutable manifest version; selection is stamped by Resolve, never trusted from remote metadata. This service does not silently choose a source or upgrade.

Sources.List(ctx, actor) and Sources.Get(ctx, actor, sourceID) recheck the same authoritative AdminCheck used for edits. List returns all configured sources in ID order, including disabled libraries and libraries with no releases or no successful refresh. Get returns ErrNotFound for an unknown source after authorization. Both return independent value copies with authentication references and verification policy, never resolved credential values. Reads are serialized with source edits and refresh; configuration remains readable after reload without refreshing.

State uses the documented per-instance designated-owner exception: only the configured owner constructs this service. Configuration persists; the bounded catalog cache is rebuilt by explicit refresh after restart. Replicas must forward to that owner, not construct independent caches or fall back on owner failure.

Run sh scripts/addons/verify.sh catalog for private bearer/basic HTTP fixtures, redirect rejection, isolated credential references, disabled sources, refresh failure, duplicate names, owner-target resolution, source persistence/removal, denied reads/edits, defensive copies and concurrent configuration reads.

Verified immutable package staging

The designated owner uses packages.Installer.Stage followed by Commit. Stage re-resolves the exact catalog selection and rejects caller-substituted policy, source, platform, artifact or digest. A stale channel selection must be resolved again explicitly. Package metadata cannot override the source administrator's verification policy. Installable code is trusted code, not a sandbox.

OCI images are distribution archives, not running containers: no Docker daemon is needed. The installer pulls the catalog-pinned per-platform manifest digest, checks image platform and compressed layer hashes, and uses the existing Sigstore verifier when the configured source requires signatures. The original reference and selected channel are retained alongside the actual digest. OCI credentials remain in the registry credential store; catalog HTTP credentials are never passed to a registry or package URL. A registry may redirect a digest-pinned blob download to a storage host, as quay.io does. The installer follows that redirect only, at most three times, without the registry credential, and checks the blob digest on the bytes it receives. Every other registry redirect is refused.

When the manifest declares ui.icon, staging opens it inside the package without following links and requires a regular file of 1 byte to 256 KiB whose first bytes are the signature of the declared type (PNG, WebP or JPEG). A missing, empty, oversize, linked or mistyped icon fails staging. The receipt records the icon's path, SHA-256 digest, size and media type; the host serves the icon only if the bytes still match that digest.

HTTPS package URLs distribute uncompressed tar archives with a required catalog SHA256. HTTP is rejected unless an administrator explicitly sets the source's allowLoopbackHTTP for a literal loopback IP (development/testing only); public insecure HTTP and credential-bearing URLs are rejected. HTTP package downloads are anonymous and do not reuse catalog credentials; redirects are refused. Archive signature verification is not implemented: a source requiring signatures must distribute OCI, and HTTP packages from that source fail closed.

Packages contain manifest.json, the declared executable, and any declared UI entry/CSS files. A service-only manifest needs no fake UI files. Host/API/protocol, data format, SDK (only with UI), identity/version and platform must match. All entrypoints/assets must be real regular files inside the immutable package; executables must have the owner execute bit. Links, special files, traversal, duplicate paths, file/directory collisions, and a supplied receipt.json are rejected. OCI layers cannot overwrite earlier paths: packages are flattened application bundles, not general-purpose container root filesystems.

Defaults cap compressed downloads and extracted content at 512 MiB (a backend image has its own limit) and archive entries at 10,000; owner configuration may set smaller positive limits. Staging is private, cancellation-aware, and removed on failure. Existing packages are never replaced or deleted. Commit publishes a unique read-only package directory with receipt.json containing source, selection/channel/reference, verified digest, verification result and manifest. The opaque stage handle is installer-bound and single-use. Root must be owner-controlled, separate from mutable data, backups, and addon-writable directories. Publication does not activate or start code; lifecycle journaling and abandoned-stage recovery belong to the owner manager.

Run sh scripts/addons/verify.sh packages for local HTTP and OCI registry fixtures, malicious archives, compatibility checks, cancellation, storage failures and immutable publication. No external container images are published or downloaded.

Package transport applies the source's allowLoopbackHTTP development opt-in at the OCI transport boundary, including registry authentication and signature requests. Without it, implicit loopback HTTP fallback is rejected before any plaintext request. Non-loopback HTTP is always denied, and a blob redirect target must use HTTPS or, with the opt-in, a literal loopback IP. Other registry redirects are denied. The expanded package byte limit counts the entire decompressed stream before TAR parsing: PAX/GNU metadata, file headers, payload, padding, and trailing bytes, cumulatively across OCI layers. Downloaded compressed bytes have a separate budget of the same configured size. Failed staging leaves installed packages intact.

The focused package tests use local OCI registries and block external trust requests. Hermetic signed fixtures include trusted Fulcio, SCT and Rekor material and exercise successful cryptographic verification, staging and commit. They also verify rejection of signed identity, digest and trust mismatches. These offline checks do not demonstrate live Sigstore verification; production signature requirements are unchanged.

Extraction bounds both the decompressed TAR stream (including metadata, padding, and trailing bytes) and cumulative file output across all OCI layers. Sparse files cannot bypass the output limit by representing large holes with small TAR entries. Failed staging removes temporary output without changing old packages.

Embedded hosts may explicitly provision Installer.SignatureTrust for a private Sigstore deployment. This is host configuration, never catalog or package data; it replaces only the trust material. Certificate identity, SCT, Rekor timestamp, and artifact digest checks remain enabled. Leaving it unset retains public Sigstore TUF trust. Offline acceptance fixtures generate local CA, signing, CT, and Rekor keys and exercise the same signature verification and staging path; they do not disable cryptographic requirements or depend on live services.

Durable lifecycle and designated ownership

The lifecycle library uses a designated-owner per-instance state exception. It does not use a rotating leader loop, election, or automatic failover. Embedded hosts construct addons.NewManager with an explicit OwnerConfig:

• InstanceID identifies this process's deployment role.
• OwnerID identifies the only role permitted to own addon storage.
• OwnerAddress is the configured HTTPS address of that owner.
• StorageID identifies the durable local volume.
• Root is an absolute, private, owner-controlled directory.
• SharedStorage must be false. Shared-storage deployments are unsupported.

The host must assign unique instance identities and a local filesystem with working advisory locks, atomic rename, and file and directory synchronization. A process holds the storage lock until Manager.Close. The stored owner identity, address, and storage identity must match on restart. A conflicting configuration fails instead of adopting existing storage. Do not mount this root on replicas. A frontend manager never creates or opens the root, even when the owner is absent. Its lifecycle calls return ErrNotOwner; forwarding transport is a separate layer. Application bootstrap accepts config.WithAddons(config.AddonConfig{...}). Enabled defaults to false; no owner is inferred from missing configuration. The owner loads protected sources and reconciles its journal before exposing Application.AddonManager(). Application shutdown releases the storage lock. Forwarding instances construct no catalog or installer and never reconcile storage. AdminCheck is mandatory on the owner and excluded from serialized settings. Host optionally sets compatibility; its zero value uses the current host and version-one contracts. OwnershipMode is named-user (the default when empty) or local-owner; see "Explicit local-owner mode" below. local-ai run sets it from --addons-ownership-mode. Like the rest of AddonConfig it is startup-only configuration with no runtime setting, and every frontend of one deployment must use the same value. Application bootstrap starts enabled addon processes after durable reconciliation.

The host supplies an authoritative catalog.AdminCheck. It must check current account, credential, and administrator status, including for duplicate requests. Actor fields do not grant administration rights. Installation rights do not grant database access.

Submit accepts install, upgrade, enable, disable, and uninstall. Install and upgrade consume the actual catalog and immutable package installer. Installation starts disabled. Enable records desired state; it does not report that a process is running. Status separates installed state, desired state, and observed health. StartEligible requires ownership, installation, enabled intent, and no unfinished operation. Runtime supervision is a separate layer.

Operation IDs are global to the owner journal. A retry must carry the same actor and complete command. Conflicting reuse returns ErrConflict. Failed operations remain failed; an explicit retry needs a new operation ID. Commands serialize through the manager lock, including commands for the same addon.

The journal synchronizes complete snapshots at intent, prepared, applied, and completed boundaries. Published packages are synchronized before prepared state. Call Reconcile after restart before admitting runtime work. Recovery fails unfinished intent rather than repeating a download or publication. Prepared and applied operations finish from their recorded generation without resolving a new catalog release. Persistence errors block further commands until reconciliation.

Uninstall clears installed and enabled intent. It retains immutable generations, addon identity, data paths, and grant metadata. Mutable data lives under data/; owner metadata lives under metadata/; immutable packages live under packages/. Reconciliation removes abandoned stage-* directories but retains published packages, including packages published just before a crash. Operators must inspect unreferenced published packages before removing them. No automatic purge runs.

Drain, backup, and migration hooks are interfaces only. The lifecycle library does not claim to execute those operations. Later runtime integration must implement these hooks before changing packages for running processes or migrating data.

Run sh scripts/addons/verify.sh lifecycle for journal interruption, ownership, idempotency, retention, and real catalog/package integration tests.

Owner bootstrap reconciles the journal and starts the enabled add-ons, and refreshes all persisted enabled sources in the background. The server does not wait for the refresh: until it ends, the release cache is empty, as after any failed refresh. The release cache is not durable. The whole startup refresh has a 30-second deadline (or the earlier application-context deadline); an offline or invalid enabled source costs a log line, not the start. The enabled add-ons start four at a time. Disabled sources are never fetched. Forwarders neither read source metadata nor contact catalog/package servers. Embedding hosts can use Application.AddonSources() (owner only; nil on forwarders) for authoritative source administration and explicit live Refresh(ctx, sourceID) calls; no background update loop is started.

Recovery rejects symlinked package roots and ancestors before enumerating staging. Staging cleanup uses directory-rooted operations and descriptor-based permission changes so links cannot redirect cleanup into external storage.

Managed runtime

Application startup attaches the process supervisor only on the configured owner, after journal reconciliation. Enabled generations start at attachment. Enable, disable, upgrade, uninstall, and application shutdown synchronize process state; disable and uninstall retain data. Forwarders do not attach a supervisor or execute packages. No automatic owner failover is provided.

An upgrade of a running add-on proves the new generation before the journal moves: the owner stops the running process, starts the new generation and waits for it to become ready. When it does not become ready, the previous generation is started again, the journal keeps it as the installed generation and the operation fails (a new operation ID retries). The published package of the failed release stays on disk; retained data is untouched. Killing the owner during that start has the same result after a restart, because the journal never named the new generation.

An enabled add-on whose process cannot start does not stop the owner. Startup records the failure for that add-on only: it stays enabled in the journal and stopped in the observed state, the other add-ons start, and commands on other add-ons are not delayed or failed by it. The command that enables or upgrades the broken add-on reports the failure, and a restart tries it again.

The owner executes the receipt's package-relative executable directly with the single argument --localai-addon-v1. There is no shell. The environment allowlist is LANG=C and TZ=UTC; host credentials and proxy variables are not inherited. Descriptor 3 contains an ephemeral listener bound to 127.0.0.1. Descriptor 4 contains one newline-terminated JSON runtime.Startup object: registration, private random bearer credential, and finite host limits. The child must monitor that descriptor after decoding and exit on EOF. EOF is cooperative liveness, not an ownership fence: a paused child may not observe it before its owner restarts.

The supervisor requires a private durable StateRoot (application default: <addon root>/runtime). A separately locked, fsynced process ledger records Linux boot identity, PID, kernel start time, and incarnation before startup credentials are sent. A pending record is persisted before exec. A restarted owner recovers all recorded children using pidfds, checks exact identity, sends termination via the retained handle, and waits for confirmed exit before any replacement launch. Reused PIDs are not signaled. Corrupt, incomplete, unreadable, or uncertain pending records fail closed; operators must establish that prior children are gone before repairing the ledger. Do not delete runtime metadata to bypass recovery. Linux with readable /proc and pidfd support is currently required; unsupported systems refuse runtime startup rather than weaken identity checks. This is not a sandbox or descendant containment: installed code is trusted, and must not detach workers outside the supervised process lifetime.

Every local HTTP request requires the bearer credential and X-LocalAI-Incarnation. The authenticated /health response must exactly match the supplied registration, including protocol 1, endpoint, generation, incarnation, and manifest capabilities. Capabilities are visible only after readiness and are withdrawn on stop, exit, or failed health. Health is checked periodically. A restart uses a fresh credential and incarnation, bounded exponential backoff, and the configured attempt budget; sustained authenticated health resets that budget.

The protocol uses POST with JSON on /read, /preview, /apply, /receipt, and /output. Read and preview accept the existing contract envelope. Apply wraps it with a revision; output wraps it with a handle; receipt accepts an operation ID. These transport wrappers are defined in the runtime package because the original contract specifies Go methods but not their HTTP request DTOs. No contract method or authority semantics changed. Receipt lookup is an internal recovery operation, not a public authority-free endpoint. The host must authorize access before calling the runtime client; model arguments cannot supply trusted authority.

The client rejects redirects, nonliteral/nonloopback endpoints, stale generations, wrong method scopes, expired authority, oversized responses, and excess concurrent requests. Streams have byte and time limits and propagate cancellation. Stop rejects new admission and withdraws registration before bounded draining. Termination uses the retained OS process handle, never a reconstructed numeric PID. An old stopped instance cannot terminate a newer generation. The addon slot stays reserved during drain and restart startup, including while the next process is not yet ready. Stop success means all slot launch workers and processes have finished; cancellation or cleanup timeout returns an error, retains the reservation until confirmed cleanup, and never authorizes an overlapping generation. A failed durable cleanup fences further launches. Stop callers wait at most two shutdown intervals; startup callers wait at most readiness plus two shutdown intervals.

Supervisor.Current(ctx, addonID, generation) and Manager.Runtime(ctx, addonID) explicitly obtain the current ready incarnation under lifecycle locks. A client previously returned by Start or lookup stays bound to its original incarnation and rejects calls after withdrawal. No stale mutation is transparently replayed on a replacement; callers must use the existing operation/receipt recovery protocol.

Default shutdown is 10 seconds; readiness is 30 seconds, restart delays range from 250 milliseconds to 30 seconds, and the restart budget is five attempts. Runtime startup failure leaves durable enabled intent intact for explicit reconciliation; it does not claim healthy service.

Shutdown withdraws admission and starts cleanup for all managed addons before waiting for any child. Cleanup errors are combined; a timeout does not release the process ledger or owner fence. Retry shutdown to join outstanding cleanup.

A completed lifecycle journal receipt records durable desired state, not proof that runtime synchronization succeeded. Replaying the same authorized operation ID retries synchronization against current desired state and returns cleanup or startup errors until synchronization succeeds. It does not repeat package publication or restore an older operation's desired state.

If lifecycle publication fails and the manager closes its runtime, that manager requires recovery. Status, lifecycle submission, and reconciliation cannot report success using retained instance handles. Repair the storage fault, successfully close the manager, and create a new owner manager and supervisor to reconcile the durable journal. The existing Linux proc/pidfd identity checks still apply; ambiguous pending launches require operator recovery rather than an unsafe kill.

Action tools

A tool of kind action does what a person asked a chat to do, at once, as that person, and in that person's own space. It is not confirmed in the browser and it is not a proposal. The sandbox add-on, which runs a command, and the office add-on, which edits a file the user owns, use it. An action tool is not for data of a shared resource: a change that others see is a write, schema or create tool, which the OpenAI-compatible route leaves as a proposal for a person to confirm.

• Declaration. The tool is in capabilities.tools with a definition of kind action, and the manifest sets capabilities.actions to true. The host refuses a manifest that declares an action tool without it. An action tool names no resource: if its schema has a resourceId, the host ignores the value the model gives and sets it to the ID of the conversation.
• Who may call it. Only the chats consumer (studio-chats-addons in the consumer table of Tables). Every other consumer neither lists the tool nor runs it: the old chat route, the direct API, agents and scripts.
• The call. POST /act with an envelope whose action and scope are act. The actor is the user of the chat. The answer is a Result. The call may run up to ActTimeout (30 minutes), is never retried by the host, and an add-on that dies during it gives the caller outcome_unknown. The request carries the chat's conversation ID in resourceId. The host sets it from the chat, the model cannot name it, and an add-on uses it only to keep state for that chat, never as authority.
• Undo. What a tool does to the user's files is a change in the history of their workspace, and it can be undone there.

Host operations

A host operation is how an add-on process asks the host to do something the add-on cannot do itself, such as starting a job of the engine. It is contract surface that third-party add-ons cannot use unless the host setting names them.

The rules:

• The host lists every operation in its code. No add-on gets an operation that is not listed, and an operation that is not registered is refused. The default is refuse.
• The host setting --addon-host-ops (LOCALAI_ADDON_HOST_OPS) names the add-ons that get the channel and the operation prefixes each may call, as a comma list of addon-id=op-prefix. The default is fine-tuning=finetune.,quantization=quantize.,sandbox=workspace.,sandbox=egress.,office=workspace.,web=workspace.,telegram=connector.. An empty value names nobody. This is a host setting, not a manifest field and not a model argument.
• An add-on that the setting names finds two optional fields in its startup object: hostEndpoint, a loopback address of the form http://127.0.0.1:<port>, and hostCredential, a secret of its own that is never the credential of the host-to-add-on channel. Both belong to one incarnation and end with it. An add-on that is not named gets neither field. The protocol version stays 1, and an add-on that ignores the fields works as before.
• A call is POST with the header X-Addon-Host-Credential and a JSON object {"operationId", "op", "arguments"}. operationId is the ID of a request the host is serving for that add-on at that moment. The host looks up the envelope it issued for that request. A call with no live request, with an expired authority or with a scope below the scope of the operation is refused. An apply operation needs an apply request; a read operation runs under a read, preview or apply request.
• The operation runs as the actor of that envelope, never as a user named in the arguments. The host refuses an argument with a name of identity or authority (for example userId, scope or approved).
• The host checks the feature of the operation for that actor, as the route middleware does for the same request. In local-owner mode the actor is the local owner and holds every feature.
• A body larger than the payload limit is refused. An operation that is marked idempotent runs once for one operation ID; a repeat gets the first result.
• A process that does not run the add-on manager offers no endpoint, so the operation fails closed.

The answer is {"result": ...} or {"error": {"code", "message"}}. The error codes are no_live_operation, scope, feature, unknown_op, bad_request, too_large, not_found, conflict and unavailable.

The channel itself ships with no operation. Each operation is added by its own change and documented with it.

Inbound operations

One class of operation has no user behind it: a message that comes from outside, such as a Telegram message pushed by an add-on. Its scope is inbound. The rules above change in these ways, and only for operations the host registers with that scope (today connector.inbound):

• No live request is needed. The authority is the credential of the add-on's live incarnation, which only that process holds and which ends with it. A missing, wrong or old credential is refused with no_live_operation (401), and so is the credential of another add-on, because each add-on has its own listener.
• The call has no actor and no feature check. The operation must find its authority in the host's own records. For connector.inbound that is the stored binding: it must name the calling add-on and be switched on, or the call is refused with scope (403), also when the binding does not exist. The owner and the agent come from the binding. An argument that names a user, an actor or an owner is refused with bad_request before the operation runs.
• The add-on ID is the one the host built the listener for, never a value of the call. The host setting must still name the operation prefix for the add-on (telegram=connector. by default).

Every other operation keeps all its checks. An operation cannot become inbound by a field of the request: the scope is part of its registration.

Connector add-ons

A connector add-on carries the messages of a messaging platform between the platform and the chats. The Telegram and Slack add-ons are first-party connector add-ons. The host keeps what is the same for every platform (the bindings, the sealed bot token, the allowlist, the sender log, one chat for each sender or thread, the turn and its notifications), so an add-on holds only the transport and never stores a token. The first-party add-ons share their Go route (addons/internal/connector) and their binding page (addons/internal/connector-ui).

The manifest declares the capability. It needs no tool and no service flag:

"capabilities": {
  "connector": {
    "kinds": [{
      "id": "telegram",
      "title": "Telegram",
      "description": "Talk to one of your saved agents from a Telegram bot.",
      "senderHint": "A chat ID, a user ID or an @username",
      "delivery": "pull",
      "page": "bots",
      "drafts": true,
      "previewMs": 400
    }]
  },
  "ui": { "...": "the page that adds and edits the bindings" }
}

Field	Meaning and bounds
kinds	1 to 4 kinds. A kind is what a binding stores in kind, so it names the platform.
kinds[].id	A lower-case identifier, unique in the manifest. A binding names the add-on too, so another add-on cannot take over the bindings of a kind.
kinds[].title, description, senderHint	Plain text for people, with the bounds of the UI texts.
kinds[].delivery	pull: the host asks for updates (long polling). push: the add-on calls the host operation connector.inbound.
kinds[].page	Optional. The add-on page that adds and edits bindings; it must be a declared page. The Connectors page links to it.
kinds[].drafts	Optional. true when the add-on serves the op draft. The host then shows the live answer of a direct conversation as a draft instead of a working message.
kinds[].previewMs	Optional. The shortest time between two previews (edits or drafts) of one binding, 250 to 10000 ms. Without it the host waits one second.

Outbound: the connector route. The host calls POST /connector on the add-on with the same bearer credential and incarnation header as every other route and a body {"op", "kind", "token", ...}. The ops are poll (offset, timeoutSeconds up to 25), send (chat, thread, text, and optionally messageId), edit (chat, messageId, text), typing (chat, thread, and optionally messageId), send_file (chat, thread, file: {"name", "mime", "data"}), me (who the bot is, for the token check on save and the Test button), media (media: one reference of a message's media; the answer is {"file": {"name", "mime", "data"}} with at most 1 MiB of data) and, for a kind that declares drafts, draft (chat, thread, draftId, text).

