server: usage api

Add a new /api/usage endpoint that shows aggregate usage statistics per model since the server started.

api/types.go

@@ -912,6 +912,19 @@ type UserResponse struct {
 	Plan      string    `json:"plan,omitempty"`
 }
 
+type UsageResponse struct {
+	// Start is the time the server started tracking usage (UTC, RFC 3339).
+	Start time.Time        `json:"start"`
+	Usage []ModelUsageData `json:"usage"`
+}
+
+type ModelUsageData struct {
+	Model            string `json:"model"`
+	Requests         int64  `json:"requests"`
+	PromptTokens     int64  `json:"prompt_tokens"`
+	CompletionTokens int64  `json:"completion_tokens"`
+}
+
 // Tensor describes the metadata for a given tensor.
 type Tensor struct {
 	Name  string   `json:"name"`

cmd/config/claude.go

@@ -6,6 +6,8 @@ import (
 	"os/exec"
 	"path/filepath"
 	"runtime"
+
+	"github.com/ollama/ollama/envconfig"
 )
 
 // Claude implements Runner for Claude Code integration
@@ -50,7 +52,7 @@ func (c *Claude) Run(model string) error {
 	cmd.Stdout = os.Stdout
 	cmd.Stderr = os.Stderr
 	cmd.Env = append(os.Environ(),
-		"ANTHROPIC_BASE_URL=http://localhost:11434",
+		"ANTHROPIC_BASE_URL="+envconfig.Host().String(),
 		"ANTHROPIC_API_KEY=",
 		"ANTHROPIC_AUTH_TOKEN=ollama",
 	)

cmd/config/clawdbot.go

@@ -9,6 +9,8 @@ import (
 	"os/exec"
 	"path/filepath"
 	"strings"
+
+	"github.com/ollama/ollama/envconfig"
 )
 
 type Clawdbot struct{}
@@ -90,7 +92,7 @@ func (c *Clawdbot) Edit(models []string) error {
 		ollama = make(map[string]any)
 	}
 
-	ollama["baseUrl"] = "http://127.0.0.1:11434/v1"
+	ollama["baseUrl"] = envconfig.Host().String() + "/v1"
 	// needed to register provider
 	ollama["apiKey"] = "ollama-local"
 	// TODO(parthsareen): potentially move to responses

cmd/config/droid.go

@@ -7,6 +7,8 @@ import (
 	"os/exec"
 	"path/filepath"
 	"slices"
+
+	"github.com/ollama/ollama/envconfig"
 )
 
 // Droid implements Runner and Editor for Droid integration
@@ -117,7 +119,7 @@ func (d *Droid) Edit(models []string) error {
 		newModels = append(newModels, modelEntry{
 			Model:           model,
 			DisplayName:     model,
-			BaseURL:         "http://localhost:11434/v1",
+			BaseURL:         envconfig.Host().String() + "/v1",
 			APIKey:          "ollama",
 			Provider:        "generic-chat-completion-api",
 			MaxOutputTokens: 64000,

cmd/config/droid_test.go

@@ -218,7 +218,7 @@ func TestDroidEdit(t *testing.T) {
 			}
 		}
 
-		if model["baseUrl"] != "http://localhost:11434/v1" {
+		if model["baseUrl"] != "http://127.0.0.1:11434/v1" {
 			t.Errorf("unexpected baseUrl: %s", model["baseUrl"])
 		}
 		if model["apiKey"] != "ollama" {
@@ -447,7 +447,7 @@ const testDroidSettingsFixture = `{
     {
       "model": "existing-ollama-model",
       "displayName": "existing-ollama-model",
-      "baseUrl": "http://localhost:11434/v1",
+      "baseUrl": "http://127.0.0.1:11434/v1",
       "apiKey": "ollama",
       "provider": "generic-chat-completion-api",
       "maxOutputTokens": 64000,

cmd/config/opencode.go

@@ -9,6 +9,8 @@ import (
 	"path/filepath"
 	"slices"
 	"strings"
+
+	"github.com/ollama/ollama/envconfig"
 )
 
 // OpenCode implements Runner and Editor for OpenCode integration
@@ -88,7 +90,7 @@ func (o *OpenCode) Edit(modelList []string) error {
 			"npm":  "@ai-sdk/openai-compatible",
 			"name": "Ollama (local)",
 			"options": map[string]any{
-				"baseURL": "http://localhost:11434/v1",
+				"baseURL": envconfig.Host().String() + "/v1",
 			},
 		}
 	}

docs/api.md

@@ -15,6 +15,7 @@
 - [Push a Model](#push-a-model)
 - [Generate Embeddings](#generate-embeddings)
 - [List Running Models](#list-running-models)
+- [Usage](#usage)
 - [Version](#version)
 - [Experimental: Image Generation](#image-generation-experimental)
 
@@ -1854,6 +1855,53 @@ curl http://localhost:11434/api/embeddings -d '{
 }
 ```
 
+## Usage
+
+```
+GET /api/usage
+```
+
+Show aggregate usage statistics per model since the server started. All timestamps are UTC in RFC 3339 format.
+
+### Examples
+
+#### Request
+
+```shell
+curl http://localhost:11434/api/usage
+```
+
+#### Response
+
+```json
+{
+  "start": "2025-01-27T20:00:00Z",
+  "usage": [
+    {
+      "model": "llama3.2",
+      "requests": 5,
+      "prompt_tokens": 130,
+      "completion_tokens": 890
+    },
+    {
+      "model": "deepseek-r1",
+      "requests": 2,
+      "prompt_tokens": 48,
+      "completion_tokens": 312
+    }
+  ]
+}
+```
+
+#### Response fields
+
+- `start`: when the server started tracking usage (UTC, RFC 3339)
+- `usage`: list of per-model usage statistics
+  - `model`: model name
+  - `requests`: total number of completed requests
+  - `prompt_tokens`: total prompt tokens evaluated
+  - `completion_tokens`: total completion tokens generated
+
 ## Version
 
 ```

docs/docs.json

@@ -102,6 +102,7 @@
             "group": "Integrations",
             "pages": [
               "/integrations/claude-code",
+              "/integrations/clawdbot",
               "/integrations/cline",
               "/integrations/codex",
               "/integrations/droid",

docs/integrations/clawdbot.mdx

@@ -0,0 +1,48 @@
+---
+title: Clawdbot
+---
+
+Clawdbot is a personal AI assistant that runs on your own devices. It bridges messaging services (WhatsApp, Telegram, Slack, Discord, iMessage, and more) to AI coding agents through a centralized gateway.
+
+## Install
+
+Install [Clawdbot](https://clawd.bot/) 
+
+```bash
+npm install -g clawdbot@latest
+```
+
+Then run the onboarding wizard:
+
+```bash
+clawdbot onboard --install-daemon
+```
+
+<Note>Clawdbot requires a larger context window. It is recommended to use a context window of at least 64k tokens. See [Context length](/context-length) for more information.</Note>
+
+## Usage with Ollama
+
+### Quick setup
+
+```bash
+ollama launch clawdbot
+```
+
+This configures Clawdbot to use Ollama and starts the gateway.
+If the gateway is already running, no changes need to be made as the gateway will auto-reload the changes. 
+
+
+To configure without launching:
+
+```shell
+ollama launch clawdbot --config
+```
+
+## Recommended Models
+
+- `qwen3-coder`
+- `glm-4.7`
+- `gpt-oss:20b`
+- `gpt-oss:120b`
+
+Cloud models are also available at [ollama.com/search?c=cloud](https://ollama.com/search?c=cloud).

