Jamie Pine
35eb522e39
Absolute paths from one developer's machine sat in an example, a shell script, recorded benchmark metadata, task docs and design docs. The example and the script were broken for anyone else as a result: the example now resolves against CARGO_MANIFEST_DIR and the script derives the repo root from git. The duplicate copy of that script under core/scripts is gone. Also drops a pointer to a repository that is not public.
Spacedrive
One computer, made of all your machines.
Your files, your software, your agents — owned, portable, and coherent across every device.
docs.spacedrive.com • Discord • Files • Getting Started • Status
Your computer used to be one box, and you owned all of it. It held your files, ran your software, and answered only to you. That computer dissolved into devices, clouds and subscriptions, and you own almost none of the result.
Spacedrive puts it back together. All your machines, storage, software and agents behave as one computer that is yours: everything visible in one place, everything portable between machines, and everything still working the day you stop paying anyone — including us.
Files is that computer's file manager, built on a filesystem index that spans every device you own. It runs standalone: that is all most people come for, and it is enough on its own.
Files
The file manager of that computer, built on an index that Spacedrive maintains continuously. The watcher feeds it as files change, so it is already current when you look — nothing is built on demand while you wait.
Every surface is a view over that one index: grid, list, columns, media, size, recents, search, knowledge. A size analysis of a folder and a photo grid of the same folder are the same data through a different lens, and each view opens current because the index already is.
- One address for everything. Every file has an
SdPath, whether it lives on this machine, another of your devices, or a cloud volume. Operations work the same across all of them. - Content identity. BLAKE3 hashing gives the same file on two machines the same fingerprint — the basis for deduplication and redundancy tracking across your machines.
- Cloud volumes as first-class storage. S3, Google Drive, Dropbox, OneDrive, Azure Blob and GCS get indexed alongside local disks.
- P2P sync. Devices connect directly over Iroh and QUIC with no server in the middle. Metadata syncs, files stay where they are, and devices that are offline remain in the index.
- Search that works. SQLite FTS5 over files and archived records, roughly 55ms on a million entries, with indexing at around 8,500 files per second.
- Preview before commit. Copies, moves and deletes are simulated first, so you see conflicts and space changes before anything runs, and the operation becomes a durable job that survives restarts.
Files runs standalone: a desktop app on macOS, Windows and Linux, iOS and Android clients, a browser client, a headless server and a CLI. The macOS install is 44.4 MB with no bundled codecs and nothing to download. V2 is in alpha; the last stable download is v1.
It is machine-scoped by design. A host has one filesystem, so it gets one index, and everything else on the machine declares against that index as a capability instead of running an indexer of its own.
Search everything, not just files
Archive indexes external sources through script adapters, which puts email, notes, bookmarks and chat history in the same index as your files. Shipped adapters: Gmail, Apple Notes, Chrome Bookmarks, Chrome History, Safari History, Obsidian, OpenCode, Slack, macOS Contacts, macOS Calendar and GitHub.
Writing one means a directory with an adapter.toml and a script in any language
that reads stdin and prints lines. See
docs/archive/README.md.
Screening
When enabled, every record passes through a safety pipeline before becoming searchable: a local classifier (Prompt Guard 2) checks external content for prompt injection, trust tiers apply stricter screening to content you did not author, flagged records are quarantined out of agent queries, and search results carry trust metadata so an agent knows what is untrusted.
Where V2 came from
2021. A personal project: one interface over files scattered across devices and cloud accounts, with content-addressed identity so the same file on two machines is the same file.
May 2022. Open sourced. Number one on GitHub Trending for three days, 10,000 stars in the first week, a $2M seed round, a team of around twelve.
Early 2025. 35,000 stars, 600,000 installs, and a codebase that could no
longer ship. The Rust Prisma client was deprecated with no migration path, libp2p
transfers hung in ways nobody could debug, and two incompatible file models meant
every operation needed writing twice.
docs/overview/history.mdx is the full post-mortem.
2025. V2, rebuilt solo against that list. SeaORM for Prisma, Iroh for libp2p,
one SdPath for every file, jobs down to about 50 lines each, and search that
actually searches.
2026. The subtraction. V2 fixed V1's engineering and kept its appetite: every install compiled in a WASM extension runtime and shipped a bundled media stack, on the theory that a file manager should carry its whole environment. It should not, and both became opt-in:
| macOS payload before | ~156 MB (73.2 MB binary plus an 83 MB framework) |
| Default install now | 44.4 MB, one file, nothing to download |
Nothing was removed from the product. The heavy subsystems became opt-in
features, macOS thumbnails route through ImageIO and QuickLook instead of bundled
FFmpeg and libheif, and half the old binary turned out to be data rather than
code. Every number is measured; the method is in
docs/plans/2026-07-29-install-size.md.