• A send with a messageId replaces that message of the bot (the host's working message) with the first part of the text and sends the rest as new messages. When the message cannot be edited (it was deleted, or is too old), the first part goes as a new message too. The shared route (addons/internal/connector) does this for every platform.
• The text of a send is the answer, in Markdown. The add-on converts it to the markup of its platform and cuts it; the shared route lets a platform send each part in its markup and fall back to plain text by itself (Telegram does). The text of an edit or a draft is a preview: it goes as plain text.
• A draft shows text as a draft: a message that the bot is still writing, which the platform replaces when the answer is sent. draftId is not zero, and the host keeps one for the drafts of one answer. When the platform refuses it, the host sends a working message instead.
• An edit and a draft are not retried after a wait: a rate_limited answer comes back at once, and the host skips that preview. A send and a send_file wait once, as before.
• A typing with a messageId says that the host already shows a working message with that ID. A platform without a typing indicator (Slack) then posts nothing.
• send_file carries one file of 1 byte to 1 MiB whose name is contract.SafeFileName of itself (no path, no control characters, at most 128 bytes). The route refuses any other file, and a platform that cannot send files. The answer names the message (Telegram) or the file (Slack) in messageId. The token is the unsealed bot token of one binding. It is passed on each call and the add-on must not write it anywhere. The answer is {"updates", "nextOffset", "messageId", "botId", "botUsername", "file"} or {"error": {"code", "retryAfter"}} with a code of unauthorized, conflict, rate_limited, too_large or failed. The host maps unauthorized to switching the binding off, repeated conflict to the same, and rate_limited to a wait. too_large is the answer of a media whose file is larger than 1 MiB or than the platform lets a bot download; the host tells the sender that the file is too large. A runtime that implements contract.ConnectorRuntime has the route; others get a 404. The add-on cuts a long text to the limit of its platform.

Inbound. A pull add-on answers poll with normalized messages (cursor, chat, thread, direct, mention, messageId, replyTo, senderId, senderName, senderUsername, text, unsupported, media, skip) and the next offset. messageId is the platform's ID of the message and replyTo the message it answers; the thread conversation mode uses them. Core keeps the offset with the binding, passes it to the next poll and stores the offset the answer names, so a new leader continues where the old one stopped. A push add-on calls the host operation connector.inbound with {"bindingId", ...message}; it is an inbound operation. Both paths end in one function of the host, so the allowlist, the sender log and the rate limit are the same.

A message is answered when it is direct (a private chat) or a mention (it addresses the bot in a group, such as a Slack mention or a reply in a thread the bot follows); any other message is logged and ignored. Which chat a message belongs to depends on the conversation mode of the binding (see How the bot answers); by default a message with a thread belongs to a chat of that thread, shared by the people who write in it, and any other message belongs to a chat of its sender. When a message has media (at most 8 references), the host reads each file with the op media before it makes the chat, and attaches it to the chat, as an upload of the chat page would. The host makes the name of each file safe again (contract.SafeFileName), and the chat saves it in its own folder. An audio file (by its extension: .ogg, .oga, .opus, .mp3, .m4a, .wav, .flac, ...) is a voice message: the host transcribes it with the transcription model of the instance and the transcript, under [Voice message, transcribed], is the text of the turn; the recording is not kept. With no transcription model, or when the transcription fails, the sender gets one sentence and no turn starts. unsupported marks a message with no text and no file that the add-on hands over. The turn runs as the owner of the binding, with the tools of the binding's saved agent.

Where it runs. Core calls the route and registers the host operation on the process that owns the add-ons. The binding loops run there, under the leader lock in a cluster. A frontend that does not own the add-ons runs no loop.

Existing bindings. A binding of kind telegram that names no add-on (made before connectors became add-ons) belongs to the add-on telegram. While that add-on does not run, the binding is listed as not running with the reason "install the Telegram add-on".

Owner forwarding and HTTP relays

The embedding host can configure AddonConfig.Transport at startup. The configuration contains the owner HTTPS address, instance identities, CA roots, leaf certificate SHA-256 pins, and Ed25519 signing keys. The owner also needs ListenAddress for its separate mTLS listener. These settings are not runtime settings or browser inputs. Without transport configuration, the public addon gateway remains unavailable.

All participating frontends must use the authoritative named-user database. The owner queries current account, credential, and feature state before dispatch. Missing storage, paused keys, revoked keys, expired sessions, and legacy synthetic identities deny access. The embedding host's permissive installation callback cannot override these checks when transport is enabled in the default named-user mode. Local-owner HTTP access exists only when the deployment sets OwnershipMode: "local-owner". A request never selects it; only the deployment does, and local-ai run does so by default when authentication is off.

The carrier requires TLS 1.3, CA verification, and explicit certificate pins. It obtains a fresh random audience from the pinned owner before signing each request. Each owner handler generates its own audience at construction. Restart therefore invalidates all previous envelopes, including envelopes whose expiry has not passed. A bounded nonce cache prevents replay within that handler. This cache is a documented per-owner exception; it does not require shared storage because no other process can use that audience.

The signed payload binds the peer, actor, credential reference, operation ID, request body, audience, nonce, and expiry. When a chat made the call, it also binds the chat (chat: the user's principal and the chat ID), so the owner runs the add-on's workspace operations with that chat's access (see Workspace operations). A changed, added or removed chat fails the signature check. A request that no chat made has no chat field and signs as before. An owner that does not know the field refuses a forwarded call that a chat made, so upgrade the owner before the frontends. The carrier never forwards browser authorization, cookies, or actor headers. It has no proxy discovery, redirects, affinity, fan-out, owner election, or failover. An absent owner returns unavailable.

Two more optional capability fields describe the UI-only resource management (see UI-only resource management).

Field	Meaning and bounds
capabilities.undo	Boolean. The add-on keeps a change set for the writes a chat makes at once and answers the host's undo calls (see Undoing a change a chat made). It changes no tool and offers nothing to a model.
capabilities.adminTools	At most 4 tool names that only the host's own UI-only endpoints may run. An admin tool is never in tools, never has a tool definition, is never offered to a model and is refused by every model, proposal and typed-read path: the declared-tool rule of those paths reads tools only.
capabilities.connector	Optional. 1 to 4 kinds of messaging binding the add-on serves, each with an id, a title, a delivery (pull or push) and optional texts and a page. See Connector add-ons.
capabilities.resources	Maps the management operations to tools: create (a declared tool of kind create), rename (a declared tool of kind schema), inspect (a declared tool of kind read whose result is an object with a name and a revision), delete (the one admin tool) and a plain-text singular noun. Every field is optional. An admin tool must be the delete of the mapping, and a mapping must name tools that exist with the kind it needs. A mapping with create or delete must also declare the read tool resource-exists in tools, because the host settles interrupted creations and deletions only on its explicit answer.

A tool definition names the resource it works on with a required resourceId argument, with two exceptions. A create tool has no resourceId: the host reserves the resource. A read tool may leave resourceId out when it answers about the person who calls it and not about a resource, for example a list of that person's own items. The host runs such a read without a resource, binds the envelope to the calling user, and the add-on must scope its answer to that user. The host operation the add-on calls must scope by the actor of the envelope; the add-on never takes a user from the arguments. A write or schema tool, and a read tool that offers resourceId, must require it.

A tool definition with "uiOnly": true is for the add-on's own page. It must be a read tool without resourceId. The host runs it for the page like any such read, but never offers it to a model, a chat, an agent or a trigger, and a model that names it gets the unknown-tool answer. The Sandbox uses it for set-backend and confirm-host: they change a setting, and the add-on asks the host whether the caller is an administrator before it acts.

Tools that replace a built-in tool

A chat has a few built-in tools that an add-on may replace: bash, web_search and web_fetch (contract.ReplaceableTools). The file tools are not on the list: they are the one file toolset of every chat. A tool definition replaces one with "shadows": "<name>":

{ "name": "bash", "kind": "action", "shadows": "bash", "description": "...", "inputSchema": { "...": "..." } }

• Checked at install. The name must be on the list, a page-only tool cannot shadow, and one manifest shadows a name once. The tool must also look like the built-in it replaces, so the model calls it the same way and the user keeps the consent split of the built-in name:

  Name	Kind	Must declare (and may require only)
  bash	action	command (string)
  web_search	read	query (string)
  web_fetch	read	url (string)

  Extra optional properties are allowed, and a shadowing tool names no resourceId. The tool adapter checks all of this again when it lists the tools: a tool that fails keeps its add-on name and replaces nothing.

• In a chat. When the add-on is installed, running and on for the chat, the model sees one tool under the built-in's name, served by the add-on, and the built-in one is hidden. The call still goes to the add-on under its own name (addon__<id>__<tool>), with the same checks and audit. An add-on that is off for the chat, stopped, not granted by the chat's agent or unavailable offers nothing, so the built-in tool serves again. The chat's Tools menu says so in one line ("Sandbox replaces bash but cannot serve this chat now ...: the built-in bash serves", or that the chat has no such tool when the built-in is off too), from fallbacks of GET /api/chats/{id}/tools. A tool without shadows keeps its add-on name.

• Who served it. A call the add-on served carries the card tool_provider with the add-on's ID, so the chat shows the provider on the tool line.

• Several add-ons. Only one add-on serves a name. An administrator picks it in Settings, Tools (runtime setting chats_tool_providers, a map from the tool to the add-on ID; the same map holds the choices for the built-in skills an add-on replaces, see Skills); with no choice, the add-on installed first serves (ties go by add-on ID, and an add-on installed before the host kept the time sorts last). A chosen add-on that is installed but cannot serve the chat (stopped, off for the chat, not granted by its agent) gives the built-in tool, not another add-on; a choice of an add-on that is not installed is ignored. With no choice, the oldest add-on that is running serves. Settings > Tools lists every installed add-on that replaces a name, running or not, oldest install first, with the install time of each, marks a stopped one (stopped) and the default as Default: installed first, and says "The add-on that replaces this tool is not running: the built-in tool serves" when none serves (stopped and note in GET /api/chats/admin/tools). The library entry and the installed view list the names an add-on replaces in shadows, and the install dialog warns when an installed add-on already replaces one: "This add-on replaces the built-in bash. Another add-on (Sandbox) already does. Only one can be active; you choose which in Settings > Tools. Until you choose, the one installed first serves." The install is still allowed.

• Install time. The owner records when an add-on was first installed (installedAt, RFC 3339 in UTC) in its lifecycle journal, next to the other install facts. An upgrade keeps it; removing the add-on and installing it again sets a new one. GET /api/addons/apps and the installed view of an administrator carry it, and More details shows it. An add-on installed before the host kept the time has none. Distributed state: shared. The time lives in the owner's journal, the one durable record of the install, and the frontends read it from the owner, so every frontend picks the same default. A choice in chats_tool_providers always wins over it.

• Other consumers. The OpenAI-compatible API, MCP and agents keep the add-on name; only chats act on shadows.

Public API

• POST /api/addons/dispatch accepts a contract request. Supported lifecycle actions are install, enable, disable, and uninstall. operation-status is read-only.
• POST /api/addons/relay accepts {request, kind, handle}. Kinds are status, progress, json, file, and asset. Handles are not filesystem paths. For file the handle is the identifier of a private file result (see Private file results), never an addon output handle.
• Both routes require named authentication and the addons feature. Lifecycle dispatch additionally requires current administration rights. Cookie requests require same-origin fetch metadata; header-authenticated clients use their validated keys.
• POST /api/addons/proposals, GET /api/addons/proposals/{id}, POST /api/addons/proposals/{id}/confirm, .../cancel, .../reconcile and .../resolve store, read, apply, withdraw and resolve data or schema proposals. See Proposals and confirmation.
• GET /api/addons/resources/{resourceId}/proposals?state=unresolved lists, for the owner role of the resource only, the proposals on it whose outcome is unresolved, whoever proposed them. See Unknown outcomes and recovery.
• GET /api/addons/artifacts/{id} and GET /api/addons/artifacts/{id}/download read the state of, and download, a private file result of the caller. See Private file results.
• GET /api/addons/apps, GET /api/addons/apps/{id}/icon?generation=N, POST /api/addons/apps/{id}/start and .../stop list and control the addons as apps. See Apps.
• GET/POST /api/addons/resources, .../resources/{resourceId}/grants, .../grants/{userId}, .../users, .../rename, .../delete, .../changes/{changeId}/undo, .../operations/{operationId} and /api/addons/conversations/{conversationId}/selection/{addonId} list, create, share, rename and delete a person's resources and keep the resource a conversation works on. They are UI-only. See UI-only resource management.

Every failure response has the body {"error":{"type","message"}}. A denial or unavailability has type addon_error and a code. An outcome_unknown response is status 503 and also carries the original operationId. Keep that operationId after an outcome_unknown response. Query operation-status; do not create another mutation. The carrier never retries mutations. Manager.OperationStatus binds the lookup to the original actor and current administrator. It does not submit, repair, start processes, or write the journal. completed describes durable lifecycle publication, not runtime health. runtimeState: not-observed makes that distinction explicit. Ambiguous legacy actor bindings deny lookup.

Progress is coarse lifecycle-state SSE, not byte counts or percentage completion. The stream sends an initial snapshot, changed snapshots, and a heartbeat snapshot. It polls no faster than 250 milliseconds and rechecks authority on each poll. It closes on a completed or failed operation. A lifecycle writer can delay observation; disconnecting still cancels the waiting lookup. Install-byte progress requires later integration.

RelayLimits has finite startup defaults: 60 seconds, 16 concurrent streams, 64 MiB for files, 8 MiB for assets, and 1 MiB for progress. Heartbeats default to five seconds. Overrides cannot remove these bounds. Streams cancel on disconnect and reject excess bytes. The receiving frontend applies the budget of the relay kind: file uses the file limit, asset the asset limit, progress and status the progress limit, and json at most 2 MiB. An unknown kind is denied on the frontend before the owner is contacted. Only content type, attachment disposition, cache policy, and nosniff headers cross the carrier. Browsers never receive an addon process address.

A denial that the owner makes before it writes the first response byte (an invalid kind, missing admission or resolver, a refused resource, or a revoked credential) is returned as a typed owner status: 403 for a denial and 503 for unavailability. The receiving frontend decodes it, so every frontend reports the same error for the same decision. Only an authorization, policy or request rejection is 403. A runtime that is not installed, enabled or serving, an owner that is busy or in recovery, an output that ends before its first byte, and any unclassified owner error are 503, never a denial. A 503 does not prove that nothing ran: the owner can fail after it recorded or applied a mutation. A frontend therefore reports a mutation that gets a 503, or a transport failure, as outcome_unknown with the original operationId, and never as plain unavailability. The owner's own frontend reports it the same way as every other frontend. When an error carries both a denial and an unavailable cause, 503 wins. After the first byte the status cannot change; the owner then aborts the stream and the frontend reports the stream as failed, never as a clean end.

Resource integration remains closed by default

Admission.Admit encloses the complete private-output stream lifetime. A host-owned resolver maps resource references and opaque output handles to trusted records. A missing admission provider or resolver denies access, including access by installation administrators. The owner installs the policy described below as both providers at bootstrap; a frontend that forwards to the owner, or a host without the policy, keeps denying. File results are held by the private artifact store described in Private file results; without it a read that returns a file output is denied whole and a file relay is denied.

Principals, roles and grants

core/services/addons/policy is the owner-local authority for resource access. Its store is a private directory (policy/ below the addon root, file mode 0600, one holder at a time) that is separate from replaceable packages, so it can be backed up and restored on its own. A resource is an opaque host-issued ID bound to the trusted addon identity and an immutable owner user ID; the addon, the model and a request never supply either.

Role	May
Viewer	read, and download private outputs
Editor	everything a Viewer may, and mutate rows
Owner	everything an Editor may, and change schema, share and delete

Grants cover whole resources, never tables or rows. Installation administration confers no access to user data: the policy has no administrator role. Only the owner grants or revokes, the owner role cannot be granted, and an action that is not classified is a mutation that needs the owner. A missing resource and a resource the caller may not touch return the same denial.

Named-user mode uses the immutable user ID and rejects anonymous and legacy synthetic principals (such as the legacy shared API key). Explicit local-owner mode disables sharing and warns that every unauthenticated caller shares one authority: any local-owner caller owns every resource. A populated store opened in the other mode is refused. policy.MigrateMode changes it only while the host is stopped (the store lock must be free), after it wrote a backup, and shared data can never return to local-owner mode.

Account and credential status are checked again at each admission. A storage outage of the account database is 503, while a missing, disabled or unauthorized account is 403, from every frontend. Only a definite denial (an error that wraps the shared forbidden sentinel) is 403. Any other error from the account or creation check, and a cancelled check, is unclassified and fails closed as 503; it is never reported as a refusal of a valid user. The policy keeps no current-user state: the actor is an argument of every call, so shared MCP sessions carry none. Writes to one resource run one at a time; reads run concurrently. A revocation (or lowering a role) is durable first, then invalidates the user's pending proposals through the ProposalInvalidator seam, cancels the user's admitted work and returns only when that work ended. If it does not end within the drain bound the call reports 503 (ErrDrainPending): the revocation is already effective for new work but is not acknowledged. The owed drain is stored with the change itself, together with the resource's authority epoch at the lowering. Every admission carries the epoch it was admitted under. Repeating the same SetGrant or Revoke after a failed attempt (or after a restart) invalidates proposals again and waits only for the user's work admitted before that lowering; it is acknowledged only after both finished, and the mark is then cleared. Work admitted after the grant became durable again is neither cancelled nor waited on by the retry. A later lowering starts a new epoch and covers it. Disabling or uninstalling an addon keeps its resources and grants; purge is a separate, explicit operation that is not available to tools.

A new resource needs explicit creation authority. At bootstrap that authority requires a current account check and an addon that is installed, enabled and not changing on the owner; an absent addon is denied and a manager that cannot answer is 503. Register records the resource pending and inaccessible before the addon creates anything; Commit activates it, and Reconcile resolves registrations left by a crash. The owner runs Reconcile at startup before it serves. It asks the add-on whether the backing exists: yes activates the record, no drops it. Reconcile is fail-soft per record. When the answer cannot be established (the add-on is disabled, failed to start or does not answer), that record stays pending and hidden, is never dropped or committed, keeps its pending quota, and the pass goes on with the other records. Startup never aborts for this, and the owner logs a warning with the settled and kept counts. A background loop re-runs Reconcile every 30 seconds while any record is pending, so a record is settled once its add-on is enabled and ready. The loop stops when the owner closes and waits for a running pass. Pending reservations of a disabled add-on therefore stay pending until the add-on is enabled again. A database with no record is never reachable.

Registrations are bounded. Per owner user at most 8 pending and 256 total resources, and at most 4096 in the store (Options.MaxPendingPerUser, MaxResourcesPerUser, MaxResources; zero selects these finite defaults). A request over a limit is denied with 403 (ErrQuota); abandoning or reconciling away a registration frees its slot.

The store file is validated on open: an unknown role, a grant to the owner, a second JSON document after the first, or an unknown field refuses the store, which is never reset. Leftover .policy-* temporary files from a crash are removed on open. If a change was renamed into place but the directory synchronization failed, the in-memory state follows the file (a revocation is not undone) and the store accepts no further writes, reporting 503, until the host restarts.

Explicit local-owner mode

Set --addons-ownership-mode=local-owner (or AddonConfig.OwnershipMode in an embedding host) to select this mode. local-ai run also selects it by default when authentication is off and you set no mode. Nothing else selects it: a missing credential, an absent account database or an unknown value never does, and an unknown value fails startup. If you turn authentication on later, set the mode to named-user and move the data with policy.MigrateMode while stopped; the store refuses the other mode. The owner logs a warning at startup that unauthenticated callers share one authority and that any caller that can reach the port can confirm addon proposals.

• Every caller that passes the addon route chain acts as one synthetic principal, local-owner, with a fixed credential local-owner. Named-user mode rejects that ID, so it cannot collide with a real user.
• That principal owns every resource and sharing is disabled: SetGrant and Revoke are denied.
• With the authentication feature disabled (no account database) there is no caller identity at all: any client that can reach the HTTP port shares the authority, including lifecycle administration. Use this only on a trusted loopback or private network.
• With an account database, the route chain still authenticates callers and applies the addons feature gate, and lifecycle changes still need an administrator; only data ownership collapses to the single local owner.
• The store records the mode. Opening a populated store in the other mode is refused; use policy.MigrateMode while stopped. Shared data cannot return to this mode.

Moving between the modes

The mode is a property of the stored data, so it is changed with the host stopped and never by changing the configuration alone:

1. Stop the owner. The migration needs the store lock, which a running owner holds.
2. Make a backup of the policy store. The migration writes one itself and refuses to continue when it cannot.
3. Name the mapping of identities. From local-owner to named users, every resource needs the account that takes it ({"local-owner": "<account id>"}; there is no wildcard and no default, and the target must be a current account). From named users to local-owner, a store that holds named-user data is refused, because that data would lose its separation.
4. Start the owner in the new mode. A store opened in the wrong mode is refused at startup, so a configuration mistake never silently changes who owns data.

A backup made in one mode restores into the other mode only with the same explicit mapping (see Data recovery).

Resource reads obtain the current ready runtime incarnation and check the exact generation. Private output copying rechecks credentials and stays inside admission until completion or cancellation. Immutable assets use installed package receipts, declared entry/CSS paths, and a contained filesystem root; an account that holds the addons feature may read the declared files of an enabled add-on it may see (see The asset route).

Mutation MCP tools remain deliberately unavailable. The later MCP integration defines safe listing/status tools for proposals; prompt text cannot authorize mutation and no tool confirms a proposal. For that reason /api/addons/dispatch, /api/addons/relay and the proposal routes are deliberately not mapped in the LocalAI Assistant tool catalog (pkg/mcp/localaitools) yet. The route map test only requires an entry for each registered tool, so no test is weakened. The routes are described in the generated Swagger document under the Addons tag.

Run sh scripts/addons/verify.sh policy for the policy suite.

Proposals and confirmation

Every data or schema mutation that comes from an API client, an MCP session or the OpenAI-compatible chat route is a stored proposal first. A proposal is applied only by an explicit confirmation of that stored operation from the person's own browser session. The chats on the server, their agents and their triggers are the exception described under Lifecycle: there the add-on switch is the consent. Sharing and database deletion are never proposals and never tools.