model/parsers/ministral.go

@@ -4,6 +4,7 @@ import (
 	"encoding/json"
 	"fmt"
 	"strings"
+	"unicode"
 
 	"github.com/ollama/ollama/api"
 )
@@ -17,12 +18,34 @@ const (
 	ministralCollectingToolArgs
 )
 
+// ministralEvent represents an event emitted during parsing
+type ministralEvent interface {
+	isMinistralEvent()
+}
+
+type ministralEventContent struct {
+	content string
+}
+
+type ministralEventThinking struct {
+	thinking string
+}
+
+type ministralEventToolCall struct {
+	name string
+	args string // raw JSON string
+}
+
+func (ministralEventContent) isMinistralEvent()  {}
+func (ministralEventThinking) isMinistralEvent() {}
+func (ministralEventToolCall) isMinistralEvent() {}
+
 type MinistralParser struct {
 	state              ministralParserState
 	buffer             strings.Builder
 	tools              []api.Tool
 	hasThinkingSupport bool
-	currentTool        *api.Tool
+	pendingToolName    string // stores tool name while collecting args
 }
 
 func (p *MinistralParser) HasToolSupport() bool {
@@ -63,74 +86,251 @@ func toolByName(tools []api.Tool, n string) (*api.Tool, error) {
 	return nil, fmt.Errorf("tool '%s' not found", n)
 }
 
-func (p *MinistralParser) Add(s string, done bool) (content string, thinking string, calls []api.ToolCall, err error) {
-	p.buffer.WriteString(s)
+const (
+	ministralToolCallsTag = "[TOOL_CALLS]"
+	ministralThinkTag     = "[THINK]"
+	ministralThinkEndTag  = "[/THINK]"
+	ministralArgsTag      = "[ARGS]"
+)
+
+// eat consumes the parser's buffer, and returns a list of any unambiguous
+// events from the current parser state. The second return value indicates
+// whether to keep looping (true when state transitions, false when waiting
+// for more data).
+func (p *MinistralParser) eat() ([]ministralEvent, bool) {
+	var events []ministralEvent
 
 	switch p.state {
 	case ministralCollectingContent:
-		if strings.Contains(p.buffer.String(), "[TOOL_CALLS]") {
-			before, _ := splitAtTag(&p.buffer, "[TOOL_CALLS]", false)
-			if before != "" {
-				return before, "", calls, nil
+		bufStr := p.buffer.String()
+
+		// Check for [TOOL_CALLS] tag
+		if strings.Contains(bufStr, ministralToolCallsTag) {
+			split := strings.SplitN(bufStr, ministralToolCallsTag, 2)
+			before := strings.TrimRightFunc(split[0], unicode.IsSpace)
+			if len(before) > 0 {
+				events = append(events, ministralEventContent{content: before})
 			}
+			after := split[1]
+			p.buffer.Reset()
+			p.buffer.WriteString(after)
 			p.state = ministralCollectingToolName
-		} else if strings.Contains(p.buffer.String(), "[THINK]") {
+			return events, true
+		}
+
+		// Check for [THINK] tag
+		if strings.Contains(bufStr, ministralThinkTag) {
+			split := strings.SplitN(bufStr, ministralThinkTag, 2)
+			before := strings.TrimRightFunc(split[0], unicode.IsSpace)
+			if len(before) > 0 {
+				events = append(events, ministralEventContent{content: before})
+			}
+			after := split[1]
+			p.buffer.Reset()
+			p.buffer.WriteString(after)
 			p.state = ministralCollectingThinkingContent
-			return "", "", calls, nil
-		} else {
-			p.buffer.Reset()
-			return s, "", calls, nil
+			return events, true
 		}
+
+		// Check for partial tag overlap with [TOOL_CALLS] or [THINK]
+		overlapToolCalls := overlap(bufStr, ministralToolCallsTag)
+		overlapThink := overlap(bufStr, ministralThinkTag)
+		maxOverlap := max(overlapToolCalls, overlapThink)
+
+		if maxOverlap > 0 {
+			// Withhold the potential partial tag
+			beforePartialTag := bufStr[:len(bufStr)-maxOverlap]
+			trailingWS := trailingWhitespaceLen(beforePartialTag)
+			ambiguousStart := len(beforePartialTag) - trailingWS
+			unambiguous := bufStr[:ambiguousStart]
+			ambiguous := bufStr[ambiguousStart:]
+			p.buffer.Reset()
+			p.buffer.WriteString(ambiguous)
+			if len(unambiguous) > 0 {
+				events = append(events, ministralEventContent{content: unambiguous})
+			}
+			return events, false
+		}
+
+		// No tag found: emit content but withhold trailing whitespace
+		whitespaceLen := trailingWhitespaceLen(bufStr)
+		ambiguousStart := len(bufStr) - whitespaceLen
+		unambiguous := bufStr[:ambiguousStart]
+		ambiguous := bufStr[ambiguousStart:]
+		p.buffer.Reset()
+		p.buffer.WriteString(ambiguous)
+		if len(unambiguous) > 0 {
+			events = append(events, ministralEventContent{content: unambiguous})
+		}
+		return events, false
+
 	case ministralCollectingThinkingContent:
-		if strings.Contains(p.buffer.String(), "[/THINK]") {
-			thinkingContent, after := splitAtTag(&p.buffer, "[/THINK]", true)
-			p.state = ministralCollectingContent
-			if after != "" {
-				p.buffer.Reset()
-				return after, thinkingContent, calls, nil
-			}
-			return "", thinkingContent, calls, nil
-		} else {
+		bufStr := p.buffer.String()
+
+		if strings.Contains(bufStr, ministralThinkEndTag) {
+			split := strings.SplitN(bufStr, ministralThinkEndTag, 2)
+			thinkingContent := split[0]
+			after := strings.TrimLeftFunc(split[1], unicode.IsSpace)
 			p.buffer.Reset()
-			return "", s, calls, nil
-		}
-	case ministralCollectingToolName:
-		if strings.Contains(p.buffer.String(), "[ARGS]") {
-			name, _ := splitAtTag(&p.buffer, "[ARGS]", false)
-
-			t, err := toolByName(p.tools, name)
-			if err != nil {
-				return "", "", calls, err
+			p.buffer.WriteString(after)
+			if len(thinkingContent) > 0 {
+				events = append(events, ministralEventThinking{thinking: thinkingContent})
 			}
-			p.currentTool = t
-			p.state = ministralCollectingToolArgs
-			return "", "", calls, nil
-		}
-		return "", "", calls, nil
-	case ministralCollectingToolArgs:
-		if strings.Contains(p.buffer.String(), "}") {
-			before, _ := splitAtTag(&p.buffer, "}", false)
-			before += "}"
-
-			var args api.ToolCallFunctionArguments
-			if err := json.Unmarshal([]byte(before), &args); err != nil {
-				// todo - throw a better error
-				return "", "", calls, err
-			}
-
 			p.state = ministralCollectingContent
+			return events, true
+		}
 