Status
Under active development, and in alpha. V2 is a rebuild rather than a patch, so treat everything below as moving.
| Files: indexing, sync, cloud volumes, desktop, mobile and web clients | working |
| Archive: record spine, durable overlay, FTS5 search, 11 adapters | working |
| Machine-scoped daemon: supervision, health, lifecycle, port leases | working |
| Mounts: read-only share on loopback, byte-range peer protocol | in progress |
| Per-source stores and the in-memory read tier | in progress |
Architecture
spacedrive/
├── core/ # Rust engine (CQRS/DDD): indexing, sync, jobs, volumes
├── crates/ # archive, supervisor, task-system, crypto, imageio, images
├── apps/
│ ├── tauri/ mobile/ # Desktop (macOS, Windows, Linux), React Native
│ ├── cli/ server/ # CLI, daemon, headless server
│ ├── native/ # GPUI client prototype
│ └── web/ # Browser client
├── packages/
│ ├── interface/ # Files' React UI
│ └── ts-client/ swift-client/ assets/
└── adapters/ # Script-based data source adapters
Every core operation — a file copy, a tag create, a search query — is a registered action or query with type-safe input and output, and Specta generates the TypeScript and Swift clients from those definitions.
| Component | Technology |
|---|---|
| Engine | Rust, Tokio |
| Database | SQLite (SeaORM + sqlx) |
| Search | SQLite FTS5 |
| Content hashing | BLAKE3 |
| P2P | Iroh (QUIC, hole-punching, local discovery) |
| Cloud storage | OpenDAL (S3, Google Drive, Dropbox, OneDrive, Azure Blob, GCS) |
| Cryptography | Ed25519, X25519, ChaCha20-Poly1305, AES-GCM |
| Media | ImageIO and QuickLook on macOS, FFmpeg/libheif/Pdfium elsewhere |
| Desktop | Tauri 2 |
| Mobile | React Native + Expo |
| Frontend | React 19, Vite, TanStack Query, Tailwind CSS v4 |
| Design system | SpaceUI |
| Type generation | Specta (TypeScript + Swift) |
Getting Started
Requires Rust (stable, MSRV 1.81), Bun 1.3+, Node 20.x, and just.
git clone https://github.com/spacedriveapp/spacedrive
cd spacedrive
# System dependencies: cmake on macOS, GTK/WebKit/gstreamer on Linux.
# Pass `mobile` to also add the iOS/Android Rust targets.
./scripts/setup.sh # .\scripts\setup.ps1 on Windows
bun install
cargo run -p xtask -- setup
xtask setup writes .cargo/config.toml, builds the release daemon, and registers
the cargo xtask, cargo daemon and cargo cli aliases, so after the first run
just setup does the same thing in one command. It downloads nothing.
On macOS that is everything you need. Thumbnails for video, HEIC, RAW and PDF come
from ImageIO and QuickLook, the same system codecs Finder uses. Elsewhere those
formats need the prebuilt codec bundle (FFmpeg, libheif, Pdfium), which is a 91 MB
download and enables the ffmpeg and heif features:
cargo run -p xtask -- setup --native-deps
Run Files
just dev-desktop # desktop app (builds and starts the daemon for you)
just dev-server # headless server + web UI
just dev-mobile # Expo dev server
Test and check
just test # cargo test --workspace
just check # cargo fmt --check + clippy
Python 3.9+ is needed at runtime by the bundled adapters, not to build Spacedrive.
Your data, your machines
Self-hosting is first-class rather than a begrudged escape hatch. Spacedrive runs on your own hardware, and your devices federate directly with each other.
- Local-first. Nothing leaves your device unless you choose to sync between your own machines.
- No telemetry. No usage signal is collected without saying so, and none is required for the software to work.
- End-to-end encrypted. All P2P traffic runs over QUIC/TLS.
- Yours to leave. The index is rebuildable from your filesystem, and nothing you typed lives anywhere you cannot read it.
The mission
The open source software the world runs on is maintained on donation economics. The better it gets, the wider the gap between the value it creates and the money reaching the people who made it. This project has tens of thousands of stars and has returned a fraction of a salary.
Fixing that is a design constraint here rather than a substitute for building something people want.
Contributing
- Join Discord to chat with developers and community
- Contributing Guide
- Archive & Adapters, how data source adapters work
- SpaceUI, shared design system
License
Releases before 2026-03-24 were published under AGPL-3.0 and remain available under that license.