Declared tools

The owner checks every request that names a tool against the manifest of the installed current generation before anything reaches the add-on. The tool name is the top-level tool key of the request arguments. The rule is the same for a read, a proposal (write or schema), the confirmation of a stored proposal and the opening of an output, on the owner's own frontend and on every forwarding frontend:

• No tool key: the request names no tool, and the rule has nothing to check.
• An exact tool key: the value must be a string that equals one entry of capabilities.tools byte for byte. A name of another add-on, a case, Unicode or white-space variant, an empty value and a value that is not a string are denied.
• A key that equals tool only after case folding or trimming white space (Tool, TOOL, tool) is denied, because an add-on decoder may read it as the tool.
• A repeated key is denied at any depth of arguments. Keys are compared after JSON unescaping, white-space trimming and Unicode case folding, so "tool" and "\u0074ool" collide, and so do "tool" and "TOOL". A runtime that keeps the first value would otherwise run another tool than the one that was checked.
• arguments must be exactly one JSON object. An array, a scalar, null, trailing data after the object, invalid JSON and nesting deeper than 64 levels are denied. This rule runs in the declared-tool check itself, so it holds for reads, proposals, confirmations and output opening whatever validation ran earlier.
• A denial is the standard 403, and the add-on is not contacted. A proposal is denied before the preview and nothing is stored. A stored proposal is checked again at confirmation, so a tool that an update withdrew is never applied.
• Only a real answer is a denial: the tool is not declared, or the generation is stale or not installed. When the declarations cannot be read, the request is unavailable (503) and nothing runs: the manager is closed or not the owner, the journal is uncertain or runtime recovery is required, the request context is done, or the source fails for any other reason. An unclassified failure is unavailable, never success and never a policy denial. The owner and every forwarding frontend give the same status. The owner reads declarations from its in-memory journal, cached for each generation.

Lifecycle

Tools reach a model through the same proposals: a chat that sets metadata.addon_tools can call the tools of enabled addons, and a change a tool proposes waits as a pending proposal. The chats on the server (/api/chats), their agents and their triggers are the one exception: turning an add-on on for a chat or an agent is the consent, so a change tool is applied at once as the chat's user. The chats adapter stores the proposal and confirms it in the same call with the credential of a chat turn, so the same admission, revision and receipt rules hold, a failed call leaves no pending proposal, and no API key, MCP session or browser request can use that credential. The result says what was applied and whether it can be undone: Tables, Fine-tuning and Quantization keep a change set for the changes they can roll back, and the result carries undoable, the changeId and a one-line undo. See Undoing a change a chat made. GET /api/addons/proposals?state=pending lists the caller's own pending proposals (newest first, at most 32). See Tables for the tool names, the approval card and the inventory of tool consumers.

POST /api/addons/proposals takes a typed write or schema request. When the arguments name a tool, the owner derives the action from the kind the manifest gives that tool (write tools are data writes, schema tools are schema changes) and refuses a request whose action disagrees, a read or create tool, and a tool the manifest does not define. The owner resolves the resource, requires write authority (schema needs the owner role), asks the addon for a preview and stores a proposal with:

• the canonical arguments, the addon generation, the resource and the proposing user;
• the schema and data revisions the preview saw;
• the exact number of effects and a bounded sample of them (the addon's preview summary, using count and sample when it has them);
• a digest of everything above (it also covers the operation ID, the proposing user and the action; the owner recomputes it from the stored record before it applies, so a record that was changed at rest is refused and never runs), and an expiry (the shorter of the addon's preview expiry and the proposal expiry limit, five minutes by default).

Nothing is applied by this call. The request's operationId becomes the proposal ID, so every later step, including a lost reply, is resolved by the same ID. Repeating the same request returns the same proposal. Reusing an ID with other arguments, another actor or another resource is a denial.

A proposal is in exactly one state:

State	Meaning
pending	Waiting for a human decision.
cancelled	Withdrawn by its actor, or invalidated when the user's authority on the resource was lowered or revoked.
expired	Not confirmed in time. The owner records it when the clock runs out, on the next proposal, on start and on recovery, whether or not anyone read it.
executing	Confirmed and handed to the addon. Durable before the addon sees the call.
applied	Committed. The receipt is stored with the proposal.
rejected	Refused by the addon (for example, the revision moved) or found stale (the addon generation changed).
outcome-unknown	The apply may or may not have run. A receipt lookup, or the resource owner's explicit acknowledgement, resolves it.
abandoned-unknown	The resource owner stopped waiting for an unresolved outcome. It no longer blocks writes and is never applied again; a later receipt lookup is still recorded.

Bounds. A user holds at most 32 live proposals (pending ones that have not expired, and unresolved ones) and at most 16 MiB of stored canonical arguments across them; the store holds at most 1024 proposals and 64 MiB of arguments. The user's own bound is checked first, so one user can exhaust only their own share. A limit reached is a typed quota denial (HTTP 403) that names which bound it was. To make room for a new proposal the owner drops settled proposals only (cancelled, expired, applied, rejected), the proposer's own first and then the oldest; a live proposal of another user is never dropped. Settled proposals are also dropped after a retention window of 24 hours by default (BrokerOptions.Retention), so a repeated request inside the window still gets its stored answer. Retention cannot be switched off.

Every transition is written atomically and fsynced by the owner before it is reported. The store lives beside the policy store and follows the same rules: one process holds it, a corrupt file is refused and never reset, and a restart recovers it.

Creating a resource

A request with action: "create" and no resourceId proposes the creation of a new resource, such as a Tables database. Nothing exists yet, so the owner works in this order:

1. It checks the creation authority of the caller and that the named tool is declared.
2. It reserves a pending resource for the proposer. A reservation is invisible to everyone, the proposer included, counts against the pending limit of the user and is bound to the proposer. The proposal records its ID as resourceId.
3. It asks the addon for a preview with the reserved ID. The preview only validates the request and creates nothing. What the human sees (count 1 and the requested name) is built by the owner and not taken from the addon.

Confirming runs the creation under the creation authority of the proposer, which is checked again, and not under a role on a resource. The addon operation ID is create- and the reserved ID, so a repeat is one operation. A committed receipt makes the proposer the private owner of the resource. Cancelling, expiry and a rejected receipt drop the reservation; the resource was never visible.

An uncertain apply keeps the reservation pending and the proposal outcome-unknown. A receipt lookup settles both. The proposer can also stop waiting (resolve): the reservation is dropped only when the addon's resource-exists tool confirms that the database does not exist, and it stays pending when the addon says it exists or cannot answer. At start the owner settles reservations of unknown or executing proposals by asking the addon, and keeps those of proposals that are still pending. A reconciliation never drops the reservation of a creation that is running in this process.

Confirm

POST /api/addons/proposals/{id}/confirm with {"digest": "..."} submits exactly the stored parameters under resource admission, once. The runtime validates the revision in the same transaction as the write, so a data change after the preview is a rejected proposal and not a silent overwrite. Confirming again returns the stored receipt and applies nothing.

The route is reachable only from a host confirmation surface:

1. The credential must be a browser session cookie of the user who proposed. A named API key, a bearer token, a legacy key or an MCP session can propose and read, and can never confirm, whatever its role.
2. The request must be a same-origin fetch (Sec-Fetch-Site: same-origin, Sec-Fetch-Dest: empty, Sec-Fetch-Mode cors or same-origin).
3. The X-LocalAI-Confirm header must equal the proposal ID, and the body must carry the digest of what was displayed.
4. A request that carries a tool, agent or MCP execution marker (Mcp-Session-Id, X-LocalAI-Tool-Call, X-LocalAI-Agent, X-LocalAI-Execution-Context) is refused.

In local-owner mode there is no session; rules 2 to 4 apply, and any API credential on the request is refused. The owner repeats the credential check, so a frontend cannot confirm for a key credential either.

Limit of the confirmation boundary in local-owner mode. Without accounts there is no identity to bind a confirmation to. The only boundary is the set of browser headers above, and a script that sets them can confirm: any caller that can reach the HTTP port shares the authority to propose and to confirm, and that includes the proposals it made itself. This is inherent to local-owner mode, where unauthenticated callers share one authority, and it is accepted and not changed. The owner logs a warning at startup that says so. Use local-owner mode only on a trusted loopback or private network, and use named-user mode wherever a confirmation must come from a particular person. Plain text, a boolean in arguments, or a model saying "yes" are only data and never approve. Clients without a confirmation UI can propose and read; they never auto-apply. No MCP tool maps to confirm: the MCP exposure of proposals (list and status only) is a later task and must keep that rule.

Unknown outcomes and recovery

Every proposal call is idempotent by the proposal ID: a proposal is a single-use state machine, and repeating a call with the same ID returns the stored state and never starts a second operation. The owner therefore answers an outage by what it knows about the proposal, and every frontend reports the same answer for the same owner state:

Response	Meaning	What to do
503 proposal_unavailable with operationId	The owner could not serve this call, and this call sent nothing to the addon. The proposal is in the state the owner recorded, which can be any state: pending, or already settled by an earlier call.	Read the proposal with GET, then repeat the same call with the same ID.
503 outcome_unknown with operationId	The request to the addon may have reached it (executing or outcome-unknown). A request that never left the owner (no free slot, a withdrawn add-on, an invalid envelope, a connection that was never written to) returns the proposal to pending and is proposal_unavailable. An add-on's 409 conflict ends the proposal as rejected.	Do not confirm again as a new operation. Reconcile, wait for the automatic lookups, or resolve.

Which type a route can emit does not depend on who proposed the operation. A proposal, a confirmation and a cancellation by the proposer can be either type. A resolution is always proposal_unavailable, because resolving never applies anything and the resolver is the resource owner, who is not always the proposer. A confirmation that the addon committed answers 200 with the applied proposal; the owner does not read the record back through a second account check that could fail after the apply.

A reply lost between a forwarding frontend and the owner for a confirmation is reported as outcome_unknown, because the frontend cannot know whether the owner started it. The lifecycle dispatch rule is unchanged: any uncertain lifecycle mutation is an unknown outcome with its original operation ID.

If the reply to a confirmation is lost, or the owner restarts while a proposal is executing, the proposal is outcome-unknown and the response is outcome_unknown with the original operationId. The mutation is never submitted again under any ID.

POST /api/addons/proposals/{id}/reconcile looks the operation's receipt up on the addon and records a committed or rejected result. It is read-only toward the addon and can be repeated. An addon that has no receipt for the operation stays outcome-unknown: a missing receipt does not prove the call never ran. While an outcome is unresolved, every other write or schema proposal on the same resource is denied; a pending proposal shows blocked: true; reads continue. A write denied for this reason is 403. For the owner role of the resource only, the body is type outcome_blocked and carries blockedBy, the operation ID to resolve. Any other user gets the plain denial and never learns another user's operation ID, also not in a proposal view. On start the owner settles proposals left executing by a crash the same way and never applies them.

Recovery has three steps and none of them submits the operation again:

1. Automatic lookups. The owner repeats the receipt lookup for each unresolved proposal with exponential backoff (from the BackoffInitial limit up to BackoffMax). They run for a finite window: the apply timeout plus five minutes, counted from the moment the apply was handed to the addon. A final receipt settles the proposal and lifts the block. After the window, only an explicit lookup remains.
2. Explicit lookup. reconcile, by the proposer, at any time.
3. Owner resolution. The owner role first finds what to resolve: GET /api/addons/resources/{resourceId}/proposals?state=unresolved lists every unresolved or abandoned proposal on the resource, whoever proposed it, with its ID, state, proposing user, action, exact effect count and bounded effect sample, expiry, digest and age in seconds, oldest first. It never returns the proposed arguments or a credential. Only the owner role of that resource may call it, from the same confirmation surface as resolve with X-LocalAI-Confirm equal to the resource ID; editors, installation administrators and API keys are denied, and no MCP tool maps to it. Then POST /api/addons/proposals/{id}/resolve with {"digest": "...", "reason": "..."} (the digest from the list, required and checked by the owner itself, not only by the HTTP route) lets the owner role of the resource stop waiting. It is accepted only from the same confirmation surface as confirm (the owner's own browser session, a same-origin fetch, X-LocalAI-Confirm equal to the proposal ID, no tool or agent marker). The proposer's role is not enough, an installation administrator without an owner role on the resource has no access, and no API key or MCP tool can call it. A final receipt found at that moment settles the proposal instead. Otherwise the proposal becomes abandoned-unknown and records who acknowledged it (resolution.actorId), when and the reason (up to 512 characters). Writes on the resource are accepted again.

An abandoned operation keeps its original operation ID. The owner keeps looking its receipt up in the background for a day and on every reconcile; if the addon reports that it ran, the proposal becomes applied with that receipt (or rejected), and the resolution stays on the record, so a late success is recorded and never hidden. An abandoned record is kept for 30 days and is not dropped to make room.

Data consistency. Resolving is the owner asserting that the outcome is not worth waiting for; it does not prove the change did not run. If it did run, it stays in the data. Later proposals preview the current revision, and the runtime refuses an apply whose revision moved, so a change prepared against the state before the unknown one is rejected and not applied on a stale view. Check the data before you resolve when that matters.

Revocation races

Lowering or removing a user's grant cancels that user's pending proposals on the resource before the change is acknowledged, and stops their admitted confirmations. A confirmation that already reached the addon becomes outcome-unknown and is resolved by receipt. Other frontends reach the proposal state only through the owner's gateway, signed and actor-bound like every other forwarded call.

Run sh scripts/addons/verify.sh proposals for the proposal suites.

UI-only resource management

A person manages their own resources, for example Tables databases, from the host's own pages. The routes below are the whole surface. They are never tools: no model tool, no MCP tool, no proposal and no typed dispatch reaches any of them, and sharing and deletion have no tool definition anywhere. The routes live in the addons feature group with the usual browser protections (CSRF, same-origin cookie requests), and they are registered in the route-feature registry.

Route	Who	What
GET /api/addons/resources?addonId=	any caller with the feature (keys too)	The resources of the add-on the caller owns or that were shared with the caller, with the caller's own role, the display name, the owner's display name for a shared one and, for the owner role only, how many people it is shared with and the revision. At most 100. The names are read at most half of the add-on's concurrency budget at a time, so a page that reads while it lists is not refused.
POST /api/addons/resources	the person's own session	Create a resource through the add-on's create tool. Caller becomes the owner; the resource is private.
GET /api/addons/resources/{resourceId}/grants	owner role, own session	Who it is shared with: immutable account ID, display name, role.
PUT /api/addons/resources/{resourceId}/grants/{userId}	owner role, own session	Give a named user viewer or editor on the whole resource.
DELETE /api/addons/resources/{resourceId}/grants/{userId}	owner role, own session	Remove the grant.
GET /api/addons/resources/{resourceId}/users?q=&limit=	owner role, own session	The user picker.
POST /api/addons/resources/{resourceId}/rename	owner role, own session	Rename through the add-on.
POST /api/addons/resources/{resourceId}/delete	owner role, own session	Delete the resource and its data.
POST /api/addons/resources/{resourceId}/apply	role of the tool, own session	One data or schema change through a declared write or schema tool, with no card.
POST /api/addons/resources/{resourceId}/changes/{changeId}/undo	the role of the change, own session	Undo one change a chat made at once, by the change set the add-on kept.
GET /api/addons/resources/{resourceId}/operations/{operationId}?kind=	owner role (rename, delete) or a role that may write (apply, undo), own session	Read-only receipt lookup of a rename, a deletion, a change or an undo.
GET/PUT/DELETE /api/addons/conversations/{conversationId}/selection/{addonId}	any caller with the feature	The resource a conversation works on.

"Own session" means the same rules as a confirmation: a browser session cookie, a same-origin fetch, X-LocalAI-Confirm equal to the resource ID (the operation ID for a creation) and no tool, agent or MCP marker. An API key, a bearer token and a legacy key are refused, whatever their role. In local-owner mode the rules are the same without a session. The grants, the picker and the receipt lookup are owner-role reads of other people's data and use the same surface.

Every call goes to the owner through the carrier, so a frontend that holds no addon state gives the owner's own answer. The state is the owner's alone and is single-runner by design: the policy store with its deleting records, the proposals and the selection store (selections/ below the addon root) are never read or written by a frontend, and a frontend that cannot reach the owner reports a mutation as outcome_unknown with the original operation ID. The owner checks the caller's role itself, with the same denial for an unknown resource and for one that is not the caller's. Installation administration is not data access: an administrator lists and changes what they own and nothing else.

Creating, renaming and deleting

These are the person's own explicit actions, so they need no approval card, but they follow every other rule of a mutation.

• Create. The owner reserves a pending resource, runs the add-on's create tool through the same preview, apply and receipt steps as a chat creation, and activates it. The resource ID is derived from the person and the operationId, so repeating the call returns the same resource and a repeat with other arguments is refused. The arguments are validated against the tool definition of the manifest exactly as a model's call is, and a crash is settled at the next start like any pending reservation. The per-person limits apply (quota_reached).

• Rename. The call runs under the resource's admission, behind an unresolved outcome (outcome_blocked for the owner role), against the generation and the revision the owner listed (stale, nothing changed). The add-on's receipt for the operation is the record: repeating an operationId replays it, and two people that chose the same ID never share one.

• Delete. The order is durable at every step:

  1. the resource becomes deleting: inaccessible to everyone, no new admission and no grant change, and admitted work is cancelled and waited for (bounded);
  2. with no work left, the owner checks the revision it listed and the confirmation: the request must carry confirm, equal to the display name exactly. A failure returns the resource to use;
  3. the deletion is marked confirmed, pending proposals of every user on the resource are cancelled, and the add-on is asked to delete under one operation identifier;
  4. only after the add-on confirmed, the record and every grant are removed.

  A crash leaves a deleting record. At the next start an unconfirmed one is returned to use, and a confirmed one is finished; a reconcile loop retries one that cannot be finished while the add-on is down. The resource stays inaccessible meanwhile. Deleting never uninstalls the add-on and never touches other resources. An outcome that cannot be established is outcome_unknown with the original operationId: repeat the same call, or look the receipt up; nothing is submitted under a new operation.

  Before it calls the delete tool, the owner asks the add-on's resource-exists tool whether the backing is still there. Only that explicit answer of the declared tool can say "gone", and then the record is removed without the delete tool. An add-on that is not installed, does not declare the tool or cannot answer never counts as "gone": the deletion stays deleting, is reported as outcome_unknown and is retried by the reconcile loop, so the data is never left behind a removed record. This applies to the request, to a repeat of it and to the recovery at start.

The delete tool is declared in capabilities.adminTools. The owner is its only caller, from the delete endpoint and its own crash recovery, under the host action delete, which no model tool, proposal or typed request can carry. The add-on refuses the tool under any other action.

Changing data and schema from the UI

POST /api/addons/resources/{resourceId}/apply is how a person's own page changes rows or the schema, for example the Tables page. The body is {operationId, generation, revision, tool, arguments}. It is not a proposal: nothing is stored for later and no card is shown, because the request in the person's own tab is the intent. It keeps every rule of a mutation:

• Tool. tool must be declared in capabilities.tools with a definition of kind write or schema. Reads, creations, the add-on's mapped rename, create and delete tools, admin tools and undeclared names are 400. The arguments are the tool's own, without the resource (the path names it) and without tool; they are validated against the definition exactly as a model's call is, so an identity, role or approval field or an unknown field never reaches the add-on.
• Role. A write tool needs at least the editor role; a schema tool needs the owner role. A viewer, an editor changing the schema, a person with no grant and an installation administrator get the answer of an unknown resource.
• Admission and revision. The change is admitted per resource, waits behind an unresolved outcome, and is checked against the generation and the revision the caller read (409 stale, nothing changed). The add-on's receipt is the record: repeating an operationId with the same request returns the same result and never applies twice; the same ID with another request is refused as 409 operation_conflict before the add-on is called, with nothing applied (never outcome_unknown). The host binds each operation ID to a request hash (requestHash: SHA-256 of the resource, the tool and the canonical arguments, per user), returns it with the result, and keeps it for a bounded number of recent operations. An add-on that refuses the reuse with 409 on /apply gets the same answer.
• Result. The answer carries the new revision and the receipt's result (for Tables, the count and the changed row ids). An outcome that cannot be established is 503 outcome_unknown with the original operationId; look it up with kind=apply and never submit a new operation. The lookup answers state plus the tool and requestHash the ID was used for. Pass requestHash=<hash of your request> to ask about your own request: a hash that differs from the recorded one is 409 operation_conflict, and a hash the host does not hold (for example after a restart) is state: unknown, which is safe to answer by repeating the same request. A client that compares requestHash before it believes committed never mistakes another change for its own.

Undoing a change a chat made

A write, schema or create tool called from a chat, an agent or a trigger is applied at once (there is no card), so undo is the safety net. An add-on that declares capabilities.undo keeps a change set for the changes it can roll back, in its own storage, and says so in the receipt of the change: the receipt payload has an undo object, {"id", "available", "summary", "reason"}. The host copies it into the tool result as undoable, changeId, undo (what an undo does, one line) and reason (why nothing was kept), and the chat shows an Undo button on the entry. The model is told that the person can undo it and that it cannot.

POST /api/addons/resources/{resourceId}/changes/{changeId}/undo takes {operationId, generation}. The change ID is the one the tool result named. It runs under these rules:

• Humans only. It accepts only the person's own browser session, with the rules of every UI-only mutation: a same-origin fetch, X-LocalAI-Confirm equal to the resource ID, and no API key, bearer token, tool, agent or MCP marker. This is deliberate. It is never a tool and no MCP tool maps to it, so a model cannot undo its own work or anyone else's without the person.
• Role. The role of the undo is the role of the change, which the add-on reports: write changes need the editor role, schema changes and creations the owner role. Other callers get the answer of an unknown resource.
• Stale. The undo is applied against the revision the add-on has just reported. If the resource moved on after the change, the answer is 409 with type stale and a message that says what changed; nothing is overwritten.
• Typed refusals. not_found (404: never kept, too large to keep, or dropped), expired (410: past the 7 days of retention), already_undone (409), cannot_undo (409: for example the job it started has finished) and stale (409). The type of the error is the code. outcome_unknown and outcome_blocked behave as for every mutation.
• Idempotent. The same operationId with the same change is one undo and returns the same result; after outcome_unknown look the receipt up with kind=undo.
• Result. {resourceId, changeId, operationId, revision, summary, deleted}. deleted is true when the change was the creation of a resource: the host deletes it with its own delete, and only while the add-on says nothing touched it since.