-			call := api.ToolCall{
+		// Check for partial overlap with [/THINK]
+		if overlapLen := overlap(bufStr, ministralThinkEndTag); overlapLen > 0 {
+			unambiguous := bufStr[:len(bufStr)-overlapLen]
+			ambiguous := bufStr[len(bufStr)-overlapLen:]
+			p.buffer.Reset()
+			p.buffer.WriteString(ambiguous)
+			if len(unambiguous) > 0 {
+				events = append(events, ministralEventThinking{thinking: unambiguous})
+			}
+			return events, false
+		}
+
+		// No tag found: emit all thinking content
+		p.buffer.Reset()
+		if len(bufStr) > 0 {
+			events = append(events, ministralEventThinking{thinking: bufStr})
+		}
+		return events, false
+
+	case ministralCollectingToolName:
+		bufStr := p.buffer.String()
+
+		if strings.Contains(bufStr, ministralArgsTag) {
+			split := strings.SplitN(bufStr, ministralArgsTag, 2)
+			toolName := split[0]
+			after := split[1]
+			p.pendingToolName = toolName
+			p.buffer.Reset()
+			p.buffer.WriteString(after)
+			p.state = ministralCollectingToolArgs
+			return events, true
+		}
+		// Wait for more data
+		return events, false
+
+	case ministralCollectingToolArgs:
+		bufStr := p.buffer.String()
+		jsonEnd := findJSONEnd(bufStr)
+
+		if jsonEnd != -1 {
+			jsonStr := bufStr[:jsonEnd+1]
+			remaining := bufStr[jsonEnd+1:]
+
+			events = append(events, ministralEventToolCall{
+				name: p.pendingToolName,
+				args: jsonStr,
+			})
+
+			p.pendingToolName = ""
+			p.buffer.Reset()
+			p.buffer.WriteString(remaining)
+			p.state = ministralCollectingContent
+			return events, true
+		}
+		// Wait for more data
+		return events, false
+
+	default:
+		panic("unexpected ministral event")
+	}
+}
+
+// parseEvents loops calling eat() until it returns false
+func (p *MinistralParser) parseEvents() []ministralEvent {
+	var all []ministralEvent
+	keepLooping := true
+	for keepLooping {
+		var events []ministralEvent
+		events, keepLooping = p.eat()
+		all = append(all, events...)
+	}
+	return all
+}
+
+func (p *MinistralParser) Add(s string, done bool) (content string, thinking string, calls []api.ToolCall, err error) {
+	p.buffer.WriteString(s)
+
+	events := p.parseEvents()
+
+	var contentBuilder, thinkingBuilder strings.Builder
+	var toolCalls []api.ToolCall
+
+	for _, event := range events {
+		switch e := event.(type) {
+		case ministralEventContent:
+			contentBuilder.WriteString(e.content)
+		case ministralEventThinking:
+			thinkingBuilder.WriteString(e.thinking)
+		case ministralEventToolCall:
+			// Validate tool exists
+			tool, toolErr := toolByName(p.tools, e.name)
+			if toolErr != nil {
+				return contentBuilder.String(), thinkingBuilder.String(), toolCalls, toolErr
+			}
+			// Parse JSON arguments
+			var args api.ToolCallFunctionArguments
+			if jsonErr := json.Unmarshal([]byte(e.args), &args); jsonErr != nil {
+				return contentBuilder.String(), thinkingBuilder.String(), toolCalls, jsonErr
+			}
+			toolCalls = append(toolCalls, api.ToolCall{
 				Function: api.ToolCallFunction{
-					Name:      p.currentTool.Function.Name,
+					Name:      tool.Function.Name,
 					Arguments: args,
 				},
-			}
-			calls = append(calls, call)
-			return "", "", calls, nil
+			})
 		}
-		return "", "", calls, nil
 	}
 
-	return p.buffer.String(), thinking, calls, nil
+	return contentBuilder.String(), thinkingBuilder.String(), toolCalls, nil
+}
+
+// findJSONEnd finds the index of the closing brace that completes a JSON object.
+// It properly handles nested objects, arrays, and strings (including escaped characters).
+// Returns -1 if the JSON is not yet complete.
+func findJSONEnd(s string) int {
+	depth := 0
+	inString := false
+	escaped := false
+
+	for i, r := range s {
+		if inString {
+			switch {
+			case escaped:
+				// If the previous character was a backslash, skip this character
+				escaped = false
+			case r == '\\':
+				// Mark the next character as escaped
+				escaped = true
+			case r == '"':
+				// End of string literal
+				inString = false
+			}
+			continue
+		}
+
+		switch r {
+		case '"':
+			// Start of string literal
+			inString = true
+		case '{', '[':
+			// Increase nesting level for objects and arrays
+			depth++
+		case '}', ']':
+			// Decrease nesting level
+			depth--
+			if depth == 0 {
+				// Reached the end of the root JSON structure
+				return i
+			}
+		}
+	}
+
+	return -1
 }