For an add-on author: declare "undo": true in capabilities, add the key undo to the receipt payload of every change, and answer two calls that are never in capabilities.tools, so no model can name them. The read tool changes-get takes {change} and answers {id, class, state, summary, detail, current, deletes} with class write or schema and state available, stale, expired, undone, cannot or not_found. The apply with the action undo and the tool changes-undo takes {change} and the revision current, and answers a committed receipt, or a rejected receipt whose payload is {"error": code, "message"}. The add-on checks the action and the scope itself. Tables, Fine-tuning and Quantization implement it; Fine-tuning and Quantization undo only the start of a job, by stopping it while it runs.

In a cluster the undo travels the same carrier as every resource call: a frontend forwards it to the owner, which calls the add-on, and the typed refusal crosses the carrier as the same typed error. The change sets are state of the add-on, in its own storage on the owner (for Tables inside each database file), so they are covered by the add-on's backups, restores and purges and need no extra shared state on the frontends; the carrier and the owner keep none.

An SDK-based page does not call the route itself. sdk.resources.apply (see the SDK) sends it for the add-on and generation the SDK is bound to.

Sharing

Grants are whole-resource and viewer or editor; the owner role cannot be granted. Named-user mode only: local-owner mode answers 409 with type sharing_disabled, and anonymous and legacy principals are never grantees. The grantee is named by the immutable account ID that the picker returned, and must be an active named account that holds the addons feature. The picker finds such accounts by a prefix of the display name or of the address (2 to 64 characters, at most 20 results) and returns only an ID and a display name, never an address; the caller and the owner are left out, and the wildcard characters of a query mean themselves. Lowering or removing a grant cancels the user's pending proposals and stops their admitted work before the call is acknowledged; a drain that does not finish in the bound is 503 and the change is already effective for new work.

The resource of a conversation

The selection is a convenience for chat, not an authority. The owner stores an opaque resource ID keyed by the authenticated user, the conversation ID and the add-on, in its own bounded store: at most 256 entries per user (the least recently used is replaced), 8192 in total, expiring after 90 days without use. It is not process-global, not shared between people or conversations, not part of the chat history and not sent to the model.

• PUT requires at least the viewer role on the resource now.
• A tool call of a chat request that names no resourceId uses the conversation's selection. The conversation ID comes from the authenticated request (the X-LocalAI-Conversation-Id header the chat sends), never from the model's arguments, and the owner authorizes the reference for the caller on every call. A reference that was revoked, whose resource was deleted, that was edited in the store or that belongs to another person gets the denial of an unknown resource: the result of kind error with code resource_unreachable, from the owner and from every frontend. The reference stays and works again if the grant does.
• With no selection the tool call returns a result of kind error and code select_resource: no operation runs and the model is told to ask the person to choose.
• A refusal the model can act on is a result of kind error, never a raw error string. The codes are invalid_request (the add-on refused the call; a message carries its short reason when it gave one, such as a missing training engine), not_permitted (the person lacks a permission), resource_unreachable (no such resource, or no access; a read and a change are denied the same way, and the result does not tell a missing resource from one the person may not use) and not_running (the add-on is starting or stopped; nothing was sent). Each note says that nothing changed and not to repeat the call. A preview that finds a conflict, such as a table that exists, is an invalid_request, not an outage. The tools of the fine-tuning and quantization add-ons are not offered at all to a person without those permissions.
• An explicit resourceId in the arguments is the person's own word and is used as before.

Errors

Status	error.type	Meaning
400	invalid_request	A missing or wrong confirmation, an operation the add-on does not support, arguments or a name it refuses, a malformed identifier.
403	addon_error	Not the person's own session, or the feature or role is denied: one answer for all of them.
403	not_permitted	The add-on checked a host permission of the person, such as fine-tuning, and it is not granted. Nothing changed.
403	resource_unreachable	The resource does not exist or the person has no access to it: one answer for both, for a read and for a change.
403	package_unverified	An install failed because the package does not match its library entry or is not a valid add-on package. Nothing was installed.
403	library_unavailable, library_conflict, release_not_found, package_fetch_failed, install_storage_failed	An install or upgrade failed at the step the type names. Nothing was installed. The server log has the detail.
403	quota_reached	The per-person resource limits.
403	outcome_blocked	Another operation on the resource has an unresolved outcome (blockedBy is shown to the owner role only).
404	not_selected	No resource is selected, or the reference is no longer usable.
409	stale	The generation or the revision the caller saw is out of date; nothing changed.
409	sharing_disabled	Local-owner mode has no named users to share with.
503	outcome_unknown	A mutation may have run. operationId is the original; repeat the same call.
503	addon_error	The owner or the add-on cannot answer.

A request that the add-on itself refuses as invalid (an unknown table, an unknown field, a value its schema rejects) is 400 with error.type invalid_request on every route that reaches the add-on: the typed apply, a read through dispatch or relay, a proposal and its confirmation, an artifact open, and the chat tool result (kind error, code invalid_request). The owner's own frontend and every forwarding frontend give the same answer. The add-on answers HTTP 400 or 422 and the runtime client types that answer as a refusal; any other non-success stays an outage (503). A refusal is deterministic and nothing was applied: it is never outcome_unknown, the same call is refused again, and the caller corrects the request instead of repeating it. An operation ID whose preview or apply was refused is not consumed: the host forgets its binding and the add-on removes the pending line it wrote for it, so the corrected request may reuse the same ID. A refused proposal is not stored. A confirmation the add-on refused ends the proposal as rejected.

A selection call has no operation: a lost answer is plain unavailability and the call is safe to repeat.

Private file results

An addon read can return file outputs, such as a CSV export. A file result is a private, temporary artifact of the user who asked for it. It is not a shared file: the model, media, upload and agent file routes and the frontend code artifacts never see these bytes, and the existing file APIs are unchanged.

What the host does with an output

1. The owner never returns the addon's output handle to a client. When a read result has file outputs, the owner reserves one artifact for each (all of them, or none) and replaces the handle in the result with the artifact ID: an opaque, random, host-generated identifier. An addon handle, URL or path is never an identifier and never names a file. The handle must be a valid ID of the contract; a URL or path in its place is refused and nothing is read.
2. The artifact records the requesting principal, the source resources (the request's resource, as the policy recorded it), the addon and its generation, the operation, the sanitized name, the media type, the declared size and the expiry. The addon handle is kept only in memory while the copy runs; only a digest of it, with the principal and the operation, is stored so that a repeated request finds the same artifact.
3. A bounded background worker copies the bytes from the registered runtime endpoint of the addon (and from nowhere else) into a private staged area under artifacts/ in the addon root: directories 0700, files 0600. The worker streams with one 32 KiB buffer and never holds the whole file. It stops at the declared size: a stream that is short, long by one byte, cancelled or failing is never published, its partial bytes are removed and the artifact is recorded as failed with a fixed reason. The bytes are published only by an atomic rename after the stream ended cleanly.
4. The copy runs under an output admission of every source resource, so a revocation cancels it and waits for it, like any other admitted work. At completion the owner checks again that the account is current and that the principal can read every source resource, that the addon is installed, enabled and at the same generation, and that the artifact is not expired. Only then is it recorded as complete (metadata is written atomically with fsync, validated on load and never reset).

Another user's artifact, an unknown ID, an expired one and a malformed one are the same 403, so identifiers cannot be probed. Read access to the source resource does not give access to another user's artifact, and sharing a database never shares its exports.

Routes

• GET /api/addons/artifacts/{id} returns the artifact view: state (transferring, complete, failed, cancelled), sanitized name, media type, size, sources, addon, generation and expiry. It is the status lookup of the background transfer. It is read-only, actor-bound and never shows a path or an addon handle.
• GET /api/addons/artifacts/{id}/download streams a complete artifact. The receiving frontend relays it from the owner through the file relay, so a download works from every frontend and no frontend holds the bytes. It needs the session or key of the requesting principal and the addons feature. Before the first byte and while streaming the owner checks the account and read access to all source resources again; a revocation cancels the stream. A failure after the first byte aborts the connection, so a truncated file never looks complete (the response carries the exact Content-Length).

No MCP tool maps to these routes in this release. A tool that returns a file result is added with the Tables export; it will return the artifact reference, not a URL.

Response safety

Every download is an attachment (RFC 6266, with an ASCII filename and a UTF-8 filename*) with a sanitized name, X-Content-Type-Options: nosniff, Cache-Control: private, no-store, a Content-Security-Policy that sandboxes the response and Content-Length. The name loses path elements, control characters, bidirectional overrides and invisible characters, shell and file system metacharacters and leading or trailing dots, is prefixed when it is a reserved device name and is at most 100 characters. The media type is one of text/csv, text/plain, application/json, application/x-ndjson, application/zip, application/pdf, image/png, image/jpeg or application/octet-stream; everything else (HTML, SVG, XML, JavaScript) is served as application/octet-stream, and parameters are dropped. A browser request with a session cookie must be a same-origin fetch or navigation like the other addon routes. Spreadsheet formula handling of a CSV is the addon's concern; the host only delivers the bytes as a download.

Limits

All limits are finite and validated at startup; a partial or unbounded override is refused.

Limit	Default
One artifact	64 MiB (at most the relay file bound; a file result larger than the configured relay file bound is refused whole at ingest, so nothing is staged that cannot be downloaded)
Declared bytes of one user / of all users	1 GiB / 4 GiB
Artifact records of one user / of all users	100 / 1000
Concurrent transfers (all / one user)	4 / 2
Concurrent downloads	16
Lifetime from creation	1 hour
One transfer	5 minutes

Quotas are enforced under one lock when the artifact is reserved, so concurrent requests cannot exceed them; a refusal reserves nothing. A quota refusal is 403, a busy store is 503. A repeated request (same principal, operation and output) returns the same artifact and uses no more quota.

Lifecycle

• Expiry. A sweep (every minute) removes artifacts at their expiry, transfers that no worker owns, artifacts of an uninstalled addon and files that no record names; opening the store also settles interrupted transfers and removes unreferenced files. An expired artifact is never served, even if the sweep did not run yet or the expiry falls while a download is being admitted or streamed. A download that meets a sweep which removes the artifact at that moment gets the same 403 as any expired or unknown identifier, never a corruption error. The sweep removes the bytes only of records it really changed: a transfer that completes while the sweep runs keeps its bytes.
• Storage errors. Only bytes that are definitely missing, not a regular file or of another size than recorded mark the artifact failed and remove the bytes (503). A transient storage error (for example too many open files or an I/O error) is 503 and leaves the bytes and the record as they are, so a later download can succeed.
• Disable. A disabled addon denies downloads and accepts no new file result. The complete bytes are kept until their normal expiry. Enabling it again serves an artifact only while it is unexpired and only to a principal that can still read every source resource.
• Uninstall. An uninstalled addon denies downloads at once and its artifacts are removed by the next sweep (they are not kept until expiry, because nothing can answer for them). An addon update to another generation does not serve the older build's files.
• Restart. Complete artifacts survive a restart. A transfer that was running is recorded as failed and its partial bytes are removed.
• Close. Closing the owner cancels running transfers and streams, waits for them and writes nothing afterwards.

Run sh scripts/addons/verify.sh artifacts for the artifact suites.

Apps

Enabled addons behave like apps: a list of what is available, with an icon, a title, a state and, when the addon has a page to open, an Open action.

Who sees what

GET /api/addons/apps needs named authentication and the addons feature. The owner builds the answer from its own journal, so every frontend gives the same answer:

• A caller sees every addon that is installed and desired-enabled. Version one has no per-addon access list; the data an addon holds stays under its own per-resource ACLs, which every call checks again. The apps a caller sees are the addons the same caller could call. Hiding an entry is never the control.
• An administrator also sees installed addons that are disabled (state stopped, enabled: false), so that a stopped addon can be started. The owner decides who is an administrator from its own current account check, never from the request. If that check cannot answer, the caller gets the reader's view.
• A headless addon is listed with its state and tools and has no mainPage, icon or title beyond its identifier.

Each entry has id, name (the same as id), title, description (the default-language text; titles and descriptions carry every language when the manifest localizes them), icon, version, generation, enabled, state, mainPage, hasUI, launchMode, tools (at most 32 names; the manifest has no tool descriptions) and toolCount. Administrators also get admin with canStart, canStop and managementUrl. Nothing of the package location, its source or the addon's own endpoint is returned.

state is an observation of the process, separate from enabled:

State	Meaning
running	The process is alive and admitting calls.
starting	A launch is between exec and readiness, at first start or restart.
failed	The process is not serving and no launch is under way. The supervisor restarts it after its backoff while its budget lasts, so failed can be followed by starting; when the budget is spent it stays failed until an administrator stops and starts the addon.
stopped	The addon is disabled, or no process of this generation is held.

An app can also carry update: "update-pending". It is a separate observation from state and is present only when the owner's cached update check says an update is available for exactly the installed package. The list answers from that cache: it never evaluates a check and never reaches a library, so an add-on nobody checked, one whose check failed and one that is up to date have no update field. See Update detection and upgrade.

mainPage is /app/addons/{id}/{page} and is present only for an enabled addon whose manifest declares a ui.mainPage that is one of its pages. icon is /api/addons/apps/{id}/icon?generation=N. The route is served by the host from the verified package, relayed from the owner, with X-Content-Type-Options: nosniff and Cache-Control: private, no-store. It is never an addon process port and never a URL the manifest names. A stale generation, an unknown or hidden app and an app without a verified icon give the same 403.

Start and stop

POST /api/addons/apps/{id}/start and .../stop take {"operationId": "...", "generation": N}. They are thin routes over the typed enable and disable lifecycle actions: the same administrator check, the same owner forwarding, the same operation identity and the same outcome-unknown rule. Stop drains the process and keeps data.

• Only administrators can start or stop. Everyone else gets 403 whatever the body says, before the body is read.
• generation is the generation the caller saw in the list. If the installed generation differs, the request is denied (403) before any operation is recorded.
• A repeated operationId replays; it never runs twice. After an outcome_unknown response, keep the operationId and poll operation-status with it.
• Health stays a separate observation: a successful start records intent and starts the process; the list's state says whether it is running.

The typed enable and disable actions of POST /api/addons/dispatch accept the same generation precondition through the request's generation field when it is not zero.

No MCP tool maps to the Apps routes in this release. The admin mutation tools for addons are a later task and will have their own review.

Built-in add-ons

GET /api/addons/builtin lists the page add-ons that ship inside the web bundle, so the My add-ons page can show them next to the installed add-ons. The answer is {"builtin": [...]}, and each entry has an id, a kind (always page), a titleKey, an icon and a mainPath, plus a feature or adminOnly when the entry has a gate. Today the list is empty: Studio and Talk, which used to be built in, are packaged add-ons now (see The Studio add-on and The Talk add-on), so a person installs and removes them like the others. The route is kept for the day a page is built in again. It needs an authenticated caller and does not need the addons feature. When authentication is off, the caller is the local owner and sees every entry. The answer depends on the caller and is sent with Cache-Control: private, no-store.

Pins

Pinned add-ons are listed in the sidebar under Add-ons, in pin order. The menu of a pinned card in My add-ons has Move earlier and Move later.

GET /api/addons/pins returns the caller's pinned add-on IDs as a JSON array, in order. PUT /api/addons/pins takes a JSON array of add-on IDs, replaces the caller's pins with it and returns the saved list. Both routes need an authenticated caller. They need no administrator role and no addons feature, like GET /api/addons/builtin. A PUT from a browser must be a same-origin request or carry a credential header. The answer depends on the caller and is sent with Cache-Control: private, no-store.

• Filtering. A pin is filtered like a tile. The list holds only IDs the caller may see now: the built-in add-ons that pass their gates (none today), plus the installed apps the owner lists for the caller. An ID the caller cannot see is left out of a read and dropped from a save, with no hint whether it exists. Duplicates are dropped.
• Hidden pins survive. A pin that is only hidden for now (an access change, or an add-on removed inside its retention window) is kept after the visible pins when the caller saves.
• Limit. A save of more than 32 IDs, or a body that is not a list of strings, is refused with 400. If the add-on owner does not answer, a save is refused with 503, so an outage cannot drop pins.
• Retention. When an add-on is removed, its pins are hidden at once and kept for 30 days. If the add-on is installed again inside that window, the pins come back. After 30 days the pins are deleted. A background sweep runs every six hours; with several frontends, one of them runs it at a time. A disabled add-on is not removed, so its pins never expire. A sweep that cannot read the add-on owner changes nothing.
• Upgrade seed. A read or a save never pins anything. Once, at the upgrade that adds this feature, the accounts that exist then get Studio and Talk pinned. Both are add-ons, so a pin is drawn once the add-on is installed. A seeded pin does not expire while its add-on was never installed, however long that takes; once the add-on was installed, its pin follows the retention rule like any other (the sweep records the first install in the table of seed runs, addon_pin_seed_runs, or in the pins file). An account created later starts with no pins. An unpin stays after a restart.
• Startup. If the pins file cannot be read (it is corrupt), the server does not start and logs the error. It does not fall back to memory, because that would lose every unpin at the next restart.
• Distributed state: shared. With an account database, the table addon_pins is shared by every frontend. With authentication off there is one local owner, and the pins are kept in addon-pins.json in the data path, per instance.

builtin and pins are reserved add-on IDs: an add-on cannot use them, because the host serves /api/addons/builtin and /api/addons/pins.

Add-ons page and management

The web UI serves one page for add-ons at /app/addons, called Add-ons. It is shown to a caller who holds the addons feature, and to anyone who can open a built-in tile (none today). Studio and Talk are add-ons: they are listed only while they are installed. The first tab is My add-ons. Administrators also reach Find add-ons and Sources. The tabs are a convenience: the owner checks administration again on every call, so a request that skips the page is denied the same way (403).

The sidebar has the items Chat, Models and Add-ons. Add-ons opens the page at /app/addons. The add-ons a user shows in the sidebar are listed under it, in order, as indented links with the add-on icon and title; nothing is listed when none is shown, and the list is hidden when the sidebar is collapsed. The user adds an add-on to the sidebar with Show in sidebar in the menu of its card, and changes the order with Move earlier and Move later in the same menu. Pins belong to the user, follow the user across browsers, and are capped at eight. A pin whose add-on is stopped opens the detail view; a pin whose add-on was removed or is hidden is not drawn. Operators who need the previous sidebar can set LOCALAI_LEGACY_NAVIGATION=true (flag --legacy-navigation); it also brings back the Add-ons hub, a one-tab bar over /app/addons that the legacy sidebar lists. The switch is a rollback aid and is removed one release after the new navigation ships; the UI reads it from the addon_navigation key of /api/features.

A tab is addressed as /app/addons?tab=browse|libraries; My add-ons has no tab. What the former Installed tab showed is the More details section of an add-on in My add-ons, so ?tab=installed opens My add-ons. The management link of an app (/app/addons?manage=ID) opens My add-ons with the details of that add-on open.

Old links

The generators that had their own page now open inside the Studio add-on, and the Studio and Talk pages open their add-ons. The old links still work and keep the model, the query string and the hash:

Old link	Opens
/app/studio	/app/addons/studio/studio
/app/studio/<tab>	/app/addons/studio/studio/<tab>
/app/talk	/app/addons/talk/talk
/app/image	/app/addons/studio/studio/images
/app/video	/app/addons/studio/studio/video
/app/3d	/app/addons/studio/studio/threed
/app/tts	/app/addons/studio/studio/tts
/app/sound	/app/addons/studio/studio/sound
/app/diarization	/app/addons/studio/studio/diarization
/app/transform	/app/addons/studio/studio/transform
/app/face	/app/addons/face/recognition
/app/face/<model>	/app/addons/face/recognition?model=<model>
/app/voice	/app/addons/voice/voice/recognition
/app/voice/<model>	/app/addons/voice/voice/recognition?model=<model>
/app/voice-library	/app/addons/voice/voice/library
/app/voice-library/new	/app/addons/voice/voice/library/new

None of these links has a core page behind it: the face, voice, Studio and Talk pages exist only in their add-ons. Without the add-on, a link opens the add-on page, which says the add-on is not available. A model id can hold / or :, which an add-on path segment refuses, so the model travels as ?model= (for example /app/voice/org%2Fname opens /app/addons/voice/voice/recognition?model=org%2Fname; the face links and the Studio links do the same).

A link with a model (/app/image/<model>) opens /app/addons/studio/studio/images?model=<model>. The /app/studio?tab= links redirect to /app/addons/studio/studio/<tab>. A viewer without the 3d, audio_diarization or audio_transform feature still lands on the app home, as before.

The legacy switch (LOCALAI_LEGACY_NAVIGATION) restores the old sidebar and the Add-ons hub only. These redirects do not depend on it: with the switch on, the old generator links also open the Studio add-on. The legacy sidebar lists Home, Models, All chats, Agents, MCP connectors, Memory and the hubs; Studio and Talk are in My add-ons, as add-ons.

The Studio add-on

The Studio add-on (studio) is a packaged add-on. It is not shipped with LocalAI: install it from Find add-ons with Install and turn on. Once it is usable, it appears in My add-ons and can be shown in the sidebar. Until then the sidebar, My add-ons and the Home /studio command do not list it, and the old Studio links open the add-on page, which says the add-on is not available.

Page	Path	Who sees it
Studio (main page)	/app/addons/studio/studio	users with the addons feature

The tab is the one path segment after the page: /app/addons/studio/studio/images, .../video, .../threed, .../tts, .../sound, .../transform and .../diarization. The overview is the page without a tab. A model travels as ?model=<id> and a result as ?work=<id>. A tab whose feature is off for the user (3d, audio_diarization, audio_transform) is not listed.

The pages call core routes with the signed-in user's own session (see "Core routes from add-ons"): /v1/images/generations, /v1/images/upscale, /video, /3d/*, /tts, /v1/audio/speech, /v1/sound-generation, /audio/transformations, /v1/audio/diarization, /api/models/*, /api/resources and /api/operations. The core keeps these routes and checks every call again, so the page never widens what a user may do. The history stays in the browser, under the same keys as before the move, so what a person made earlier is still there when the add-on is installed. The words are the media locale files of the web UI, in every language the web UI has.

The Talk add-on

The Talk add-on (talk) is a packaged add-on, separate from Studio. Install it from Find add-ons with Install and turn on. It adds the page /app/addons/talk/talk (main page) for users with the addons feature. The page opens a WebRTC session to a realtime pipeline through /v1/realtime/calls and reads the pipeline list from /api/pipeline-models, with the signed-in user's own session; the core keeps both routes. The page uses the MCP servers that the person added in the browser, the same list as Chat. The words are the English talk locale file of the web UI; other languages read the English text.

The Voice add-on

The Voice add-on (voice) is a packaged add-on. It is not shipped with LocalAI: install it from Find add-ons with Install and turn on. Once it is usable, it appears in My add-ons and can be shown in the sidebar.

The add-on has one page, /app/addons/voice/voice, and one entry in My add-ons. Its sections are tabs of that page, and the tab is the next segment of the address:

Tab	Path	Who sees it
Recognition (default)	/app/addons/voice/voice/recognition	users with the voice_recognition feature; the tab shows an off notice for anyone else
Library	/app/addons/voice/voice/library (the form to add a voice is .../library/new)	administrators
Text to speech	/app/addons/voice/voice/speech	users with the addons feature; the server refuses the speech call without audio_speech
From a recording	/app/addons/voice/voice/recording	users with the addons feature

The earlier page addresses (/app/addons/voice/recognition, .../library, .../create and .../speech) still open: each one redirects to its tab and keeps the query string (?model=, ?selected=, ?voice=).

All of these also need the addons feature to see the tile. The pages call core routes with the signed-in user's own session (see "Core routes from add-ons"): /v1/voice/* for recognition (verify, analyze, embed, register, identify, forget), /api/voice-profiles for the library and /v1/audio/speech for the speech page. The core keeps the voice registry and these routes; the server checks every call again, so an add-on page never widens what a user may do.

The Recognition tab is the workbench: who is this, same person, the people this browser enrolled, and the analyze and embedding tools under More tools. Library manages the reference recordings that text-to-speech can use and is shown to administrators only. Text to speech speaks text with a voice model. From a recording links to the diarization workspace of Studio. The Face add-on shares the same views, built from one source directory into each add-on.

The Face add-on

The Face add-on (face) is a packaged add-on, separate from the Voice add-on. It is not shipped with LocalAI: install it from Find add-ons with Install and turn on. Once it is usable, it appears in My add-ons and can be shown in the sidebar.

Page	Path	Who sees it
Recognition (main page)	/app/addons/face/recognition	users with the face_recognition feature; the page shows an off notice for anyone else

The tile needs the addons feature. The page calls core routes with the signed-in user's own session (see "Core routes from add-ons"): /v1/face/* (verify, analyze, embed, register, identify, forget) and the model list. The core keeps these routes and the face registry; the server checks every call again, so the page never widens what a user may do.

The page holds the face recognition workbench and the management of the known faces: the people this browser enrolled, the enrol form and the removal behind Undo. It has no administrator-only part, so every user who holds face_recognition can use all of it. The detect-and-analyze and embedding tools are under More tools. The Voice add-on shares the same views, built from one source directory into each add-on.

The Fine-tuning and Quantization add-ons

The Fine-tuning add-on (fine-tuning) and the Quantization add-on (quantization) are packaged add-ons. They are not shipped with LocalAI: install them from Find add-ons with Install and turn on. Once usable, they appear in My add-ons and can be shown in the sidebar.

Add-on	Page	Path	Who sees it
fine-tuning	Fine-Tune	/app/addons/fine-tuning/jobs	users with the fine_tuning feature; the page shows an off notice for anyone else
quantization	Quantize	/app/addons/quantization/jobs	users with the quantization feature; the page shows an off notice for anyone else

The old links /app/fine-tune and /app/quantize open the add-on page. Without the add-on, they show the add-on page that says the add-on is not available. The pages call /api/fine-tuning/* and /api/quantization/* with the signed-in user's own session (see "Core routes from add-ons"); the core keeps these routes and checks every call again. Both add-ons offer chat tools that start a job. In the chats a tool call starts it at once; on other surfaces it is a proposal that waits for a person to confirm. They reach the engine through Host operations, and a tool can control only a job that a tool started. See Fine-tuning and Quantization.

My add-ons

My add-ons is a grid of cards with a search box. The built-in cards (none today) come first, then the installed add-ons by identifier. A card shows the icon, the name, the state (On, Off, or the state the host observed, such as Starting or Failed), the description, Adds a page when the add-on has a page and the number of chat tools when it has tools. A running add-on with a page also has an Open link. Each card has one overflow menu with Show in sidebar (or Remove from sidebar), Details, and, for an administrator, Remove. A pinned add-on also has Move earlier and Move later in the menu. Showing a ninth add-on in the sidebar opens a dialog that asks the user to remove one first. Selecting a card, or Details, opens the preview pane beside the list (/app/addons?detail=ID) and never the main page, because the main page needs a live process. The pane shows the name and state, the description, screenshots only when the host lists images for the add-on (it lists none today, so no slot is drawn), what the add-on adds (its page and its chat tools), Turn on or Turn off, Remove and the version and source. For an administrator it also has More details: health (with the install time), trust, compatibility, storage, data recovery and updates. Administrators also see Find add-ons. A caller without the addons feature sees only the built-in cards (none today), and the page does not call GET /api/addons/apps for that caller.

With nothing installed, an administrator sees three add-ons from the sources, each with Install and turn on, and a link to Find add-ons. A search that matches nothing says that no add-on matches it.

The pane also lists the pages the add-on names in ui.navigation. The page draws GET /api/addons/apps: the icon from the host's icon route, the title in the interface language (falling back to the default text and then to the identifier), the description, a state chip (running, starting, failed, stopped), an update pending chip when the owner's cached check reports one, an Open link only when the entry has a mainPage, and a tool summary. Icons always come from the same origin; the page never loads an image from a manifest or a library.

The list is read again every 5 seconds (30 seconds in a hidden tab). A failed read doubles the wait up to 60 seconds, keeps the last list on screen and says it may be out of date. Leaving the page cancels the pending read. Administrators see Turn on and Turn off on an entry that carries the admin object. They send a fresh operation ID with the generation that was listed and keep that ID for retries of that one action: once the operation is final (completed, failed or refused) the ID is dropped, so the next click is a new operation and is never answered from an old one. After outcome_unknown the page reads the original operation with operation-status and never submits a second one. If the reads end without a final answer, the add-on shows Check status (one more read of the same operation) and Dismiss, and no other action on that add-on is sent until one of them is used. After a completed status the page reads the list again and says so when the add-on did not change, because the owner answers a replayed operation with its old status and applies nothing. A 403 (a stale generation or no permission) and a 503 are shown with the same words everywhere on the page, and the list is read again.

The Open link points at /app/addons/{id}/{page}, the page of the add-on described in Add-on pages below.

Management actions

All of these go through POST /api/addons/dispatch, are administrator-only on the owner, and take strict arguments: an unknown field is a denial. The reads change nothing. The writes keep the operation identity and outcome-unknown rule of the lifecycle actions.

Action	Kind	Arguments	Payload
catalog-search	read	text, kind (addon by default, backend or all), sourceId, offset, limit (at most 100)	{releases, total}; only releases for the owner's configured target platform are listed, and total counts that set
catalog-details	read	request addonId; sourceId, version	{release, installed}; installed is null or {version, generation, enabled}
catalog-sources	read	none	{sources}; each source has refresh once the owner fetched it
addons-installed	read	none	{addons}, the installed view below
catalog-source-save	write	one source: id, name, url, enabled, policy, authReference; removed: true deletes it	the stored source
catalog-source-refresh	write	sourceId	{sourceId, count, checkedAt, lastSuccess}
upgrade	write	sourceId, version	the operation status

A library carries a reference to a credential, never a secret. authReference must be the name of a credential the owner configured; any other value is denied and is not stored or shown back. The URL of a library cannot carry credentials or a query. A browser session cannot turn on the plaintext loopback opt-in for a library that does not already have it. Removing a library never touches installed addons: they stay installed, keep their data and keep running.

A refresh is a write because it makes the owner fetch a remote document. A library that is disabled, unreachable or invalid answers 503 with outcome_unknown and the original operation ID, like any other write whose result cannot be proved; the refresh object of catalog-sources says what the owner last learned (checkedAt, lastSuccess, count, failed).

Installed view

Each entry of addons-installed has id, title, description, icon, version, generation, enabled, state (the process observation, separate from enabled), sourceId, channel, selection, verification (sha256 or sha256+sigstore), digest, platform (os, arch), compatibility (the host, api, protocol and dataFormat ranges of the manifest), tools, toolCount, hasUI, mainPage, launchMode, generations (retained installed generations), restarts (supervisor restart attempts), storage (bytes and truncated for the retained data) and update.

The size of the retained data is read in small batches, never follows symbolic links, and stops after 20000 entries or about 0.75 seconds, whichever comes first; truncated is then true and bytes is a lower bound. The owner keeps the last result for 30 seconds, so a page that polls does not walk the tree again.

update is the owner's update check of the add-on: state, available, the candidate version and digest, checkedAt, lastSuccess, stale, reasons and a redacted error class. It is absent when the owner cannot answer for the add-on, and then the page says that the update status is unknown. The page never compares version labels itself and never reads a failed check as current. See Update detection and upgrade. Process output is not collected in this release, so the page has no log view; the state and restarts fields are the health evidence it shows.

Uninstalling stops the addon and removes its package from use. Its data, identity and grants are kept, so a later install of the same addon picks them up. The page says so before it submits the uninstall.

Data recovery

An add-on's data is the data the owner keeps for it under <root>/data/<id>, the owner's records of its resources and shares, and the proposals made against it. Recovery keeps that recoverable. It is administrator-only, it is never a tool, and the owner decides every step. Replacing a package alone never restores data: the previous package is kept, but it does not make the data it could not read readable again.

Backups

POST /api/addons/{id}/recovery/backup makes one backup. The owner blocks new admission for the add-on's resources, makes its pending proposals invalid, drains the work that was admitted (bounded), stops the add-on and copies its data: every SQLite file is written with VACUUM INTO, so the write-ahead log is part of the copy and the copy holds no log of its own. Scratch directories named exports and tmp are not copied, and symbolic links are never followed. The owner's access state (resources and shares), the proposals of the add-on and the conversation selections are copied beside it.

A backup is a directory <root>/backups/<id>/<backupId> of mode 0700 with files of mode 0600. Its manifest.json records the creation time, the reason (manual, upgrade or restore), the ownership mode, the package digest, trusted source, channel, verification level and data format, and the size, sha256 and, for a SQLite file, the schema version of every file. The backup is verified while it is still private: every digest is read again and every SQLite file passes PRAGMA integrity_check. Only then is it renamed into place and flushed. A backup that fails, is cancelled or does not fit leaves nothing behind, and the add-on is started again if it was meant to run.

The free space is checked before anything is copied: it must be at least the size of the data plus a margin of 16 MiB or a tenth of the data. A backup that does not fit answers 503 with the type insufficient_storage and changes nothing.

Retention is explicit: the owner keeps BackupKeep backups per add-on (default three) and removes older ones, and with BackupMaxAge also those older than that. The newest verified backup and the recovery copy of an upgrade are never removed by retention, and POST .../recovery/delete-backup refuses them with sole_recovery_copy: the only verified recovery copy is never removed silently. A purge with an explicit scope is the way to remove them. POST .../recovery/prune applies the policy now.

Upgrade of an add-on that has data

An upgrade of an add-on that has a data directory runs this order, and every step is recorded before it runs:

1. The trusted identity of the staged package is compared with the installed one: the same add-on from the same source and channel with the same verification level. A different one is refused before anything stops. The same rule holds for a reinstall of retained data. A package that declares an older data format than the data is refused.
2. Admission is blocked, proposals are invalidated, admitted work drains and the add-on stops.
3. A verified backup is made.
4. The package is published and the candidate generation is started while nothing else touches the data. That start is the migration under exclusive ownership: an add-on migrates its own data when it starts and writes its progress to migration.json in its data directory (state is running, complete or failed, with from, to and error). The start is the runtime's health check; a migration.json that does not say complete fails the upgrade. Tables does this with Store.Migrate(ctx, fromVersion, toVersion): each database changes in one transaction, the version of the store is recorded last, a run that stopped is resumed, and data in a newer format than the executable reads is refused and left as it is.
5. Only when the candidate is healthy is the generation published, with the journal ordering of a plain upgrade.

A failure at any step before the publication is undone. When the data was not touched, the old generation runs again. When the candidate began to migrate and the format is the same, the verified backup is put back before the old generation starts, and what the candidate left is kept as a recovery copy. When the migration changes the data format and fails, the add-on stays unavailable in the state needs-recovery: the previous package is retained and so is the verified backup, but the previous package cannot read data the new one began to change, so nothing starts it. The administrator restores the backup or purges the data. While the state lasts, the backup that was made for it is always kept: retention, a prune and a delete do not remove it, and a manual backup is refused, because a copy of half migrated data must not look like the newest good one. A restart finds the operation at its recorded phase and applies one rule: a generation that was not published is undone, and a published one is finished.

Add-ons without data keep the plain upgrade and start the candidate as before.

Restore

POST /api/addons/{id}/recovery/restore restores one backup. The add-on must not be running, and the request carries the add-on ID typed as confirm. The backup must belong to the add-on named in the route: a backup of another add-on is refused with invalid_request before anything is touched. The backup is verified before anything changes. Its files are staged beside the data and swapped in atomically; the data they replace stays as a recovery copy, named previous-<id>, until an administrator purges it. A crash at any point is settled by the owner on restart: before the swap the restore is undone, from the swap on it is finished.

The answer lists warnings: the time the data goes back to, how many changes the owner applied after the backup (their effects are gone), and the shares that were not restored. Restoring never gives back access:

• Every owner and every grant is checked again against the current accounts. An account that is gone, paused or pending is not restored.
• The owner keeps a durable log of lowered access (a removal or a lowering of a role), compacted per person and resource: a later, stricter lowering replaces the ones it covers, so frequent changes of other people never push a removal out. A grant that was removed after the backup is not restored, and one that was lowered is restored at the lower role. The cutoff is the moment the access state was read for the backup, taken under the same lock, and the add-on's admission hold also stops changes of its grants.
• The log has a hard size limit. When it must drop entries, it records the time up to which it may be incomplete. A restore from a backup older than that time restores the data and no access at all: no owner has to be trusted to be still valid, so the resources come back with their owner only and the answer says so (historyTruncated). The owner shares them again.
• Proposals that were pending are cancelled. Operations that were executing, or whose outcome was unknown, become outcome_unknown and are looked up by receipt of the restored add-on; they are never applied again.

A backup made in local-owner mode needs an explicit ownership mapping, {"local-owner": "<account id>"}, to be restored in named-user mode; there is no wildcard and no default, and the target must be a current account. Data of named users is refused in local-owner mode. The mode of a stopped installation can be changed with policy.MigrateOwnership: it writes a backup of the policy store first, takes the same mapping, and refuses to move populated named-user data to the local owner.

Repeating the operationId of a restore returns its receipt and does not restore again.

Purge of retained data

POST /api/addons/{id}/purge removes retained data. It is not the deletion of a database and not an uninstall, which never remove retained data. The add-on must be uninstalled or disabled and not running, the request carries an explicit scope (data, backups, recoveryCopies) and the add-on ID typed as confirm. The purge is recorded before it starts and finished by the owner after a crash. It touches no other add-on and removes the owner's records of the add-on's resources only when data is in the scope. Purging the data clears needs-recovery.

Routes and rules

Route	What it does
GET /api/addons/{id}/recovery	state, backups with their verification state, retention
POST /api/addons/{id}/recovery/backup	make a backup
POST /api/addons/{id}/recovery/verify	read a backup completely and record the result
POST /api/addons/{id}/recovery/restore	restore a backup
POST /api/addons/{id}/recovery/delete-backup	delete a backup that is not the sole recovery copy
POST /api/addons/{id}/recovery/prune	apply the retention policy now
POST /api/addons/{id}/purge	purge retained data

All of them need a named administrator. Every mutation is accepted only from the administrator's own browser session: a same-origin fetch whose X-LocalAI-Confirm header equals the operationId; an API key, a tool call or an MCP session is refused. Each is decided by the owner, so a forwarding frontend gives the same typed answer, and a lost reply is outcome_unknown with the original operationId. Typed refusals are needs_recovery, addon_running, busy, backup_not_found, backup_corrupt, incompatible_format, identity_changed, mapping_required, mapping_invalid, downgrade_unsupported, sole_recovery_copy, confirmation_mismatch, operation_conflict, invalid_request and insufficient_storage. None of these is a tool: no MCP tool, model tool, proposal or typed dispatch of a model maps to them, and the manifest of an add-on cannot declare them.

The Add-ons page shows the recovery state of an installed add-on: the backups with their size and verification state, a banner for needs-recovery with the actions restore and purge, the phase of an upgrade in progress and the warnings of the last restore.

Update detection and upgrade

The owner knows what is installed and which library and channel it came from, so the owner decides whether an update exists. A browser never compares versions and never chooses the library. Detection and upgrade are two different things: a check only reads, and an upgrade is one explicit administrator action that names the exact candidate.

States

GET /api/addons/{id}/update returns one of these states:

State	Meaning
unknown	No check has run for the installed package.
up-to-date	The library offers nothing newer for the selection the add-on was installed from.
update-available	A compatible candidate exists and differs from the installed package.
incompatible	A candidate exists that this host cannot run. reasons names platform, host-version, api-version, protocol-version, sdk-version or data-format.
check-failed	The owner could not tell. error is a class: source-unavailable, refresh-failed, credential-unavailable, not-listed, ambiguous-release, canceled or internal, and lastSuccess is when the library was last read. A failed check is never shown as up to date.
integrity-error	The library lists the installed immutable version with another digest. It is never an update.

The status also carries installedVersion, installedDigest, candidateVersion, candidateDigest, the sourceId, channel and selection the add-on came from, checkedAt, stale (the answer rests on data older than the time to live) and throttled (an explicit refresh was not run because one ran moments ago).

How the candidate is chosen

• A moving channel (development, a branch-* channel, latest) compares resolved digests. A different digest is an update even when the version label is the same.
• A development build installed from a mutable entry (a moving tag such as 0.0.0-development, with no digest in the library) is compared by the digest that its tag names. Its label is not compared as a version. An explicit check (Check for updates, Check again or the automatic upgrade) resolves the tag in the registry and keeps the digest, so a later status read compares against it and does not use the network. Before the first explicit check the state is unknown. When the tag names another build, the state is update-available with that digest as candidateDigest and the label with the first 12 hex digits of the digest as candidateVersion. A tag that cannot be resolved is check-failed with refresh-failed. The upgrade resolves the tag again and refuses with candidate_changed when the tag moved after the check.
• An add-on installed from a version label is immutable. The owner offers the highest greater version of the same major version, and never a pre-release to a stable install. The catalog entry of the installed version with another digest is an integrity-error: an immutable version was rewritten.
• The library and the channel never change by themselves. An entry of another channel or another library is never offered, whatever its version.
• A catalog entry may carry an optional requires object with host, api, protocol, dataFormat and sdk ranges ({"min": 1, "max": 2}). It is a hint for the check only: the manifest of the verified package is checked again when the candidate is staged. A candidate whose data format ends below the installed one is incompatible. The sdk range is only checked for a package that declares one, so a package without a UI is not refused for it.

Cache, refresh and credentials

The check and its cache live on the owner only, like the journal they read: a forwarding frontend asks the owner and never keeps an answer of its own. The cache is bounded for each add-on (default 256 entries, 15 minutes) and ends with the owner. A status read answers from it or evaluates the library data the owner already holds; it never uses the network. POST /api/addons/{id}/update/check fetches the library again through the credential reference the library is configured with, at most once per interval for an add-on and for a library (default 15 seconds; UpdateCacheTTL and UpdateRefreshInterval of the add-on configuration change the two defaults), and then returns the new status. A secret, an address or the text of a failure never appears in a status. Checks never apply an update.

Who sees what

Anyone who holds the addons feature and may see the add-on reads the status: an administrator any installed add-on, everyone else an enabled one. A caller who is not an administrator gets the state, the versions, the times and the reasons, without the library, channel, selection or digests. Checking and upgrading are administrator-only and accepted only from the administrator's own browser session (same-origin fetch, X-LocalAI-Confirm equal to the operationId); an API key is refused, whatever its role. The owner decides again on every call, from every frontend.

Routes

Route	Who	What
GET /api/addons/{id}/update	Feature holders	The owner's status, redacted for non-administrators.
POST /api/addons/{id}/update/check	Administrator, browser session	Fetch the library again and return the status. Applies nothing.
POST /api/addons/{id}/upgrade	Administrator, browser session	Upgrade to the candidate that was checked.
POST /api/addons/dispatch	Administrator	The typed actions update-status, update-check and update-upgrade (arguments expectedDigest and candidateDigest) give the same answers; the plain upgrade action of earlier releases still works and is not pinned.

No MCP tool maps to these routes. A read-only status tool may be exposed later.

The upgrade

POST /api/addons/{id}/upgrade takes operationId, expectedDigest (the digest of the installed package the caller saw), candidateDigest (the digest of the candidate the check resolved) and optionally generation. The owner then:

1. refuses, before it records anything, an installed package that is not the expected one (409 stale_installed), a candidate that is not the one the library resolves now (409 candidate_changed), an incompatible candidate (409 incompatible), a rewritten immutable version (409 integrity_error), and the same transition under another operation ID that has not failed (409 upgrade_conflict). A library that cannot be read is 503 check_failed;
2. stages and verifies the candidate, never changing the library, the channel or the trusted identity of the package;
3. for an add-on with data, backs it up, drains it and migrates it under the data recovery described above, and publishes the new generation only when it starts.