model/parsers/ministral_test.go

@@ -0,0 +1,545 @@
+package parsers
+
+import (
+	"reflect"
+	"testing"
+
+	"github.com/ollama/ollama/api"
+)
+
+func TestMinistralParserStreaming(t *testing.T) {
+	type step struct {
+		input      string
+		wantEvents []ministralEvent
+	}
+
+	cases := []struct {
+		desc  string
+		tools []api.Tool
+		steps []step
+		think bool // whether to enable thinking support
+	}{
+		// Content streaming
+		{
+			desc: "simple content",
+			steps: []step{
+				{input: "Hello, how can I help you?", wantEvents: []ministralEvent{
+					ministralEventContent{content: "Hello, how can I help you?"},
+				}},
+			},
+		},
+		{
+			desc: "streaming content word by word",
+			steps: []step{
+				{input: "Hello,", wantEvents: []ministralEvent{ministralEventContent{content: "Hello,"}}},
+				{input: " how", wantEvents: []ministralEvent{ministralEventContent{content: " how"}}},
+				{input: " can I help?", wantEvents: []ministralEvent{ministralEventContent{content: " can I help?"}}},
+			},
+		},
+
+		// Simple tool calls
+		{
+			desc:  "simple tool call",
+			tools: []api.Tool{{Function: api.ToolFunction{Name: "get_weather"}}},
+			steps: []step{
+				{input: `[TOOL_CALLS]get_weather[ARGS]{"location": "San Francisco"}`, wantEvents: []ministralEvent{
+					ministralEventToolCall{name: "get_weather", args: `{"location": "San Francisco"}`},
+				}},
+			},
+		},
+		{
+			desc:  "tool call with nested object",
+			tools: []api.Tool{{Function: api.ToolFunction{Name: "create_entities"}}},
+			steps: []step{
+				{input: `[TOOL_CALLS]create_entities[ARGS]{"entities": [{"entityType": "Person", "name": "Jack", "observations": ["Works as a baker"]}]}`, wantEvents: []ministralEvent{
+					ministralEventToolCall{name: "create_entities", args: `{"entities": [{"entityType": "Person", "name": "Jack", "observations": ["Works as a baker"]}]}`},
+				}},
+			},
+		},
+		{
+			desc:  "tool call with deeply nested objects",
+			tools: []api.Tool{{Function: api.ToolFunction{Name: "update_config"}}},
+			steps: []step{
+				{input: `[TOOL_CALLS]update_config[ARGS]{"settings": {"user": {"profile": {"name": "John", "age": 30}}, "theme": "dark"}}`, wantEvents: []ministralEvent{
+					ministralEventToolCall{name: "update_config", args: `{"settings": {"user": {"profile": {"name": "John", "age": 30}}, "theme": "dark"}}`},
+				}},
+			},
+		},
+		{
+			desc:  "tool call with array of objects",
+			tools: []api.Tool{{Function: api.ToolFunction{Name: "process_items"}}},
+			steps: []step{
+				{input: `[TOOL_CALLS]process_items[ARGS]{"items": [{"id": 1}, {"id": 2}, {"id": 3}]}`, wantEvents: []ministralEvent{
+					ministralEventToolCall{name: "process_items", args: `{"items": [{"id": 1}, {"id": 2}, {"id": 3}]}`},
+				}},
+			},
+		},
+		{
+			desc:  "tool call with escaped quotes in string",
+			tools: []api.Tool{{Function: api.ToolFunction{Name: "search"}}},
+			steps: []step{
+				{input: `[TOOL_CALLS]search[ARGS]{"query": "say \"hello\""}`, wantEvents: []ministralEvent{
+					ministralEventToolCall{name: "search", args: `{"query": "say \"hello\""}`},
+				}},
+			},
+		},
+		{
+			desc:  "tool call with braces inside string",
+			tools: []api.Tool{{Function: api.ToolFunction{Name: "format"}}},
+			steps: []step{
+				{input: `[TOOL_CALLS]format[ARGS]{"template": "Hello {name}!"}`, wantEvents: []ministralEvent{
+					ministralEventToolCall{name: "format", args: `{"template": "Hello {name}!"}`},
+				}},
+			},
+		},
+		{
+			desc:  "empty JSON object",
+			tools: []api.Tool{{Function: api.ToolFunction{Name: "no_args"}}},
+			steps: []step{
+				{input: `[TOOL_CALLS]no_args[ARGS]{}`, wantEvents: []ministralEvent{
+					ministralEventToolCall{name: "no_args", args: `{}`},
+				}},
+			},
+		},
+		{
+			desc:  "JSON with newlines in string",
+			tools: []api.Tool{{Function: api.ToolFunction{Name: "write"}}},
+			steps: []step{
+				{input: `[TOOL_CALLS]write[ARGS]{"content": "line1\nline2\nline3"}`, wantEvents: []ministralEvent{
+					ministralEventToolCall{name: "write", args: `{"content": "line1\nline2\nline3"}`},
+				}},
+			},
+		},
+		{
+			desc:  "backslash in string value",
+			tools: []api.Tool{{Function: api.ToolFunction{Name: "path"}}},
+			steps: []step{
+				{input: `[TOOL_CALLS]path[ARGS]{"dir": "C:\\Users\\test"}`, wantEvents: []ministralEvent{
+					ministralEventToolCall{name: "path", args: `{"dir": "C:\\Users\\test"}`},
+				}},
+			},
+		},
+
+		// Content after tool call
+		{
+			desc:  "content after tool call",
+			tools: []api.Tool{{Function: api.ToolFunction{Name: "test"}}},
+			steps: []step{
+				// NOTE: It's unclear if this is valid Ministral output, but the parser
+				// currently treats text after a tool call as regular content. This test
+				// documents that behavior so we notice if it changes.
+				{input: `[TOOL_CALLS]test[ARGS]{"a": 1}some content after`, wantEvents: []ministralEvent{
+					ministralEventToolCall{name: "test", args: `{"a": 1}`},
+					ministralEventContent{content: "some content after"},
+				}},
+			},
+		},
+
+		// Multiple tool calls
+		{
+			desc: "multiple tool calls in sequence",
+			tools: []api.Tool{
+				{Function: api.ToolFunction{Name: "get_weather"}},
+				{Function: api.ToolFunction{Name: "get_time"}},
+			},
+			steps: []step{
+				{input: `[TOOL_CALLS]get_weather[ARGS]{"location": "NYC"}[TOOL_CALLS]get_time[ARGS]{"timezone": "EST"}`, wantEvents: []ministralEvent{
+					ministralEventToolCall{name: "get_weather", args: `{"location": "NYC"}`},
+					ministralEventToolCall{name: "get_time", args: `{"timezone": "EST"}`},
+				}},
+			},
+		},
+		{
+			desc: "multiple tool calls streamed separately",
+			tools: []api.Tool{
+				{Function: api.ToolFunction{Name: "tool_a"}},
+				{Function: api.ToolFunction{Name: "tool_b"}},
+			},
+			steps: []step{
+				{input: `[TOOL_CALLS]tool_a[ARGS]{"x": 1}`, wantEvents: []ministralEvent{
+					ministralEventToolCall{name: "tool_a", args: `{"x": 1}`},
+				}},
+				{input: `[TOOL_CALLS]tool_b[ARGS]{"y": 2}`, wantEvents: []ministralEvent{
+					ministralEventToolCall{name: "tool_b", args: `{"y": 2}`},
+				}},
+			},
+		},
+
+		// Streaming tool calls
+		{
+			desc:  "streaming tool call with nested objects",
+			tools: []api.Tool{{Function: api.ToolFunction{Name: "create_entities"}}},
+			steps: []step{
+				{input: "[TOOL_CALLS]create_entities[ARGS]", wantEvents: []ministralEvent{}},
+				{input: `{"entities": [{"entityType": "Person",`, wantEvents: []ministralEvent{}},
+				{input: ` "name": "Jack",`, wantEvents: []ministralEvent{}},
+				{input: ` "observations": ["Works`, wantEvents: []ministralEvent{}},
+				{input: ` as a baker"]}`, wantEvents: []ministralEvent{}},
+				{input: `]}`, wantEvents: []ministralEvent{
+					ministralEventToolCall{name: "create_entities", args: `{"entities": [{"entityType": "Person", "name": "Jack", "observations": ["Works as a baker"]}]}`},
+				}},
+			},
+		},
+		{
+			desc:  "streaming with incomplete JSON waits for completion",
+			tools: []api.Tool{{Function: api.ToolFunction{Name: "test"}}},
+			steps: []step{
+				{input: "[TOOL_CALLS]test[ARGS]{", wantEvents: []ministralEvent{}},
+				{input: `"a": {`, wantEvents: []ministralEvent{}},
+				{input: `"b": 1`, wantEvents: []ministralEvent{}},
+				{input: `}`, wantEvents: []ministralEvent{}},
+				{input: `}`, wantEvents: []ministralEvent{
+					ministralEventToolCall{name: "test", args: `{"a": {"b": 1}}`},
+				}},
+			},
+		},
+
+		// Partial tag handling
+		{
+			desc: "partial tool tag fakeout",
+			steps: []step{
+				{input: "abc[TOOL", wantEvents: []ministralEvent{ministralEventContent{content: "abc"}}},
+				{input: " not a tag", wantEvents: []ministralEvent{ministralEventContent{content: "[TOOL not a tag"}}},
+			},
+		},
+		{
+			desc:  "tool call tag split across chunks",
+			tools: []api.Tool{{Function: api.ToolFunction{Name: "test"}}},
+			steps: []step{
+				{input: "[TOOL_", wantEvents: []ministralEvent{}},
+				{input: "CALLS]test[ARGS]{}", wantEvents: []ministralEvent{
+					ministralEventToolCall{name: "test", args: `{}`},
+				}},
+			},
+		},
+		{
+			desc:  "content before tool call",
+			tools: []api.Tool{{Function: api.ToolFunction{Name: "get_weather"}}},
+			steps: []step{
+				{input: "hello [TOOL_CALLS]get_weather[ARGS]{}", wantEvents: []ministralEvent{
+					ministralEventContent{content: "hello"},
+					ministralEventToolCall{name: "get_weather", args: `{}`},
+				}},
+			},
+		},
+		{
+			desc:  "whitespace between content and tool call is trimmed",
+			tools: []api.Tool{{Function: api.ToolFunction{Name: "test"}}},
+			steps: []step{
+				{input: "content \n [TOOL_CALLS]test[ARGS]{}", wantEvents: []ministralEvent{