Repeating an operationId returns the recorded result and does not run again, from any frontend. Two requests with the same pins and different operation IDs run once: the second is upgrade_conflict, in flight or finished. Two upgrades to different candidates from the same installed package are serialized by the installed digest, and the loser is stale_installed. A candidate that fails to stage, fails its compatibility check at execution, does not start or is canceled keeps the last good generation running, or leaves the add-on in the explicit recovery state described under data recovery. The previous installation is never deleted. 409 upgrade_failed says that the attempt failed; a new attempt needs a new operation ID. After outcome_unknown, read operation-status with the same ID and do not submit a new upgrade.

Automatic upgrade

By default add-ons upgrade only when an administrator starts the upgrade. The setting Upgrade add-ons automatically (auto_upgrade_addons in the runtime settings, flag --auto-upgrade-addons or LOCALAI_AUTO_UPGRADE_ADDONS) is off by default, like the matching setting for backends. When you turn it on, the owner checks the library of every installed add-on shortly after start and then every six hours, the same schedule as the backend update check. When a library lists a newer version for the channel the add-on was installed from, the owner runs the same pinned upgrade that the Upgrade action runs. The digest and signature checks, the compatibility checks, the data backup and the recovery behavior above all apply, and an upgrade that fails any of them changes nothing and is written to the log. An automatic upgrade never changes the library, the channel or the trusted signer. A candidate that failed is not tried again until the library lists a different one. With the setting off, nothing upgrades by itself.

In the page

The Installed list and the management sheet show a badge with the candidate version and the short digest, and the compatibility reasons of an incompatible candidate. An administrator gets Check for updates with the time of the last check, and Upgrade when the state is update-available. Upgrade opens a confirmation that names the library, the channel, the installed and the new version and digest, and says that the data of the add-on is backed up first. It then shows progress and, on failure, the error and the kept version. Anyone else sees the status without actions.

The Upgrade action is on the card of the add-on in the list and on its entry in Find add-ons, next to the update notice, and in its details. All of them open the same confirmation.

When the notice is shown but the upgrade cannot start, the card says why instead of hiding the action:

What the card says	Cause	What to do
Only an administrator can upgrade it	Your account is not an administrator.	Ask an administrator.
The update is not confirmed yet	The check did not resolve a candidate to pin (no digest).	Select Check again.
The library names no build to install	A check ran and the owner still reports an update without a digest to pin.	Check again later.
The last check failed	The library could not be read.	Select Check again.
The update cannot run here, with the reasons	The state is incompatible.	None until the host or the platform fits.
The library lists this version with different contents	The state is integrity-error. It is never an update.	None.
The data needs recovery first	The add-on is in the recovery state.	Restore a backup in the details, then upgrade.
The latest check found no update	The list showed a cached notice that a newer evaluation no longer supports.	Select Check again.

After Check again, the card shows the answer of the check: "Checked just now" (or the time of the check) and what the check found, for example that the add-on is up to date. When the check found an update with a digest, Upgrade replaces the reason. A check that fails shows why, and you can run it again.

Upgrade all (N) shows in the bar of the list for administrators when at least one add-on has an update. It opens a dialog that lists every add-on with an update: the name, the installed and the new version, and, for one that cannot start, the reason above. Only add-ons that can start can be selected; Select all and Select none change the selection. Upgrade selected then runs the single-add-on upgrade for each selected add-on, one after another, never in parallel. Each is its own operation with its own operation ID and its own pinned digests, so the owner decides every one again. The dialog shows each add-on as queued, upgrading, done, failed (with the reason the owner gave) or skipped. A failure does not stop the others. You can close the dialog while it runs: the upgrades continue and the bar shows the progress and then a summary. Retry failed runs the failed add-ons again as new operations.

Check for updates in the bar of the list reads the library of every installed add-on again; it applies nothing, and an add-on that cannot be checked does not stop the others.

An upgrade started from the list, the details, Find add-ons or Upgrade all keeps running when you leave the page. It appears in the Activity page under the Add-ons filter, with progress, Cancel and Retry. Cancel stops the request, and the page then reads the operation to report what the owner did. The list of upgrades is kept by the browser that started them, so it is lost on a page reload.

While the add-on service restarts (an upgrade restarts it), the page cannot read the operation and the server answers 503. The page then waits longer before each new read, 2, 4, 8 and up to 30 seconds, and the upgrade says Waiting for the add-on service to start with the time of the next check; Check now in the details reads it at once. The Tools menu of a chat does the same for the state of the Sandbox: it says that the Sandbox is starting and reads its state again by itself, with the same back-off.

Add-on pages

An add-on with a ui object contributes pages to the web UI. The code of those pages is the packaged ES module the manifest names, served by the owner, and the host loads it after the host itself was built: nothing about an add-on is known to the UI build. The host never resolves imports for it, never hands it a router, a store or a request function, and never lets it replace a route.

The UI entry of an app

An enabled app whose installed package has a sha256 digest carries a ui object in GET /api/addons/apps. A disabled app, a headless app and an app without a digest do not.

Field	Meaning
entry	The same-origin URL of the module: /api/addons/{id}/assets/{digest}/{path}.
css	The URLs of the declared style sheets, under the same digest.
sdk	{min, max}, the host SDK versions the module accepts.
pages, navigation, chatActions	The identifiers the manifest declares. The host decides where each is shown.

digest is the package sha256 in lower-case hex. A new package is a new URL, which is why the files are immutable.

The asset route

GET /api/addons/{id}/assets/{digest}/{path} streams one declared file (the entry module or a style sheet) from the verified package on the owner, relayed by the frontend that received the request. It is a read for every caller who holds the addons feature.

• The file is served only while the add-on is installed and enabled and the digest is the current one. An older package, another add-on's digest, a disabled or uninstalled add-on, a file the manifest does not declare, a traversal, a directory and a missing feature all give the same denial without any file content, and a denial is never cached. (A path that ends in a slash is stripped by the app's router first, so it gives 404 instead of 403.)
• The owner serves the bytes, never the add-on process, and a file that is not a regular file of the package is refused.
• The response is text/javascript or text/css with X-Content-Type-Options: nosniff and Cache-Control: private, max-age=31536000, immutable.
• The browser authenticates with its same-origin session cookie. An API key goes in the Authorization header. A credential in the URL is never accepted, and a cookie request without same-origin fetch metadata is refused.
• A replica relays the response from the owner: the same bytes, headers and denials.

There is no MCP tool for this route, as for the other Apps routes.

The host SDK

The module exports createAddon(sdk) and nothing it needs comes from anywhere else. createAddon returns { Component, chatActions, dispose }; Component and chatActions are optional and dispose() is required. The SDK is a frozen object:

Member	What it is
React	The host's own React. Use sdk.React.createElement and sdk.React.useState: hooks run on the host's React instance. A bare react import fails to load, and a second bundled React would break hooks.
ui	Wrappers over host components: Button, Spinner, EmptyState, Badge, Stack and Text. A wrapper has a fixed type, so React never remounts it.
Icon, iconNames	A semantic icon. A name outside iconNames draws a neutral icon.
theme	tokens (CSS custom property references with stable names), mode() and subscribe.
locale	language(), subscribe, define(language, bundle) and t(key, fallback, vars). Bundles belong to this add-on only.
navigation	navigate(path), href(path), current(). Only the add-on's own declared pages resolve; anything else throws.
api	read(request) and propose(request), both bound to this add-on and its installed generation by the SDK. A read is always the typed read, with the resourceId of the resource the person named; a write is only ever a stored proposal that the host confirms. There is no raw request. Operation identifiers are op- and a UUID, which is a namespace name the owner accepts.
resources	The person's own resources of this add-on: list, create, rename, remove, apply, operation, grants, setGrant, revokeGrant, searchUsers and selection.get/set/clear. See the resource calls of the SDK.
notify	info, success, warning, error.
markDirty(flag), isDirty()	Unsaved edits. The shell asks before it leaves or replaces the page.
layers	openDialog(spec) and openPanel(spec): the host draws the dialog or side panel and owns its focus, Escape and unmount.
signal, onDispose(fn)	The signal aborts when the add-on is withdrawn; disposers run once.

A chat action is {id, label, icon, open(context)}, where id must be one of ui.chatActions. context is exactly {conversationId, resourceRef?}. There is no history, message or send hook, and an add-on cannot intercept submission or override a route. An action that is undeclared, duplicated, without a label or without open is ignored; an error inside open never reaches the host.

Compile an add-on with classic JSX transformed to React.createElement over sdk.React, inside the createAddon factory. Do not import a JSX runtime. The fixture packages under tests/addons/fixture/ui are written in that form.

The resource calls of the SDK

sdk.resources is the page of an add-on talking to the UI-only resource management of the host. It is a fixed list of typed calls: the add-on never supplies an address or a header, so it cannot reach another route or another add-on. The host shapes each call, adds the own-session confirmation header that the route needs and sends the add-on's installed generation, so a stale tab is refused by the server.

Call	Route
list()	GET /api/addons/resources?addonId=
create({arguments})	POST /api/addons/resources
rename(id, {name, revision})	POST .../{id}/rename
remove(id, {confirm, revision})	POST .../{id}/delete
apply(id, {tool, arguments, revision})	POST .../{id}/apply
operation(id, kind, operationId)	GET .../{id}/operations/{operationId}?kind=
grants(id), setGrant(id, userId, role), revokeGrant(id, userId), searchUsers(id, query)	the grant and picker routes
selection.get/set/clear(conversationId, ...)	/api/addons/conversations/{conversationId}/selection/{addonId}

A call that changes something takes an operation ID. The SDK makes a new one for each call; pass {operationId} in the options to repeat the same operation after an unknown_outcome answer. A malformed identifier, role or revision is refused in the SDK before any request. After the signal aborts every call throws and nothing is sent. Errors are AddonsError with a kind (stale, denied, invalid, unknown_outcome, sharing_disabled, not_selected, unavailable, network), the HTTP status, the server's type and, for an unknown outcome, the original operationId.

A chat request that has add-on tools enabled carries the conversation in the X-LocalAI-Conversation-Id header (when it has the shape the server accepts), which is how a tool call that names no resource finds the person's selection.

Core routes from add-ons

An add-on page can call the core API of LocalAI, for example /v1/face/verify, with sdk.core.fetch(path, init). The call runs with the signed-in user's own session, through the same authentication and authorization as any request the browser makes. The add-on gets no extra credential and no elevated role: a route that needs the administrator role stays closed to a user without it, and a route behind a feature (for example face_recognition) stays closed to a user whose feature is off. Those calls answer 403, and 401 when the session is missing. The add-on cannot widen what the user may do.

sdk.core.fetch is fetch with a fixed policy. It sends the cookies of the page (credentials: 'same-origin'), refuses other origins (mode: 'same-origin'), refuses redirects, and does not use the cache. It removes any Authorization and Cookie header that the add-on sets, and it stops when the add-on is withdrawn.

The path must be a same-origin path that starts with one /. These are refused with an invalid_path error before any request is made:

• an absolute URL, a path that starts with //, a backslash, or a control character;
• a . or .. segment, also when it is written with percent-encoding;
• the add-on gateway (/api/addons) and the auth routes (/api/auth). The add-on gateway has its own typed calls, and the auth routes hold account credentials. The check uses the decoded path, so /api/%61ddons/apps is refused too;
• an encoded / or \ in the first two segments, because it can hide a segment boundary. In a later segment it is a normal identifier (/api/voice-profiles/a%2Fb).

How the host loads it

1. The SDK range must contain the host SDK version (1). Otherwise nothing is imported and the page says the add-on needs another interface version.
2. Every URL must be a same-origin asset of that add-on's own package, all under one digest. Under a path prefix the host puts its base path in front, as for every API call.
3. Style sheets are read as text and scoped: every selector becomes :where([data-addon="{id}"]) …, :root, html and body map to the add-on's root, @import, @font-face and every other at-rule except @media, @supports, @container, @layer and @keyframes are dropped, and keyframe names are renamed. The sheet reaches only the add-on's own root. The same rule is applied to every declaration, including those inside keyframes, nested rules and @media, @supports, @container and @layer blocks, and to at-rule conditions: a declaration is dropped when its text, after comments and CSS escapes are resolved, contains url(, src(, image-set( (also with a vendor prefix), cross-fade(, element(, paint(, image(, expression(, behavior or -moz-binding, or when it has a backslash escape outside a quoted string. data: images are dropped as well; an add-on that needs an image uses an asset of its own package from its code.
4. The module is imported by URL with a timeout and an abort, createAddon is called with the SDK, and the result is checked.
5. The page component is drawn inside an error boundary in the add-on's root. A throw draws an error card and nothing else.

Revalidation, updates and unsaved edits

The shell reads the apps list while an add-on page, the Add-ons navigation or a chat action needs it (every 5 seconds, 30 in a hidden tab). An add-on that is no longer listed, was disabled, failed, was replaced by another generation or package, or has no UI any more is withdrawn at once: its SDK signal aborts, its component unmounts, its slots (navigation entries and chat actions) disappear, dispose() runs, its style sheets are removed, and the page says why with a Reload action. The owner already denies its code and calls, so an old copy that stays in memory cannot do anything.

A reader no longer sees a stopped add-on at all, so for a reader a disabled add-on reads as "not installed, or you can no longer use it".

If the add-on reported unsaved edits with markDirty, it is not unmounted: the page stays on screen, frozen and with its signal aborted, with a banner saying the edits cannot be saved. Reload then asks before it discards them. Leaving the page asks as well, and so does closing the tab.

Navigation entries

A page named in ui.navigation is linked in the detail view of the add-on and under its pinned entry while the add-on is enabled, not failed and compatible with the host SDK. The same registry keeps the validated chat actions per add-on and generation and removes them with the add-on. Declared chat actions are drawn by the host in the chat page (see Extension points and slots); a bare identifier has no placement and is not drawn.

Verification

sh scripts/addons/verify.sh esm-ui runs the node unit tests of the SDK, loader, style scoping and registry, the Go specs of the asset route through real stored keys and sessions, and a browser spec against the UI test server running with authentication and a real owner. That spec installs the fixture packages after the build, from a catalog on loopback, with the protocol fixture as their process.

Extension points and slots

An add-on extends the existing LocalAI pages through declarations in its manifest. The host validates every declaration at install and draws the contributions itself. Nothing is registered while the add-on runs: the module supplies only the components and the handlers its manifest names, and the host ignores the others (and logs them in a development build).

Declarations

All four lists are inside ui and are optional. A manifest without them is valid.

Field	Entry	Rules
ui.chatActions	A bare identifier, or {id, label, icon, placement, target}	id is a namespace name. label is plain text (60 characters at most) or a language map; markup characters are rejected. icon is a host icon name from the SDK list or the packaged ui.icon path; a remote or absolute address is rejected. placement is composer or message. target has exactly one of {tool} (a tool in capabilities.tools) and {component} (an export of the module).
ui.slots	{id, slot, component}	slot is one of the host slot identifiers below. id is a namespace name that no chat action uses.
ui.renderers	{kind, tool, component}	kind is approval-card or tool-result. tool is a tool of the same add-on. One renderer for each kind and tool.
ui.controls	A host control ID	One of the IDs in Host controls. At most 8, no repeats. The add-on supplies no component for it.

A component name is a PascalCase identifier of at most 64 characters. Each list has at most 16 entries. Identifiers are unique across chat actions and slots. Any unknown field in an entry is rejected.

Host slots

The slot list is closed. A new slot is a host change with its own review: it is never added by a manifest. A slot passes frozen props and nothing else. No slot passes message text, history, the model or any host state.

Slot	Where the host draws it	Props
chat.composer.actions	In an overflow menu next to the attach buttons of the composer. Actions of placement composer are menu items. Components are drawn in the menu.	conversationId, disabled
chat.message.actions	In the action row of each message. Actions of placement message are icon buttons.	conversationId, messageId, role
app.nav	On the Add-ons page, below the tile grid. Under the legacy switch: at the end of the second navigation row of the Add-ons hub, after the links of ui.navigation.	none
settings.sections	On the add-on's own page: the Settings tab of its details on the Add-ons page (/app/addons?detail=<id>&view=settings), one section for each component. Only the add-on's own sections are drawn there, and only for an administrator.	none
files.preview	On the Files page, in the preview of the selected file, between its size and its download link. The first component that handles the file's extension is drawn; the host's own view stays.	path, name, extension

A settings.sections component is drawn on the Settings tab of its add-on's details, not on the Settings page. The Settings page has an Add-ons group, shown when at least one add-on declares a section, that lists those add-ons, each with a link to its Settings tab. An add-on that has a page of its own gets a Settings link from the host above that page. The tab and the link are for administrators; a section is never drawn for anyone else, so a section does not need to hide itself from other users. The old address of a section, /app/settings#ad-section-<add-on>-<contribution>, opens the add-on's Settings tab. Below the add-on's own sections the tab draws the host controls the add-on has, such as the network of the commands it runs: LocalAI's own runtime settings, drawn and saved by LocalAI. Draw the section with the settings kit.

messageId is the position of the message in the conversation. role is user, assistant, system or tool.

A files.preview contribution declares the extensions it handles, lower case and without a dot ("extensions": ["docx", "xlsx"], one to 16, no repeats). No other slot has the field. path is the path of the file inside the user's files, name its name and extension its extension. The component reads the file through a read tool of its own add-on, as the user; the host never hands it file content.

Host controls

A host control is a component that LocalAI ships and draws itself, bound to its own settings. An add-on asks for one by ID in ui.controls; it supplies no code, value or label for it, so it cannot change what the control does. The list of controls is closed, like the slot list. Any add-on can use any control it is entitled to: the host never reads the add-on's ID to decide.

Control	Where the host draws it	The add-on must
network-policy	On the add-on's Settings tab: the network of the commands it runs (limit, preset of a new chat, ports, extra registries, always allowed hosts, private networks, audit retention and the audit; see Network access). In the chat's Tools menu: the Network row, while the add-on runs the chat's commands, with a link to the add-on's Settings tab for an administrator. On a network card: the same link.	Be granted a host operation that covers egress.check (for example devbox=egress. in LOCALAI_ADDON_HOST_OPS) and declare a tool that replaces bash.
tool-providers	On the add-on's Settings tab: the rows of Settings > Tools for the built-in tools it replaces (who serves the tool, the choice when several add-ons replace it, and the switch of the built-in tool).	Declare at least one tool that replaces a built-in tool (shadows).

The apps list carries the controls in ui.controls, with only those the add-on is entitled to now. A control whose grant was removed is dropped from the list, without an error. Declaring a control grants nothing: the host operation, the limit and the audit work the same with or without it. A manifest without ui.controls is valid, and an add-on with a control has a Settings tab even when it declares no settings.sections.

A second add-on that runs commands through the same proxy gets the same block, row and link with one manifest line and one grant, and no change to LocalAI:

"ui": { "entry": "ui/devbox.js", "sdk": { "min": 2, "max": 2 }, "controls": ["network-policy", "tool-providers"] }

The grant is one more entry of LOCALAI_ADDON_HOST_OPS, next to the ones it has: devbox=egress.. The setting replaces the default list, so keep its other entries.

What the host guarantees

• A contribution is drawn only while the apps list the server returned says the add-on is enabled, not failed and compatible with the host SDK, and only for the generation and package the loaded module belongs to. A disable, a crash, an update or a revocation removes every contribution at the next list. A re-enabled add-on is loaded again on the next page load.
• The page reads the apps list once for the slots. It keeps reading it only while a usable add-on contributes.
• Each contribution is in its own error boundary. A component that throws leaves a small "Unavailable" marker, and the page keeps working.
• A contribution is drawn in a container of its own. It cannot replace or hide a host node. The style sheet of the add-on is scoped to its root.
• A tool action is one typed read call of the same add-on, for the tool the manifest named, with {tool, context} as arguments. The context has conversationId and, for a message action, messageId and role. The server authorizes the call as for every other call: hiding a button is never the control. The result is drawn by the declared tool-result component or, without one, as bounded text. A result that names a stored proposal ({"proposalId": "..."}) opens the approval frame.
• A component action opens the exported component in a host dialog, with the same context and a close function.
• The server, not the page, limits a tool action to the tools the add-on declared: a request for an undeclared tool is a 403 (see Declared tools).

Add-on UI code is trusted code

The UI module of an add-on is native code that runs in the origin of the host page. It is not sandboxed. The host cannot technically prevent it from reading the page, calling the same routes as the person or driving host buttons, including the buttons of the approval frame. The controls are the trust decision when the person installs the add-on, the server-side authorization of every request (roles, grants, declared tools and the revocation of credentials) and the approval frame that the host owns: a confirmation needs the proposer's own browser session and the digest of the stored proposal. Install only add-ons that you trust.

Approval cards

A proposal of an add-on is shown in a frame the host owns. The frame loads the stored proposal, shows who asks, how many changes it has, when it expires and its state, and draws the approval-card component declared for the tool inside it. The component receives plain data (proposalId, tool, effects, arguments). It gets no callback and no digest. Confirm and Cancel are the frame's buttons. A confirmation sends the X-LocalAI-Confirm header and the digest that was displayed, and the server accepts it only from the proposer's own browser session, as in Proposals and confirmation. An add-on asks for the frame with sdk.approvals.present(proposalId). The proposal must belong to the same add-on.

Private files in a page and in chat

A read tool may answer with file results. LocalAI replaces each output of the add-on by a private artifact identifier of its own before anything reaches a browser or a model. sdk.files.read(request, options) is sdk.api.read for such a tool: it resolves to { files, payload }, where each file is { id, name, mediaType, size } as the host recorded it. The add-on's handle, a path and a URL are never part of it. The add-on draws the file with sdk.ui.ArtifactCard, which takes only artifactId: the card loads name, size, state and expiry from GET /api/addons/artifacts/{id} and offers the authenticated download of the host route. A chat tool result of kind artifact is drawn with the same card. The identifier is the only thing the chat stores. An add-on cannot supply a link, a name or a path that the card shows.