+					ministralEventContent{content: "content"},
+					ministralEventToolCall{name: "test", args: `{}`},
+				}},
+			},
+		},
+		{
+			desc:  "tabs and newlines before tool call are trimmed",
+			tools: []api.Tool{{Function: api.ToolFunction{Name: "test"}}},
+			steps: []step{
+				{input: "content\t\n\t[TOOL_CALLS]test[ARGS]{}", wantEvents: []ministralEvent{
+					ministralEventContent{content: "content"},
+					ministralEventToolCall{name: "test", args: `{}`},
+				}},
+			},
+		},
+		{
+			desc:  "non-breaking space before tool call is trimmed",
+			tools: []api.Tool{{Function: api.ToolFunction{Name: "test"}}},
+			steps: []step{
+				// \u00a0 is non-breaking space, which unicode.IsSpace considers whitespace
+				{input: "content\u00a0[TOOL_CALLS]test[ARGS]{}", wantEvents: []ministralEvent{
+					ministralEventContent{content: "content"},
+					ministralEventToolCall{name: "test", args: `{}`},
+				}},
+			},
+		},
+		{
+			desc: "whitespace before THINK tag is trimmed",
+			steps: []step{
+				{input: "content \n [THINK]thinking[/THINK]after", wantEvents: []ministralEvent{
+					ministralEventContent{content: "content"},
+					ministralEventThinking{thinking: "thinking"},
+					ministralEventContent{content: "after"},
+				}},
+			},
+		},
+		{
+			desc: "trailing whitespace withheld then emitted",
+			steps: []step{
+				{input: "Hello ", wantEvents: []ministralEvent{ministralEventContent{content: "Hello"}}},
+				{input: "world", wantEvents: []ministralEvent{ministralEventContent{content: " world"}}},
+			},
+		},
+		{
+			desc: "trailing newline withheld then emitted",
+			steps: []step{
+				{input: "Hello\n", wantEvents: []ministralEvent{ministralEventContent{content: "Hello"}}},
+				{input: "world", wantEvents: []ministralEvent{ministralEventContent{content: "\nworld"}}},
+			},
+		},
+
+		// Thinking support
+		{
+			desc:  "thinking content",
+			think: true,
+			steps: []step{
+				{input: "thinking here[/THINK]", wantEvents: []ministralEvent{
+					ministralEventThinking{thinking: "thinking here"},
+				}},
+				{input: "content after", wantEvents: []ministralEvent{
+					ministralEventContent{content: "content after"},
+				}},
+			},
+		},
+		{
+			desc:  "thinking with whitespace after end tag",
+			think: true,
+			steps: []step{
+				{input: "my thoughts[/THINK]  \n  response", wantEvents: []ministralEvent{
+					ministralEventThinking{thinking: "my thoughts"},
+					ministralEventContent{content: "response"},
+				}},
+			},
+		},
+		{
+			desc:  "non-breaking space after think end tag is trimmed",
+			think: true,
+			steps: []step{
+				// \u00a0 is non-breaking space
+				{input: "thinking[/THINK]\u00a0response", wantEvents: []ministralEvent{
+					ministralEventThinking{thinking: "thinking"},
+					ministralEventContent{content: "response"},
+				}},
+			},
+		},
+		{
+			desc:  "partial think end tag",
+			think: true,
+			steps: []step{
+				{input: "thinking[/THI", wantEvents: []ministralEvent{ministralEventThinking{thinking: "thinking"}}},
+				{input: "NK]after", wantEvents: []ministralEvent{ministralEventContent{content: "after"}}},
+			},
+		},
+		{
+			desc:  "think tag fakeout",
+			think: true,
+			steps: []step{
+				{input: "thinking[/THI", wantEvents: []ministralEvent{ministralEventThinking{thinking: "thinking"}}},
+				{input: "not end tag", wantEvents: []ministralEvent{ministralEventThinking{thinking: "[/THInot end tag"}}},
+			},
+		},
+		{
+			desc:  "thinking then tool call",
+			think: true,
+			tools: []api.Tool{{Function: api.ToolFunction{Name: "test"}}},
+			steps: []step{
+				{input: "let me think[/THINK][TOOL_CALLS]test[ARGS]{}", wantEvents: []ministralEvent{
+					ministralEventThinking{thinking: "let me think"},
+					ministralEventToolCall{name: "test", args: `{}`},
+				}},
+			},
+		},
+
+		// Content then THINK tag transition
+		{
+			desc: "content then think tag",
+			steps: []step{
+				{input: "content[THINK]thinking[/THINK]more", wantEvents: []ministralEvent{
+					ministralEventContent{content: "content"},
+					ministralEventThinking{thinking: "thinking"},
+					ministralEventContent{content: "more"},
+				}},
+			},
+		},
+
+		// Unicode handling
+		{
+			desc: "unicode content",
+			steps: []step{
+				{input: "你好 🌍 مرحبا", wantEvents: []ministralEvent{
+					ministralEventContent{content: "你好 🌍 مرحبا"},
+				}},
+			},
+		},
+		{
+			desc:  "unicode in tool args",
+			tools: []api.Tool{{Function: api.ToolFunction{Name: "greet"}}},
+			steps: []step{
+				{input: `[TOOL_CALLS]greet[ARGS]{"message": "你好 🌍"}`, wantEvents: []ministralEvent{
+					ministralEventToolCall{name: "greet", args: `{"message": "你好 🌍"}`},
+				}},
+			},
+		},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.desc, func(t *testing.T) {
+			parser := MinistralParser{}
+			parser.hasThinkingSupport = tc.think
+			parser.Init(tc.tools, nil, nil)
+
+			for i, step := range tc.steps {
+				parser.buffer.WriteString(step.input)
+				gotEvents := parser.parseEvents()
+
+				if len(gotEvents) == 0 && len(step.wantEvents) == 0 {
+					// avoid deep equal on empty vs. nil slices
+					continue
+				}
+
+				if !reflect.DeepEqual(gotEvents, step.wantEvents) {
+					t.Errorf("step %d: input %q: got events %#v, want %#v", i, step.input, gotEvents, step.wantEvents)
+				}
+			}
+		})
+	}
+}
+
+func TestMinistralParser_Errors(t *testing.T) {
+	t.Run("unknown tool returns error", func(t *testing.T) {
+		p := &MinistralParser{}
+		p.Init([]api.Tool{{Function: api.ToolFunction{Name: "known_tool"}}}, nil, nil)
+
+		_, _, _, err := p.Add(`[TOOL_CALLS]unknown_tool[ARGS]{"a": 1}`, true)
+		if err == nil {
+			t.Fatal("expected error for unknown tool")
+		}
+	})
+
+	t.Run("invalid JSON returns error", func(t *testing.T) {
+		p := &MinistralParser{}
+		p.Init([]api.Tool{{Function: api.ToolFunction{Name: "test"}}}, nil, nil)
+
+		_, _, _, err := p.Add(`[TOOL_CALLS]test[ARGS]{invalid json}`, true)
+		if err == nil {
+			t.Fatal("expected error for invalid JSON")
+		}
+	})
+}
+
+func TestFindJSONEnd(t *testing.T) {
+	tests := []struct {
+		name     string
+		input    string
+		expected int
+	}{
+		{
+			name:     "simple object",
+			input:    `{"a": 1}`,
+			expected: 7,
+		},
+		{
+			name:     "nested object",
+			input:    `{"a": {"b": 2}}`,
+			expected: 14,
+		},
+		{
+			name:     "array inside object",
+			input:    `{"items": [1, 2, 3]}`,
+			expected: 19,
+		},
+		{
+			name:     "braces in string",
+			input:    `{"template": "Hello {name}!"}`,
+			expected: 28,
+		},
+		{
+			name:     "escaped quotes",
+			input:    `{"msg": "say \"hi\""}`,
+			expected: 20,
+		},
+		{
+			name:     "incomplete object",
+			input:    `{"a": {"b": 1}`,
+			expected: -1,
+		},
+		{
+			name:     "deeply nested",
+			input:    `{"a": {"b": {"c": {"d": 1}}}}`,
+			expected: 28,
+		},
+		{
+			name:     "object with trailing content",
+			input:    `{"a": 1} extra`,
+			expected: 7,
+		},
+		{
+			name:     "array",
+			input:    `[{"a": 1}, {"b": 2}]`,
+			expected: 19,
+		},
+		{
+			name:     "escaped backslash before quote",
+			input:    `{"path": "C:\\"}`,
+			expected: 15,
+		},
+		{
+			name:     "empty string",
+			input:    "",
+			expected: -1,
+		},
+		{
+			name:     "no opening brace",
+			input:    "hello world",
+			expected: -1,
+		},
+		{
+			name:     "only opening brace",
+			input:    "{",
+			expected: -1,
+		},
+		{
+			name:     "unclosed string",
+			input:    `{"key": "unclosed`,
+			expected: -1,
+		},
+		{
+			name:     "double escaped backslash then quote",
+			input:    `{"path": "C:\\\\"}`,
+			expected: 17,
+		},
+		{
+			name:     "unicode in key and value",
+			input:    `{"키": "값"}`,
+			expected: 13,
+		},
+		{
+			name:     "nested arrays",
+			input:    `{"matrix": [[1, 2], [3, 4]]}`,
+			expected: 27,
+		},
+		{
+			name:     "mixed nesting",
+			input:    `{"a": [{"b": {"c": [1, 2, 3]}}]}`,
+			expected: 31,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			result := findJSONEnd(tt.input)
+			if result != tt.expected {
+				t.Errorf("findJSONEnd(%q) = %d, want %d", tt.input, result, tt.expected)
+			}
+		})
+	}
+}
+
+func TestMinistralParser_HasToolSupport(t *testing.T) {
+	p := &MinistralParser{}
+	if !p.HasToolSupport() {
+		t.Error("expected HasToolSupport to return true")
+	}
+}
+
+func TestMinistralParser_HasThinkingSupport(t *testing.T) {
+	p := &MinistralParser{hasThinkingSupport: false}
+	if p.HasThinkingSupport() {
+		t.Error("expected HasThinkingSupport to return false")
+	}
+
+	p = &MinistralParser{hasThinkingSupport: true}
+	if !p.HasThinkingSupport() {
+		t.Error("expected HasThinkingSupport to return true")
+	}
+}