Verification

sh scripts/addons/verify.sh host-slots runs the node unit tests of the slot registry, the component validation of the loader and the approval rules; the Go specs of the manifest declarations and the apps list; and a browser spec against the UI test server with authentication and a real owner. The spec installs the slots fixture package after the build. It contributes to every slot, with a contribution that throws and a style sheet that collides with host class names.

First-party package and release channels

The first-party add-ons (Tables, Voice, Face, Fine-tuning, Quantization, Studio and Talk) are each distributed as an OCI image of one platform for each supported operating system and architecture, joined by an image index. The host installs it as any other OCI package: it extracts layers into a private directory and starts the executable as a managed process. No container runtime runs the add-on. Docker or ORAS only construct and move the distribution.

Package format

sh scripts/addons/package.sh --layout <dir> --version <semver> --platform <os/arch> writes a local OCI layout with one image and a receipt.json of its digests. The tool never pushes, tags, signs or logs in.

Add --config addons/<id>/package/platforms.yaml to package another add-on. The script reads the add-on name, the executable and the UI directory from that file and, without --executable, builds ./addons/<id>/cmd/<executable> for the host platform. Tables is built with CGO_ENABLED=1. The other add-ons are pure Go and are built with CGO_ENABLED=0, so their executables need no C toolchain on any platform. The default configuration is the one of Tables.

Part	Value
Image manifest	OCI image manifest, artifactType application/vnd.localai.addon.package.v1.
Annotations	io.localai.addon.id, io.localai.addon.version, io.localai.addon.platform, org.opencontainers.image.version, and org.opencontainers.image.revision when the commit is known.
Config	OCI image config with the os and architecture of the platform. No history, no timestamp.
Layer executable	The entrypoint (mode 0755) and the runtime files the platform needs (mode 0644), as application/vnd.oci.image.layer.v1.tar+gzip.
Layer ui	Only when the manifest declares a UI: every file under the directory of ui.entry. The declared entry, style sheets and icon are required. The layer is absent otherwise.
Layer manifest	manifest.json with version, os and arch set to the release version and platform.
Image index	One entry for each enabled platform, artifactType application/vnd.localai.addon.package.index.v1, with the version and commit as annotations.

Every part is deterministic: the same inputs give the same digest. A package is rejected, before it can be published, when:

• the entrypoint, a declared UI asset or a required runtime library is missing;
• the manifest, the image config, the annotations or the receipt name another platform or version, or the artifact type differs;
• a layer has an unexpected media type, role, order or file, or a blob does not hash to its digest;
• a file has set-id, sticky or group/world-write bits, a non-entrypoint file is executable, the entrypoint is not, or a path is absolute, escapes the root, repeats or is not a regular file;
• the UI bundles its own React runtime (the host provides React through the UI SDK);
• the icon is not an image of its declared type.

addonpkg extract restores a verified layout to a directory with the validated executable mode. The installer applies the same path rules and ignores the modes in the archive other than the executable bit.

Release platform matrix

Each first-party add-on has its own addons/<id>/package/platforms.yaml. It lists the platforms of that add-on. A platform is shipped only while enabled: true, and it needs a native runner and the executable-start native check. Until a platform's native check passes on a runner of that platform, it stays listed with enabled: false. The workflow fails closed on an empty matrix and on a missing repository input (ADDON_OCI_REPOSITORY). There is no default registry. The add-on repository is quay.io/go-skynet/local-ai-addons; every other backend image repository is refused.

Every platform matrix lists four platforms: linux/amd64 on ubuntu-24.04, linux/arm64 on ubuntu-24.04-arm, darwin/arm64 on macos-15 and darwin/amd64 on macos-15-intel. Voice, Face, fine-tuning, quantization, Studio and Talk enable all four. A test fails when a platform is enabled without a recorded check.

A package that ships skills names their folder with skills: (for example skills: addons/office/skills). The build lays it at the package root as skills/, in a layer of its own, on every platform, and checks it with the rules the owner applies at install: the folders must be the skills the manifest declares, and every file must pass the checks of Skills. A package that the owner would refuse is never built.

Each package also has a catalog template in gallery/addons/<id>.yaml.tmpl, so it can be listed in Find add-ons once a maintainer renders and publishes its entries. See gallery/addons/README.md for the commands.

All add-ons publish to one OCI repository, the value of ADDON_OCI_REPOSITORY. Each add-on is a set of tags of that repository, and every tag starts with the add-on id: <id>-sha-<short sha> for a commit, <id>-<version> for a release, and <id>-development or <id>-latest for a channel alias. For example, the Fine-tuning build of one commit is oci://quay.io/go-skynet/local-ai-addons:fine-tuning-sha-99d1dca2c869.

A release entry of the official library is pinned by digest. A development entry is mutable, as a development backend is: it names the moving tag (oci://quay.io/go-skynet/local-ai-addons:fine-tuning-development), sets mutable: true and has no digest. At install LocalAI resolves the tag to a digest, verifies the signature against that digest, installs by it and records it in the receipt. A mutable entry must be an OCI artifact. The update check compares a mutable entry by the digest its tag names (see How the candidate is chosen).

Set ADDON_OCI_REPOSITORY to quay.io/go-skynet/local-ai-addons. The value has no tag and no digest. A value that is a backend image repository makes the add-on tools refuse it. addonpkg target --repository <repository> --addon <id> --tag <tag> prints the reference one add-on publishes to, without a network request.

The index entries pin each platform image by digest, so the tag only names the artifact. The host accepts a tag reference (oci://<repository>:<tag>) and reads the repository with any number of path elements, so an index that still names the old repository-per-add-on form (oci://<repository>/<id>:<tag>) keeps working.

Publishing events (a push to master, a version tag and a published release) fail closed for an add-on until at least one of its platforms is enabled in its platforms.yaml. Pull requests and other pushes publish nothing, so an add-on without an enabled platform only gets no package legs there. A missing or malformed platforms.yaml fails the plan on every event.

Channels and tags

scripts/addons/channel.sh --event <name> --ref <full ref> --sha <commit> is the one place that decides what an event may publish. Its output is what the workflow uses.

Event	Immutable tag	Moving tag
Pull request	none	none: nothing is pushed, no secret or OIDC token is available
Push to master	sha-<12 hex of the commit>	development
Push to another branch	none	none: a branch push never publishes, whatever its name
Push of a v<semver> tag	<semver>	none: a tag alone never advances latest
Published, stable, non-draft release	<semver>	latest, after complete publication
Published pre-release	<semver>	none

The names in this table are the part after the add-on id. The registry tag is <id>-<name>, for example tables-sha-4edf30cda051 or voice-development.

Names are in disjoint namespaces: latest and development are reserved, <semver> and sha-<12 hex> are immutable, and every branch channel starts with branch-. The branch channel is branch- + a label + - + the first 12 hex characters of the SHA-256 of the full original ref (refs/heads/...). The label is the branch name in lower case, every run of characters outside a-z0-9 replaced by one dash, cut to 40 characters (x when nothing is left). The hash makes Feature/X, feature-x and feature/x three channels. Only the exact ref refs/heads/master is development. Nothing built from another branch is published, and no new branch channel is created. Channels that an earlier run published stay where they are.

Immutable tags are never overwritten. A second publication of a version with other content is refused. The one exception is a rebuild of the same version and commit after an earlier run published it: the published content is kept (--adopt-existing), and the workflow verifies its signature before it promotes anything.

Publication and signing

The workflow .github/workflows/addons-package.yml has these jobs.

Job	What it does	Credentials
validate	One leg for each add-on. Builds the UI of the add-on, and runs the packaging checks (which cover every add-on) on every event, including pushes and releases. A red validate stops package, publish and promote.	none
plan	Runs channel.sh and, for each add-on, checks its platform matrix. It joins the matrices into one list of legs and names the add-ons that have legs. It checks the registry input.	none
package	Needs plan and validate. One native runner for each enabled platform of each add-on: builds the executable and UI, runs package.sh and the native runtime check, uploads the layout.	none
publish	One leg for each add-on. Needs plan, validate and every package leg. Joins the legs of the add-on, pushes platform digests and the index by digest to the repository of the add-on, tags the immutable version, signs by digest (keyless, OCI 1.1 referrers, new bundle format) and verifies with the exact identity of this workflow.	registry login and id-token: write, in a protected environment
promote	One leg for each add-on. Needs publish, and through it validate. Moves aliases, serialized by add-on and channel.	registry login, in a protected environment

An add-on with no enabled platform gets no legs. On a pull request or a branch push this fails nothing. On an event that publishes, the plan fails for that add-on and nothing is published, so a release never leaves one add-on behind without notice.

A pull request, including one from a fork, never reaches publish or promote, and pull_request_target is not used. A missing or failed leg, or a failed validate, leaves every alias unchanged: nothing is published from a partial set of legs, latest moves only after complete validation, and aliases move only after signing. The workflow policy (addonpkg policy) requires this chain in the job graph. It rejects a publish that does not need validate, and a validate job with continue-on-error or an always(), failure() or cancelled() condition.

Promotion is serialized by a concurrency group for each channel and is checked against the registry: development and branch channels do not move to a commit that is an ancestor of the one they name, and latest does not move to a lower version or to a pre-release. A slow older run therefore ends as a no-op.

A force-pushed branch is the new truth of that branch. When the commit a branch channel names is no longer in the checkout (git answers "not a valid commit name"), the promotion succeeds only if the run is a push to the exact ref that owns the channel (--ref and --event of addonpkg promote; the channel name is derived from the ref). For any other event or ref the alias stays where it is and the command fails with E_UNREACHABLE. development follows the strict rule: an unreachable commit is always refused with E_UNREACHABLE, and it needs a manual decision. Any other git failure fails the promotion. A slow run of a pre-rewrite commit that finds the rewritten tip in its checkout cannot be told apart from a new diverged commit, so a promotion of the old commit after a force-push can briefly move the channel back until the next push. The registry has no atomic compare-and-swap for tags, so the group, not the check, is what closes the window between two runs.

The signer identity is the workflow file of this package pipeline, not the backend image workflows: https://github.com/<owner>/<repo>/.github/workflows/addons-package.yml@<ref> with the issuer https://token.actions.githubusercontent.com. The identity contains the ref of the run, which differs for each release tag.

Catalog entries

gallery/addons/tables.yaml.tmpl and addonpkg catalog render one entry for each platform from a verified merged layout. An entry pins the digest of the platform image, names the repository and the immutable tag, carries the channel it is current for and holds the expected signature policy. The catalog resolves a selection by version or by channel, so a channel is metadata that names the entry and its digest, never a tag that is read at install time. See gallery/addons/README.md.

An entry may also carry a requires object, the compatibility hint of the update check (see Update detection and upgrade).

The host applies the signature policy of the library (source), not the one of an entry: a refresh replaces an entry's policy with the policy the administrator configured on the source. Executable verification (the signature of the package digest against the workflow identity) and the trust of the catalog document itself (how the library is fetched) are two separate things, and the host tests them separately.

Workspace operations

An add-on that holds sandbox=workspace., office=workspace. or web=workspace. in the host operation setting can ask for the files of the user it is working for:

• workspace.materialize builds a read-only tree of some folders of the user in a directory of the add-on's own data directory.
• workspace.commit takes the upper.tar layer the add-on left in such a directory and commits it as one change set. It needs an act request.
• workspace.changes lists the change sets after a revision.
• workspace.admin answers only when the user of the request is an administrator, and is refused for anyone else. An add-on uses it to guard a setting; it never decides who is an administrator itself.
• workspace.export copies one file of the user into the data directory, and workspace.import (an act request) writes one file of the data directory into the files of the user, with the version check of the files API.

Every call runs as the user of the live request, never as a user named in the arguments. dir is a relative folder inside the add-on's data directory; an absolute path, a parent reference or a symbolic link is refused.

What a call can reach depends on who made the request:

• A call from a chat. The host binds every add-on call that a chat makes to that chat, for every tool kind: a read tool that names no resource (such as the Office add-on's read-docx) as well as an action or a change. The calls act with the chat's own access, the same access as the chat's file tools:
  • In a chat whose agent has no folder, / is the user's home, as before.
  • In a chat whose agent has a folder (see [One folder for an agent]({{%relref "features/chats-agents#one-folder-for-an-agent" %}})), / is the agent's folder. A path names a file inside it: /Other/x is Other/x inside the folder, not a folder of the home. The add-on can also name the chat's own folder as /workspace/chat/x, and the shared folders as /workspace/shared-<name>/x (read only). A path with a parent reference, NUL or a control character is refused. An encoded or look-alike separator (..%2f, %2e%2e, a backslash, U+2024, U+FF0F) is part of a name. A path outside these folders is not found. The agent's mode applies: a read-only folder takes no file.
  • When an add-on mounts its / at home, the host also mounts the chat's folder at chat in a chat whose agent has a folder, so the Sandbox sees /workspace/home and /workspace/chat as the chat's file tools do. No add-on needs a change for this.
  • workspace.changes lists only the change sets inside the chat's folders, and a workspace.commit entry outside every mount comes back as an outside-view conflict.
  • A call whose chat the host cannot resolve (the chat is gone, its agent's folder is missing, or the binding names another user than the actor) is refused with bad_request. It never falls back to the home. An action always works for one chat, so an action request that carries no binding and whose resource names no chat of the user is refused too.
• A call that no chat made, such as a call from the add-on's own page or the OpenAI-compatible route, works on the user's whole home, as the Files page does. The person acts with their own rights there.

The binding stays on the host. It is kept next to the envelope of the live request, and between frontends it rides in the signed forwarded request (see Owner forwarding and HTTP relays). It is never a field of the envelope the add-on receives, which installed add-ons decode with unknown fields refused, so the add-on contract does not change. The access is resolved again from the stores at each call (stateless per request).

Network operations

An add-on that holds sandbox=egress. in the host operation setting (the default) can ask about the network connections of the commands it runs (see Network access):

• egress.check (act, feature chats) takes {host, port, kind, tls}, with kind connect or http, and answers {action, reason, allowPrivate, self}, with action allow or block. An allow carries self: the addresses of the server as CIDR strings (its interfaces, loopback, and the addresses of the host of the external base URL), the same list web_fetch refuses. The add-on refuses them on every dial, besides the interfaces of its own machine, so a command cannot reach the LocalAI API through the server's public name or address. When a card asks the person (see the cards under Network access), the call waits for the answer, at most 60 seconds, before it answers. The chat is the resource of the live envelope and the user is its actor: arguments named chat, user, preset or allow, or any other unknown name, are refused. The host writes the audit line of the verdict itself. The operation may wait up to 60 seconds for a card and has a 75-second timeout.
• egress.audit (act, feature chats) takes {events: [...]}, at most 500 lines of {host, port, kind, verdict, reason, addr, bytesUp, bytesDown, durationMs, at}. verdict is done (an allowed connection ended; the bytes it moved) or block (the add-on refused it itself: an address that is private, the server's own or a metadata address, a TLS name that is not the host, a limit, or a connection while no call ran).

Simple Office

Simple Office is a lightweight option for basic office-file work, without requiring Sandbox. It reads and edits documents in the user's files, from a chat. It is a package like the others (office in the add-on library) and uses the workspace.* host operations to take a file in and give the result back.

Tool	Kind	What it does
read-docx, read-xlsx, read-pptx, read-pdf	read	Return the text of a file: paragraphs with their index, cells row by row (a formula that has no stored result shows as the formula, such as =SUM(A1:A2)), slides with their shapes, pages. The answer is cut to about 48 KiB and says so.
edit-docx, edit-xlsx, edit-pptx	action	Apply a list of changes (at most 200) to a file. All of them are saved, or none.
create-docx, create-xlsx, create-pptx, create-pdf	action	Make a new file. create-docx and create-pdf take a markdown string; create-xlsx takes a single-sheet csv string. All three take path and optional title strings; XLSX also accepts sheet. The former blocks and sheets arguments are no longer supported. PDF writes Latin-1 text with the built-in font. See Simple Office for the contract and limits.

The action tools run at once, as the user of the chat, with no confirmation, and only in a chat. Every tool, a read tool too, works with the chat's own access: in a chat whose agent works in one folder, / is that folder (see Workspace operations). Without accounts, the files are the same ones the chat and the Files page use. With accounts and local-owner mode, the workspace.* operations are refused, because the local owner is no account and has no files; use named-user mode for the Office and Sandbox add-ons. A file that someone else changed after the tool read it is never overwritten: the tool's version is saved as a second file next to it, and the answer says so. A file that is not of the type its name says, or a path outside the user's files, is refused with a reason the model can read.

The add-on also ships four skills, xlsx, docx, pptx and pdf (see Skills). Each one replaces the built-in skill of the same name while the add-on can serve the chat. They are short procedures written for the tools above: read first, change only what must change, create a file only when none exists, and say what the tools cannot do (charts, tracked changes, form fields and the like) instead of describing a manual method. They need no command runner, so they are never "unavailable here". When the add-on is stopped, off for the chat or removed, the built-in skills come back. A test checks that every tool a skill names is a tool of the manifest.

Web

The Web add-on serves the chat tools web_search and web_fetch in place of the built-in ones. Its tools search and fetch declare "shadows": "web_search" and "shadows": "web_fetch", so a chat with the add-on on and running sees one web_search and one web_fetch, served by the add-on. It adds two things the built-in tools do not have: search providers that need a key (Brave Search, Tavily, Google Programmable Search), tried in an order the administrator chooses, and the text of PDF files. Both tools run the code of the built-in tools for the rest: the same SSRF guard, limits and untrusted-content label. See [Web search and web pages]({{%relref "features/web#the-web-add-on" %}}) for what it does in a chat and how to set it up.

Tool	Kind	What it does
search	read	Replaces web_search. Arguments query, maxResults, freshness, site.
fetch	read	Replaces web_fetch. Arguments url, maxChars, format. Reads PDF too.
status, set-settings, test-search	read, page only (uiOnly)	The settings of the add-on, on the Settings tab of its details on the Add-ons page: the providers, their order and keys, and a test of each provider. A change and a test need an administrator, which the add-on asks the host with workspace.admin.

The contract allows no underscore in an argument name, so the add-on's tools name max_results and max_chars of the built-in tools maxResults and maxChars. The model reads the add-on's schema.

The keys are kept in settings.json in the add-on's data folder (mode 0600). They are never returned by a tool, written to a log or put in a model's prompt or a tool result: the Settings tab shows only whether a key is set. A key reaches its provider in a request header (X-Subscription-Token, Authorization or X-Goog-Api-Key), never in a URL, and the clients of the keyed providers follow no redirect, so a key is never sent to another host.

The add-on runs on the add-on owner, like every add-on. Its settings, its search cache and its rate limits live in that one process, so they are the same for every frontend of a cluster.

Sandbox

The Sandbox add-on lets a chat run commands and code in isolation. It has one tool for the model, the action tool bash, which replaces the chat's built-in bash ("shadows": "bash"), so a chat with the Sandbox sees one bash and it runs isolated. The model reads and changes files with the chat's own file tools (read_file, write_file, edit_file, list_files, search_files and the others; see [Chats]({{%relref "features/chats#running-commands" %}})): they work on the same files, so a script written with write_file runs with bash at once, and what bash writes is in the user's files after the call. bash runs inside an isolated sandbox of the user and the chat, with the user's files mounted under /workspace/home, where each call starts. The skill folders are not mounted in the sandbox: to run a skill's script there, the model copies it into the home with the file tools first. The status tool keeps its add-on name (addon__sandbox__status). Without the Sandbox, or with it off or unable to run, a chat may still run commands on the server itself, without isolation, when an administrator allows it; see [Commands on this machine]({{%relref "features/chats#commands-on-this-machine" %}}).

An older package of the add-on also served read, write, edit, glob and grep. They are gone: reinstall the package to drop them. A chat's switch for the add-on is kept by add-on ID, so nothing a chat stored changes.

• One sandbox per user and chat. The key is the pair of the user and the conversation. The user comes from the host. Nobody else reaches the sandbox, and the model cannot name another user or chat.
• Changes are one change set. After every call the add-on commits what the code changed, as one change set in the history of the user's files, so it can be undone there. If someone else changed the same file meanwhile, their version stays and the sandbox's version is saved as a copy; the answer names the file and the copy. A file written outside /workspace/home (where bash starts) is not saved, and the answer names it too. In a chat whose agent works in [one folder]({{%relref "features/chats-agents#one-folder-for-an-agent" %}}), /workspace/home is that folder (read only when the agent can only read it), and the chat's own folder is at /workspace/chat, where files are saved as well; nothing else of the user's files is in the sandbox. The change set belongs to the chat that ran the command, so the chat lists the files it saved (see [Files that a command saved]({{%relref "features/chats#files-that-a-command-saved" %}})). A file change by someone else makes the sandbox rebuild its view of the files before the next call.
• Limits. At most 4 sandboxes for a user (the oldest idle one gives way to a new one) and 16 in all. When the 16 are used, the oldest suspended sandbox of any user gives way to a new one, so sleeping sandboxes never keep a user out; a call is refused with a busy answer only when every sandbox is running. A sandbox that is idle for 10 minutes is suspended and wakes on the next call. A sandbox is removed 24 hours after it was made. An add-on restart removes every sandbox it left, suspended ones too. When the add-on stops (it is disabled, or the server shuts down) it first stops all its sandboxes, for at most 8 seconds. The supervisor waits for that before it kills the add-on process; a server that is killed with SIGKILL cannot do this, and the next start removes what was left. When a suspended sandbox is removed, the host backend also stops the processes it left running.
• Output. A tool answer is cut at 64 KiB, and a bash call stops after its timeout (120 seconds by default, 28 minutes at most, so the answer is back before the 30-minute limit of an action call).
• Without a backend. When the host has no way to run sandboxes, nothing runs. A chat is not offered the Sandbox tools then: it gets one stand-in tool that tells the model why, so the model answers that it cannot run code instead of trying. Without the add-on, or with it switched off for a chat, the chat's model is told that it cannot run commands, and the chat's Tools menu says so to the user; see [Running commands]({{%relref "features/chats#running-commands" %}}). A tool that is called anyway answers "This host cannot run commands." The same happens when the Files workspace is off, because a sandbox needs the user's files.
• Where you see the state. The Sandbox page tells every user which backend runs the code, or that no backend works, or, while the host backend is in use, that commands run directly on this machine with no isolation. An administrator also sees, on that page and in Settings, the reason each backend does not work and what to do. An administrator can choose the host backend there: the page shows the warning that commands then run on this machine without isolation, and asks for the typed confirmation before it saves. A user who is not an administrator is told that only an administrator can change this.
• Undo. Each bash call that changes files is one change set in the history of the user's files (Files, History), and it is undone there, one call at a time. A whole turn is not one change set: the sandbox runs in the add-on process and the Files service commits each write of a tool on its own. The file tools of the chat itself work the same way, with an Undo note under the call.
• Shared code. The Host backend, the manager that keeps one sandbox per user and chat, and bash live in core/services/studio/sandbox (host, runner). The add-on runs them over its host operations; the chats' built-in bash runs the same code over the workspace store directly.