model/parsers/parsers.go

@@ -3,6 +3,7 @@ package parsers
 import (
 	"strings"
 	"unicode"
+	"unicode/utf8"
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/harmony"
@@ -114,3 +115,33 @@ func splitAtTag(sb *strings.Builder, tag string, trimAfter bool) (string, string
 	sb.WriteString(after)
 	return before, after // return events
 }
+
+// overlap returns the longest overlap between the suffix of s and the prefix of delim
+func overlap(s, delim string) int {
+	max := min(len(delim), len(s))
+	for i := max; i > 0; i-- {
+		if strings.HasSuffix(s, delim[:i]) {
+			return i
+		}
+	}
+	return 0
+}
+
+// trailingWhitespaceLen returns the length in bytes of trailing whitespace in s
+func trailingWhitespaceLen(s string) int {
+	remaining := s
+	total := 0
+	for len(remaining) > 0 {
+		r, size := utf8.DecodeLastRuneInString(remaining)
+		// if it's an invalid utf8 rune, assume it isn't whitespace
+		if r == utf8.RuneError && size == 1 {
+			break
+		}
+		if !unicode.IsSpace(r) {
+			break
+		}
+		total += size
+		remaining = remaining[:len(remaining)-size]
+	}
+	return total
+}

model/parsers/qwen3coder.go

@@ -11,7 +11,6 @@ import (
 	"strconv"
 	"strings"
 	"unicode"
-	"unicode/utf8"
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/logutil"
@@ -194,36 +193,6 @@ func eat(p *Qwen3CoderParser) ([]qwenEvent, bool) {
 	}
 }
 
-// TODO(drifkin): move this to a shared location
-// longest overlap between suffix of s and prefix of delim
-func overlap(s, delim string) int {
-	max := min(len(delim), len(s))
-	for i := max; i > 0; i-- {
-		if strings.HasSuffix(s, delim[:i]) {
-			return i
-		}
-	}
-	return 0
-}
-
-func trailingWhitespaceLen(s string) int {
-	remaining := s
-	total := 0
-	for len(remaining) > 0 {
-		r, size := utf8.DecodeLastRuneInString(remaining)
-		// if it's an invalid utf8 rune, assume it isn't whitespace
-		if r == utf8.RuneError && size == 1 {
-			break
-		}
-		if !unicode.IsSpace(r) {
-			break
-		}
-		total += size
-		remaining = remaining[:len(remaining)-size]
-	}
-	return total
-}
-
 type XMLFunctionCall struct {
 	XMLName    xml.Name       `xml:"function"`
 	Name       string         `xml:"name,attr"`

server/routes.go

@@ -85,6 +85,7 @@ type Server struct {
 	addr    net.Addr
 	sched   *Scheduler
 	lowVRAM bool
+	usage   *UsageTracker
 }
 
 func init() {
@@ -273,6 +274,10 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		c.Header("Content-Type", contentType)
 
 		fn := func(resp api.GenerateResponse) error {
+			if resp.Done {
+				s.usage.Record(origModel, resp.PromptEvalCount, resp.EvalCount)
+			}
+
 			resp.Model = origModel
 			resp.RemoteModel = m.Config.RemoteModel
 			resp.RemoteHost = m.Config.RemoteHost
@@ -579,6 +584,8 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 					}
 					res.Context = tokens
 				}
+
+				s.usage.Record(req.Model, cr.PromptEvalCount, cr.EvalCount)
 			}
 
 			if builtinParser != nil {
@@ -1590,6 +1597,8 @@ func (s *Server) GenerateRoutes(rc *ollama.Registry) (http.Handler, error) {
 	r.HEAD("/api/blobs/:digest", s.HeadBlobHandler)
 	r.POST("/api/copy", s.CopyHandler)
 
+	r.GET("/api/usage", s.UsageHandler)
+
 	// Inference
 	r.GET("/api/ps", s.PsHandler)
 	r.POST("/api/generate", s.GenerateHandler)
@@ -1658,7 +1667,7 @@ func Serve(ln net.Listener) error {
 		}
 	}
 
-	s := &Server{addr: ln.Addr()}
+	s := &Server{addr: ln.Addr(), usage: NewUsageTracker()}
 
 	var rc *ollama.Registry
 	if useClient2 {
@@ -1875,6 +1884,10 @@ func (s *Server) SignoutHandler(c *gin.Context) {
 	c.JSON(http.StatusOK, nil)
 }
 
+func (s *Server) UsageHandler(c *gin.Context) {
+	c.JSON(http.StatusOK, s.usage.Stats())
+}
+
 func (s *Server) PsHandler(c *gin.Context) {
 	models := []api.ProcessModelResponse{}
 
@@ -2033,6 +2046,10 @@ func (s *Server) ChatHandler(c *gin.Context) {
 		c.Header("Content-Type", contentType)
 
 		fn := func(resp api.ChatResponse) error {
+			if resp.Done {
+				s.usage.Record(origModel, resp.PromptEvalCount, resp.EvalCount)
+			}
+
 			resp.Model = origModel
 			resp.RemoteModel = m.Config.RemoteModel
 			resp.RemoteHost = m.Config.RemoteHost
@@ -2253,6 +2270,8 @@ func (s *Server) ChatHandler(c *gin.Context) {
 					res.DoneReason = r.DoneReason.String()
 					res.TotalDuration = time.Since(checkpointStart)
 					res.LoadDuration = checkpointLoaded.Sub(checkpointStart)
+
+					s.usage.Record(req.Model, r.PromptEvalCount, r.EvalCount)
 				}
 
 				if builtinParser != nil {

server/routes_generate_test.go

@@ -88,19 +88,39 @@ func TestGenerateChatRemote(t *testing.T) {
 		if r.Method != http.MethodPost {
 			t.Errorf("Expected POST request, got %s", r.Method)
 		}
-		if r.URL.Path != "/api/chat" {
-			t.Errorf("Expected path '/api/chat', got %s", r.URL.Path)
-		}
 
 		w.WriteHeader(http.StatusOK)
 		w.Header().Set("Content-Type", "application/json")
-		resp := api.ChatResponse{
-			Model:      "test",
-			Done:       true,
-			DoneReason: "load",
-		}
-		if err := json.NewEncoder(w).Encode(&resp); err != nil {
-			t.Fatal(err)
+
+		switch r.URL.Path {
+		case "/api/chat":
+			resp := api.ChatResponse{
+				Model:      "test",
+				Done:       true,
+				DoneReason: "load",
+				Metrics: api.Metrics{
+					PromptEvalCount: 10,
+					EvalCount:       20,
+				},
+			}
+			if err := json.NewEncoder(w).Encode(&resp); err != nil {
+				t.Fatal(err)
+			}
+		case "/api/generate":
+			resp := api.GenerateResponse{
+				Model:      "test",
+				Done:       true,
+				DoneReason: "stop",
+				Metrics: api.Metrics{
+					PromptEvalCount: 5,
+					EvalCount:       15,
+				},
+			}
+			if err := json.NewEncoder(w).Encode(&resp); err != nil {
+				t.Fatal(err)
+			}
+		default:
+			t.Errorf("unexpected path %s", r.URL.Path)
 		}
 	}))
 	defer rs.Close()
@@ -111,7 +131,7 @@ func TestGenerateChatRemote(t *testing.T) {
 	}
 