Sandbox backends and settings

The add-on runs commands in one of these backends, chosen by an administrator on the Settings tab of the Sandbox details on the Add-ons page (the Sandbox page links to it):

Choice	What it is
Automatic	KVM (Cloud Hypervisor) when the host has it, otherwise gVisor, otherwise nothing. It never picks the host backend, not even as a last resort.
gVisor	runsc with user namespaces. Needs runsc (the package carries it), taskset, and on Ubuntu 24.04 and later an AppArmor profile (the package carries gvisor-profile, which prints it).
KVM	Cloud Hypervisor with virtiofsd, mke2fs, debugfs and /dev/kvm. The package carries Cloud Hypervisor and, on amd64, virtiofsd. bwrap is used when present and not in a container.
Off	No sandbox. Every tool answers that the host cannot run commands.
Host	Runs commands on the host as the user LocalAI runs as, with no isolation.

The Settings tab lists the backends of the host, each with its state (in use, ready or unavailable); the reason a backend does not work is under Details. When gVisor fails with exit status 128 and cannot read client sync file, the row also says that gVisor needs an AppArmor profile that allows user namespaces and a container that may create them, and links to The Sandbox in the LocalAI container. The add-on asks for the network-policy and tool-providers host controls, so the same tab also holds the network of the Sandbox commands (the limit, the preset of a new chat, the ports, the extra registries, the always allowed hosts, private networks, the audit retention and the audit) and the bash row of Settings > Tools, and saves them like the Settings page. An explicit choice is never replaced by another when it does not work: the tools say they cannot run.

The Host backend is for a machine that can run neither of the others. Choosing it takes two steps. The page shows a plain warning and asks the administrator to type the sentence I understand that commands run on this host without isolation. The add-on then saves the choice. The confirmation is forgotten when another backend is chosen. The backend refuses to start when LocalAI runs as root. Every command is logged with the name of the chat user, in the LocalAI log with the prefix [sandbox] (LocalAI writes everything an add-on prints to its output or error into its own log, one line at a time, cut at 4096 bytes), and the Settings tab of the Sandbox details (for administrators) shows a notice for as long as the host backend is the one in use. The model sees it in the answer of the status tool. Treat the machine as one that code can harm.

The choice is kept in the add-on's data directory. Changing it removes the running sandboxes; what they did is already in the users' files.

The guest image, runsc and the KVM programs are not part of the add-on's source. A release sets main.guestImage (the image, by digest) at build time. The checksums of the runsc and of the Cloud Hypervisor and virtiofsd it was tested with are compiled into the add-on (runscPins and kvmPins in addons/sandbox/cmd/sandbox), and tests keep them equal to package/runsc.json and package/kvm.json. A build without a guest image reports that on the page, and only the host backend can run.

What the guest has

The guest image carries exactly these programs and libraries for the model to use (the model is told this list in the prompt of each chat that has the Sandbox):

python3 (pandas, openpyxl, docx, pptx, pypdf, Pillow), pip, jq, ripgrep (rg), sqlite3, zip, unzip, file, tree, poppler (pdftotext, pdftoppm), git, curl

It does not have Node.js or npm, ImageMagick, ffmpeg or OCR. The command line runs as an unprivileged user and apt is blocked by the proxy, so nothing can be installed from inside the guest with a system package manager. pip can install Python packages within the chat's network setting (see Network access). Adding a program means a new guest image. The build of the image fails when one of these programs is missing or does not start (addons/sandbox/images/guest/check-tools.sh).

Network access

The guest of a sandbox has no network device: no NIC, no TAP device and no user-mode network stack. A command reaches the network only through one proxy, on guest port 3128, which the guest image sets in HTTP_PROXY, HTTPS_PROXY, pip, apt, git, npm and GOPROXY. The guest agent forwards each connection to that port to the add-on, over vsock (KVM) or a Unix socket (gVisor), and the add-on's proxy asks LocalAI about each connection with the host operation egress.check. LocalAI decides; the add-on dials. The proxy never decrypts TLS and holds no certificate the guest trusts.

The guest image has curl, git, pip and the CA certificates, and all of them use the proxy (the full list of its programs is under What the guest has). The model is told so at each turn, with the chat's preset. To read or search web pages it is told to prefer web_fetch and web_search when the chat has them (the built-in ones or the Web add-on's), and to use curl for APIs, downloads and installs. With the network off, the note says that nothing is reachable.

Commands that run on this machine (the host backend, or the chats' built-in bash) have the network of the server: none of this applies to them, and the settings say so.

Presets. A chat has one of three presets, and an administrator sets the ceiling, the widest preset any chat may use:

Preset	What a command reaches
Off (default)	Nothing.
Package registries only	The HTTPS hosts of the package registries below, on port 443.
Allow all	Every public host on an allowed port.

The preset of a chat is the lower of its own choice and the ceiling (chats_egress_ceiling, default off). A new chat starts with chats_egress_default, capped by the ceiling. The preset lives in the chat's record on the server, never in a file of the workspace: the sandbox can write the home, so a file there could be changed by the code it is meant to confine. A preset never reaches a private, loopback or link-local address, nor the server's own addresses, nor a cloud metadata address. The server's own addresses are the same list for web_fetch and the Sandbox: the interfaces of the server, loopback, and the addresses that the host of the external base URL resolves to (read again every 60 seconds); the add-on also refuses the interfaces of its own machine. With chats_egress_allow_private on, a host that an administrator allowed by a rule or a card answer may reach private addresses; the metadata addresses stay refused. See [Runtime settings]({{%relref "features/runtime-settings#tools" %}}) for the six chats_egress_* settings.

Package registries. The list ships with LocalAI (core/services/studio/egress/registries.json) and changes only by a reviewed pull request that also changes this table and is named in the release notes. Every entry is an exact host name, HTTPS on port 443, with no pattern. Upload hosts (such as upload.pypi.org) are not on it, and neither is raw.githubusercontent.com, which serves any user's content. An administrator adds exact hosts with chats_egress_extra_registries, for example an internal mirror (which is private and so also needs the private switch and an administrator's rule). Extra registries are registries: they follow the Package registries rules (HTTPS only, and only with that preset or a card). To allow one more host for any request, use the list below.

Always allowed hosts. An administrator can list exact hosts in chats_egress_allowed_hosts (at most 64; host or host:port to pin the port; no pattern). A host on the list is allowed, with no card, in every chat whose effective preset is Package registries only or Allow all, on any request kind (plain HTTP as well as HTTPS) and on the allowed ports (chats_egress_ports), or only on the pinned port. With the preset off in a chat nothing on the list is allowed there, and with the ceiling off the list does nothing at all. The list does not widen the guard: a private, loopback, link-local, metadata or server address is not allowed by it, only by the private switch together with an administrator's rule or card answer, and the TLS server name must still match, on the pinned port as on 443. An administrator's entry also wins over a deny the person stored in one chat before the host was listed. The audit line of a request this way has the reason allowed_host. The chat's Network panel shows the list to the person, read only.

Registry	Hosts
Python (PyPI)	pypi.org, files.pythonhosted.org
JavaScript (npm)	registry.npmjs.org, registry.yarnpkg.com
Go modules	proxy.golang.org, sum.golang.org
Rust (crates.io)	index.crates.io, static.crates.io, crates.io
Ruby (RubyGems)	rubygems.org, index.rubygems.org
Java (Maven Central)	repo.maven.apache.org, repo1.maven.org
GitHub releases and archives	github.com, codeload.github.com, objects.githubusercontent.com, release-assets.githubusercontent.com

What the proxy checks. The host of a request goes through the same guard as web_fetch ([Addresses inside your network]({{%relref "features/web#addresses-inside-your-network" %}})): a canonical name or literal, no address zone, no numeric tricks, no localhost. A host the guard refuses is never resolved, so it does not leak through DNS. A name is resolved once; when any answer is not allowed the connection is refused, and the proxy dials the address it checked. The addresses of the node's own interfaces, the whole loopback range and the server's own addresses that LocalAI sends with each allow (see egress.check) are refused even when public. Only ports 80 and 443 by default (chats_egress_ports; 22 and 25 are always refused). HTTPS goes as CONNECT host:443: after the tunnel opens, the first TLS record must be a ClientHello whose server name is the CONNECT host, or the tunnel is closed (this stops a tunnel to one tenant of a CDN that talks to another). A CONNECT to another allowed port (an always-allowed host:8443, or a port added to chats_egress_ports) may carry any protocol, and the server may speak first, so the proxy waits for nothing; but when the first bytes the command sends are a TLS record, they must be a ClientHello whose server name is the CONNECT host too, or the tunnel is closed with the audit reason sni. Plain HTTP goes as an absolute http:// URL: each request of a kept-alive connection is decided on its own, the Proxy-* and hop-by-hop headers are dropped, a Host header that names another host is refused, and every request gets a new upstream connection. CONNECT to port 80, the origin form, https:// in absolute form and Upgrade are refused with 400. The proxy follows no redirect: the guest's client follows it with a new request, which is checked again.

Cards. When a command asks for a host that the chat's preset does not admit but the ceiling would, and a person watches the chat, LocalAI stores a card in the chat and the person sees it in the chat: A command in this chat wants to reach host:port (an international host name is checked and shown in its ASCII form, so a look-alike name such as pаypal.com with a Cyrillic а reads xn--pypal-4ve.com; a name with no valid ASCII form is refused), with Allow once (the rest of that command), Allow for this chat (a rule) and Deny (a deny rule, so the host does not ask again; it can be removed in the network panel). The connection waits for the answer for up to 60 seconds, and never past the call's own timeout, so pip goes on by itself after the click; without an answer it gets the 403 line with the user did not answer. A late answer still counts: Allow for this chat stores the rule, and Allow once lets the next call of the chat reach the host within ten minutes. A card nobody answered expires when the chat's turn ends; it stays visible as Not answered and can still be answered, with the same buttons, under the call that asked and in the network panel of the Tools menu. Ceiling off never shows a card; ceiling registries shows one only for a registry host (so a chat on Off can approve pypi.org); ceiling all for any public host. A host inside your network shows a card only with chats_egress_allow_private on, and only an administrator can allow it. A chat shows at most three cards at a time and one per host and port; a second connection to the same host waits on the same card. A card names the turn that asked it, and waits only while that turn holds the chat and its 60 seconds have not run out. When the frontend that ran a turn dies, its cards do not stay pending: they read as Not answered at once, do not count toward the three, and expire in the chat record when the reaper ends that turn, when the next turn starts, or when the next card is asked, whichever comes first. They can still be answered late, like any expired card. Unattended runs (triggers, connectors) and sub-agents get no card: the block is in their output. The card is written to the chat record before the answer is sent on the chat bus, and the waiting check reads it again every two seconds, so an answer given through any frontend reaches it within two seconds even when the bus message is lost. When the ceiling goes down while a card waits, the card is refused, and every answer to it gets 400 egress_above_ceiling (the same answer twice to an answered card is 204; another answer is 409 already_answered). A card the ceiling cannot admit shows no buttons: it says that network access is off (or limited) on this server, names who can change it, and gives an administrator a link to the Settings tab of the add-on that runs the commands (the network-policy host control). The Network row of the chat's Tools menu says the same (Off: the administrator has not allowed network access.) when the ceiling is off or below all.

What the model reads. The Sandbox bash says that the network is HTTP and HTTPS through the proxy only. When the Sandbox serves bash, each turn tells the model the preset in one line, such as Network for commands: package registries only (Python (PyPI), ...). A blocked host asks the user; do not try to get around the proxy. When no card can appear (a ceiling no wider than the preset, an unattended run, a sub-agent), the line says instead that a blocked host is refused and nothing asks the user, so the model names the host it needs in its answer. A blocked request gets 403 with one line of text in the tool output, for example Blocked by LocalAI: example.com:443 is not allowed for this chat (off). Ask the user to allow it. With ceiling off, or a chat preset above the ceiling, the note of the turn says that the server's setting keeps the network narrower, that no approval changes it, and that the model should tell the user instead of waiting or retrying. A request blocked for that reason gets the line blocked by the server's setting for the Sandbox network, not by a missing approval; a block that waits on a person keeps the the user did not answer line.

Only while a command runs. The proxy knows the call in progress of each sandbox. A connection while no call runs (a background process left by an earlier command) is blocked with the reason idle, nothing is asked, and it is audited at the next call. When a call ends, every tunnel it opened is closed. An allow verdict is cached for the call in progress for at most 30 seconds; a block is never cached.

Limits. These are fixed in this release:

Limit	Value
Concurrent connections of one sandbox	32 (the guest agent caps all its ports at 64)
New connections of one sandbox	20 per second, burst 40
Concurrent connections of the add-on process	256
Dial timeout	10 s
Idle tunnel (no bytes either way)	120 s
Bytes from the guest, per call	64 MiB
Bytes to the guest, per call	4 GiB

Audit. Every verdict is a line of the audit (who, which chat, host, port, kind, verdict, reason), and every allowed connection adds a done line with the address it reached, the bytes each way and how long it lasted. The reason names the rule that decided: preset, allowed_host (the administrator's always-allowed list), rule, once, denied, ceiling, port, private and the others the proxy adds, such as idle and sni. Lines older than chats_egress_audit_days (default 30) are removed daily, with the lines of the audit of the commands run on this machine.

What fails. Tools that ignore the proxy variables, SSH remotes (git@...), raw TCP and UDP, ICMP and DNS from the guest fail, as before.

Residual risks. The presets bound what a command reaches; they do not make every reachable host safe:

• A registry that takes uploads (npm publish with a token the command holds) can carry data out under "Package registries only", and so can any host under "Allow all". The preset is the person's choice of risk; the per-call upload cap bounds the volume, not the risk.
• Inside a TLS tunnel the Host header is encrypted and cannot be checked; a CDN that routes by it can still be reached through an allowed name.
• A command that sees the proxy can try many hosts; each one is checked and audited.

Distributed mode. Each piece of state has one mode:

State	Mode
Ceiling, private switch, ports, extra registries, default	shared: runtime settings, like the other chats_* settings
A chat's preset, rules and cards	shared and durable: the chat record in the chat store (the database in a cluster)
A card that waits	stateless per request: the host operation on the owner waits; it wakes on the answer's bus message or on a re-read of the card every two seconds
The decision of one connection	stateless per request: the host operation reads the settings and the chat record at each check
The map from a call's resource to its chat	per instance: a cache of a fixed mapping (a hash of the chat ID); each frontend fills its own, and a miss only costs a search of the user's chats
Audit	shared: the studio_egress_events table in a cluster, a JSON file under the data path, or memory without one
Audit pruning	single-runner: one frontend at each daily tick, under an advisory lock
Verdict cache, rate limits, connection counts	per instance: the add-on process on the owner; an allow lives at most 30 seconds and only within one call, so a settings change applies within 30 seconds

With --addons-role frontend a frontend runs no proxy; the owner does, and the host operations run on the owner's host.

Packaging and setting up the Sandbox add-on

The Sandbox package is for Linux on amd64 and arm64. The macOS platforms are listed in its platform file and stay disabled.

• runsc. addons/sandbox/package/runsc.json pins one gVisor release (20261012.0) with the URL, the checksum the release publishes for each tarball, and the checksum of the runsc inside it. addons/sandbox/package/fetch-runtime.sh <os/arch> <dir> downloads it, checks both checksums, and the release workflow puts runsc and gvisor-bin/ in the package. At run time the add-on checks runsc against the pinned checksum, copies it with its helpers to runsc/ in its data directory as executable files, and uses that copy. A runsc that does not match is never used. Without one in the package, a runsc in a system folder (/usr/local/sbin, /usr/local/bin, /usr/sbin, /usr/bin, /sbin, /bin) is used as it is. The add-on does not read PATH: LocalAI starts it with none, so it looks for runsc, taskset and the KVM tools in those folders only.

• AppArmor. On Ubuntu 24.04 and later, a program needs an AppArmor profile to make user namespaces. The package carries gvisor-profile, a small program that prints the profile of one runsc. The add-on copies it next to runsc, and when gVisor fails for this reason the page tells an administrator the exact step, with the real paths. Run it once, as root, on the machine that runs LocalAI (not inside a container):

  /path/to/data/addons/sandbox/runsc/gvisor-profile /path/to/data/addons/sandbox/runsc/runsc \
    | sudo tee /etc/apparmor.d/localai-runsc
  sudo apparmor_parser -r /etc/apparmor.d/localai-runsc
  

  The profile names the binary by path, so a new runsc at a new path needs a new profile. Inside a container, AppArmor is that of the host: use a profile that allows user namespaces for the container (or --security-opt apparmor=unconfined), together with a seccomp profile that allows unshare and clone with CLONE_NEWUSER.

• The guest image. The guest runs a Debian userland, the guest kernel and the guest agent. The sandbox guest image workflow (.github/workflows/sandbox-images.yml) builds addons/sandbox/images/guest/Dockerfile for amd64 and arm64, pushes it to the project registry (quay.io/go-skynet/sandbox-guest, which must allow anonymous pulls) and outputs the reference by digest. The package build passes that reference in SANDBOX_GUEST_IMAGE=<repository>@sha256:<digest>, and scripts/addons/package.sh builds the add-on with -ldflags "-X main.guestImage=...". package.sh refuses a Sandbox build without the variable; --local-check builds with a stand-in reference no registry serves, for a check that does not publish. To build the image by hand: docker buildx build -f addons/sandbox/images/guest/Dockerfile . from the repository root.

• KVM. addons/sandbox/package/kvm.json pins Cloud Hypervisor v54.0 (amd64 and arm64) and virtiofsd v1.14.0 (amd64) with checksums, and fetch-runtime.sh puts them in the package. The add-on checks each against the checksum compiled into it and copies it, as an executable file, to kvm-bin/ in its data directory. Upstream publishes no static virtiofsd for arm64, so on arm64 install virtiofsd 1.13 or later in a system folder. The KVM backend also needs mke2fs and debugfs from e2fsprogs (the LocalAI image installs them) and access to /dev/kvm for the user LocalAI runs as. Bubblewrap is optional in a container: the add-on does not use it when it finds /.dockerenv or /run/.containerenv, or when SANDBOX_NO_BWRAP=1 is set in its environment (SANDBOX_NO_BWRAP=0 forces it). addons/sandbox/images/host/Dockerfile still builds an image with the same programs for a product that wants them in one place.

• Tests in CI. The sandbox tests workflow (.github/workflows/sandbox-tests.yml) runs the Go tests and the end-to-end suite, the gVisor integration suite on amd64 and arm64 (with the pinned runsc and its profile loaded), the KVM integration suite on a runner that has /dev/kvm (it skips, and says so, on one without), and the release smoke test.

• Release smoke test. scripts/addons/sandbox_smoke.sh takes the package that package.sh made, starts the add-on from the extracted files under the real supervisor with the default settings, and runs a command through the tool path of a chat. It passes when the command ran in gVisor or KVM, or, on a host that cannot, when every tool answers "This host cannot run commands." and nothing ran. --require isolated makes the second case a failure.

The Sandbox in the LocalAI container

gVisor, KVM and the host backend refuse to run as root, so the sandbox needs the add-on process to run as an unprivileged user. The LocalAI container image starts as root only long enough to run entrypoint.sh, which hands the data folders (/models, /backends, /configuration, /data, /tmp/generated) to the user localai (uid and gid 10001) and then runs LocalAI as that user, with the groups that own /dev/kvm, /dev/kfd and /dev/dri/* added. The add-on runs as the same user.

• Existing volumes. A volume created by an older image is owned by root. On the first start with the new image the entrypoint changes its owner to uid 10001 (chown -R), which takes a while on a large models folder, and logs it. Nothing needs to be done by hand. If the volume cannot be changed (a read-only or network mount that refuses chown), set its owner to 10001:10001 on the storage side, or run the container with --user set to the owner of the data.
• Another user. Set LOCALAI_UID and LOCALAI_GID to use another id. A container that already starts as a non-root user (docker run --user, Kubernetes runAsUser) is left as it is; the data folders must then be writable by that user.
• Keep root. Set LOCALAI_RUN_AS_ROOT=true to run as root as before. The sandbox is then unavailable, and the page says why.
• Devices. Pass /dev/kvm to the container (--device /dev/kvm); the entrypoint adds the group that owns it. In Kubernetes, request the device through your device plugin and give the pod a fsGroup or supplementalGroups entry for it.
• User namespaces. gVisor needs the container to be allowed to create user namespaces, and on an Ubuntu 24.04 or later host an AppArmor profile that allows them (see AppArmor in Packaging and setting up the Sandbox add-on). Without them runsc stops at once, and the backend's reason ends with exit status 128 and cannot read client sync file ... EOF. Docker's default seccomp profile blocks this: use a profile that allows it, or --security-opt seccomp=unconfined on a host you trust. KVM needs /dev/kvm and no extra privileges.
• Cluster. The add-on runs on the add-on owner node only. In Kubernetes, give that pod the settings above; a node pool dedicated to it is the way to keep the isolation requirements off the other nodes.

What the source tree does not prove: the images and the runsc download are not built in the test suite, and the gVisor and KVM backends need a host that allows them. Their tests skip with the reason on a host that does not. The host backend and the whole path from a chat to a file in the user's files run in the tests.