 	t.Setenv("OLLAMA_REMOTES", p.Hostname())
-	s := Server{}
+	s := Server{usage: NewUsageTracker()}
 	w := createRequest(t, s.CreateHandler, api.CreateRequest{
 		Model:      "test-cloud",
 		RemoteHost: rs.URL,
@@ -159,6 +179,61 @@ func TestGenerateChatRemote(t *testing.T) {
 			t.Errorf("expected done reason load, got %s", actual.DoneReason)
 		}
 	})
+
+	t.Run("remote chat usage tracking", func(t *testing.T) {
+		stats := s.usage.Stats()
+		found := false
+		for _, m := range stats.Usage {
+			if m.Model == "test-cloud" {
+				found = true
+				if m.Requests != 1 {
+					t.Errorf("expected 1 request, got %d", m.Requests)
+				}
+				if m.PromptTokens != 10 {
+					t.Errorf("expected 10 prompt tokens, got %d", m.PromptTokens)
+				}
+				if m.CompletionTokens != 20 {
+					t.Errorf("expected 20 completion tokens, got %d", m.CompletionTokens)
+				}
+			}
+		}
+		if !found {
+			t.Error("expected usage entry for test-cloud")
+		}
+	})
+
+	t.Run("remote generate usage tracking", func(t *testing.T) {
+		// Reset the tracker for a clean test
+		s.usage = NewUsageTracker()
+
+		w := createRequest(t, s.GenerateHandler, api.GenerateRequest{
+			Model:  "test-cloud",
+			Prompt: "hello",
+		})
+		if w.Code != http.StatusOK {
+			t.Fatalf("expected status 200, got %d", w.Code)
+		}
+
+		stats := s.usage.Stats()
+		found := false
+		for _, m := range stats.Usage {
+			if m.Model == "test-cloud" {
+				found = true
+				if m.Requests != 1 {
+					t.Errorf("expected 1 request, got %d", m.Requests)
+				}
+				if m.PromptTokens != 5 {
+					t.Errorf("expected 5 prompt tokens, got %d", m.PromptTokens)
+				}
+				if m.CompletionTokens != 15 {
+					t.Errorf("expected 15 completion tokens, got %d", m.CompletionTokens)
+				}
+			}
+		}
+		if !found {
+			t.Error("expected usage entry for test-cloud")
+		}
+	})
 }
 
 func TestGenerateChat(t *testing.T) {

server/usage.go

@@ -0,0 +1,62 @@
+package server
+
+import (
+	"sync"
+	"time"
+
+	"github.com/ollama/ollama/api"
+)
+
+type ModelUsage struct {
+	Requests         int64
+	PromptTokens     int64
+	CompletionTokens int64
+}
+
+type UsageTracker struct {
+	mu     sync.Mutex
+	start  time.Time
+	models map[string]*ModelUsage
+}
+
+func NewUsageTracker() *UsageTracker {
+	return &UsageTracker{
+		start:  time.Now().UTC(),
+		models: make(map[string]*ModelUsage),
+	}
+}
+
+func (u *UsageTracker) Record(model string, promptTokens, completionTokens int) {
+	u.mu.Lock()
+	defer u.mu.Unlock()
+
+	m, ok := u.models[model]
+	if !ok {
+		m = &ModelUsage{}
+		u.models[model] = m
+	}
+
+	m.Requests++
+	m.PromptTokens += int64(promptTokens)
+	m.CompletionTokens += int64(completionTokens)
+}
+
+func (u *UsageTracker) Stats() api.UsageResponse {
+	u.mu.Lock()
+	defer u.mu.Unlock()
+
+	byModel := make([]api.ModelUsageData, 0, len(u.models))
+	for model, usage := range u.models {
+		byModel = append(byModel, api.ModelUsageData{
+			Model:            model,
+			Requests:         usage.Requests,
+			PromptTokens:     usage.PromptTokens,
+			CompletionTokens: usage.CompletionTokens,
+		})
+	}
+
+	return api.UsageResponse{
+		Start: u.start,
+		Usage: byModel,
+	}
+}

server/usage_test.go

@@ -0,0 +1,136 @@
+package server
+
+import (
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"sync"
+	"testing"
+
+	"github.com/gin-gonic/gin"
+	"github.com/ollama/ollama/api"
+)
+
+func TestUsageTrackerRecord(t *testing.T) {
+	tracker := NewUsageTracker()
+
+	tracker.Record("model-a", 10, 20)
+	tracker.Record("model-a", 5, 15)
+	tracker.Record("model-b", 100, 200)
+
+	stats := tracker.Stats()
+
+	if len(stats.Usage) != 2 {
+		t.Fatalf("expected 2 models, got %d", len(stats.Usage))
+	}
+
+	lookup := make(map[string]api.ModelUsageData)
+	for _, m := range stats.Usage {
+		lookup[m.Model] = m
+	}
+
+	a := lookup["model-a"]
+	if a.Requests != 2 {
+		t.Errorf("model-a requests: expected 2, got %d", a.Requests)
+	}
+	if a.PromptTokens != 15 {
+		t.Errorf("model-a prompt tokens: expected 15, got %d", a.PromptTokens)
+	}
+	if a.CompletionTokens != 35 {
+		t.Errorf("model-a completion tokens: expected 35, got %d", a.CompletionTokens)
+	}
+
+	b := lookup["model-b"]
+	if b.Requests != 1 {
+		t.Errorf("model-b requests: expected 1, got %d", b.Requests)
+	}
+	if b.PromptTokens != 100 {
+		t.Errorf("model-b prompt tokens: expected 100, got %d", b.PromptTokens)
+	}
+	if b.CompletionTokens != 200 {
+		t.Errorf("model-b completion tokens: expected 200, got %d", b.CompletionTokens)
+	}
+}
+
+func TestUsageTrackerConcurrent(t *testing.T) {
+	tracker := NewUsageTracker()
+
+	var wg sync.WaitGroup
+	for i := 0; i < 100; i++ {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			tracker.Record("model-a", 1, 2)
+		}()
+	}
+	wg.Wait()
+
+	stats := tracker.Stats()
+	if len(stats.Usage) != 1 {
+		t.Fatalf("expected 1 model, got %d", len(stats.Usage))
+	}
+
+	m := stats.Usage[0]
+	if m.Requests != 100 {
+		t.Errorf("requests: expected 100, got %d", m.Requests)
+	}
+	if m.PromptTokens != 100 {
+		t.Errorf("prompt tokens: expected 100, got %d", m.PromptTokens)
+	}
+	if m.CompletionTokens != 200 {
+		t.Errorf("completion tokens: expected 200, got %d", m.CompletionTokens)
+	}
+}
+
+func TestUsageTrackerStart(t *testing.T) {
+	tracker := NewUsageTracker()
+
+	stats := tracker.Stats()
+	if stats.Start.IsZero() {
+		t.Error("expected non-zero start time")
+	}
+}
+
+func TestUsageHandler(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	s := &Server{
+		usage: NewUsageTracker(),
+	}
+
+	s.usage.Record("llama3", 50, 100)
+	s.usage.Record("llama3", 25, 50)
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Request = httptest.NewRequest(http.MethodGet, "/api/usage", nil)
+
+	s.UsageHandler(c)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d", w.Code)
+	}
+
+	var resp api.UsageResponse
+	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("failed to unmarshal response: %v", err)
+	}
+
+	if len(resp.Usage) != 1 {
+		t.Fatalf("expected 1 model, got %d", len(resp.Usage))
+	}
+
+	m := resp.Usage[0]
+	if m.Model != "llama3" {
+		t.Errorf("expected model llama3, got %s", m.Model)
+	}
+	if m.Requests != 2 {
+		t.Errorf("expected 2 requests, got %d", m.Requests)
+	}
+	if m.PromptTokens != 75 {
+		t.Errorf("expected 75 prompt tokens, got %d", m.PromptTokens)
+	}
+	if m.CompletionTokens != 150 {
+		t.Errorf("expected 150 completion tokens, got %d", m.CompletionTokens)
+	}
+}