apple-local-llm

Call Apple's on-device Foundation Models from JavaScript — no servers, no setup.

Works with Node.js, Electron, and VS Code extensions.

Requirements

• macOS 26+ (Tahoe)
• Apple Silicon (M Series)
• Apple Intelligence enabled in System Settings

Installation

npm install apple-local-llm

Quick Start

Simple API

import { createClient } from "apple-local-llm";

const client = createClient();

// Check compatibility first
const compat = await client.compatibility.check();
if (!compat.compatible) {
  console.log("Not available:", compat.reasonCode);
  // Handle fallback to cloud API
}

// Generate a response
const result = await client.responses.create({
  input: "What is the capital of France?",
});

if (result.ok) {
  console.log(result.text); // "The capital of France is Paris."
}

Streaming

for await (const chunk of client.stream({ input: "Count from 1 to 5." })) {
  if ("delta" in chunk) {
    process.stdout.write(chunk.delta);
  }
}

API Reference

createClient(options?)

Creates a new client instance.

const client = createClient({
  model: "default",               // Optional: model identifier (currently only "default")
  onLog: (msg) => console.log(msg), // Optional: debug logging
  idleTimeoutMs: 5 * 60 * 1000,     // Optional: helper idle timeout (default: 5 min)
});

Defaults:

• Helper auto-shuts down after 5 minutes of inactivity
• Helper auto-restarts up to 3 times on crash (with exponential backoff)
• Request timeout: 60 seconds (configurable via timeoutMs)

You can also import and instantiate the class directly:

import { AppleLocalLLMClient } from "apple-local-llm";
const client = new AppleLocalLLMClient(options);

client.compatibility.check()

Check if the local model is available. Always call this before making requests.

const result = await client.compatibility.check();
// { compatible: true }
// or { compatible: false, reasonCode: "AI_DISABLED" }

Reason codes: | Code | Description | |------|-------------| | NOT_DARWIN | Not running on macOS | | UNSUPPORTED_HARDWARE | Not Apple Silicon | | AI_DISABLED | Apple Intelligence not enabled | | MODEL_NOT_READY | Model still downloading | | SPAWN_FAILED | Helper binary failed to start | | HELPER_NOT_FOUND | Helper binary not found | | HELPER_UNHEALTHY | Helper process not responding correctly | | PROTOCOL_MISMATCH | Helper version incompatible with client |

client.capabilities.get()

Get detailed model capabilities (calls the helper).

const caps = await client.capabilities.get();
// { available: true, model: "apple-on-device" }
// or { available: false, reasonCode: "AI_DISABLED" }

client.responses.create(params)

Generate a response.

const result = await client.responses.create({
  input: "Your prompt here",
  model: "default",         // Optional: model identifier
  max_output_tokens: 500,   // Optional: limit response tokens
  stream: false,            // Optional
  signal: abortController.signal, // Optional: AbortSignal
  timeoutMs: 60000,         // Optional: request timeout (ms)
  response_format: {        // Optional: structured JSON output
    type: "json_schema",
    json_schema: {
      name: "Result",
      schema: { type: "object", properties: { ... } }
    }
  }
});

Structured Output Example:

const result = await client.responses.create({
  input: "List 3 colors",
  response_format: {
    type: "json_schema",
    json_schema: {
      name: "Colors",
      schema: {
        type: "object",
        properties: {
          colors: { type: "array", items: { type: "string" } }
        }
      }
    }
  }
});
const data = JSON.parse(result.text); // { colors: ["red", "blue", "green"] }

response_format is not supported with streaming.

Returns ResponseResult on success, or an error object:

// Success:
{ ok: true, text: "...", request_id: "..." }
// Error:
{ ok: false, error: { code: "...", detail: "..." } }

Note: The return type is a discriminated union, not the exported ResponseResult interface.

Error codes: | Code | Description | |------|-------------| | UNAVAILABLE | Model not available (see reason codes above) | | TIMEOUT | Request timed out (default: 60s) | | CANCELLED | Request was cancelled via AbortSignal | | RATE_LIMITED | System rate limit exceeded | | GUARDRAIL | Content violated Apple's safety guidelines | | INTERNAL | Unexpected error |

client.stream(params)

Async generator for streaming responses.

for await (const chunk of client.stream({ input: "..." })) {
  if ("delta" in chunk) {
    // Partial content
    console.log(chunk.delta);
  } else if ("done" in chunk) {
    // Final complete text
    console.log(chunk.text);
  }
}

client.responses.cancel(requestId)

Cancel an in-progress request.

const result = await client.responses.cancel("req_123");
// { ok: true } or { ok: false, error: { code: "NOT_RUNNING", detail: "..." } }

client.shutdown()

Gracefully shut down the helper process.

await client.shutdown();

TypeScript Types

All types are exported:

import type {
  ClientOptions,
  ReasonCode,
  CompatibilityResult,
  CapabilitiesResult,
  ResponsesCreateParams,
  ResponseResult,
  JSONSchema,
  ResponseFormat,
} from "apple-local-llm";

CLI Usage

The fm-proxy binary can also be used directly from the command line:

# Simple prompt
fm-proxy "What is the capital of France?"

# Streaming output
fm-proxy --stream "Tell me a story"
fm-proxy -s "Tell me a story"

# Limit output tokens
fm-proxy --max-tokens=50 "Count to 100"

# Start HTTP server
fm-proxy --serve
fm-proxy --serve --port=3000

# Other options
fm-proxy --help      # Show usage (or -h)
fm-proxy --version   # Show version (or -v)
fm-proxy --stdio     # stdio mode (used internally by npm package)

HTTP Server Mode

Run fm-proxy --serve to start a local HTTP server:

fm-proxy --serve --port=8080

Endpoints:

| Endpoint | Method | Description | |----------|--------|-------------| | /health | GET | Health check and availability status | | /generate | POST | Text generation (supports streaming) |

Options:

| Option | Description | |--------|-------------| | --port=<PORT> | Set server port (default: 8080) | | --auth-token=<TOKEN> | Require Bearer token for /generate |

You can also set AUTH_TOKEN environment variable instead of --auth-token.

CORS: All endpoints support CORS with Access-Control-Allow-Origin: *.

Examples:

# Health check
curl http://127.0.0.1:8080/health
# Response: {"status":"ok","model":"apple-on-device","available":true}

# Simple generation
curl -X POST http://127.0.0.1:8080/generate \
  -H "Content-Type: application/json" \
  -d '{"input": "What is 2+2?"}'
# Response: {"text":"2+2 equals 4."}

# With max_output_tokens
curl -X POST http://127.0.0.1:8080/generate \
  -H "Content-Type: application/json" \
  -d '{"input": "Count to 100", "max_output_tokens": 50}'

# With structured output (response_format)
curl -X POST http://127.0.0.1:8080/generate \
  -H "Content-Type: application/json" \
  -d '{"input": "List 3 colors", "response_format": {"type": "json_schema", "json_schema": {"name": "Colors", "schema": {"type": "object", "properties": {"colors": {"type": "array", "items": {"type": "string"}}}}}}}'

# With authentication
curl -X POST http://127.0.0.1:8080/generate \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer <token>" \
  -d '{"input": "Hello"}'

Streaming (SSE)

Add "stream": true to get Server-Sent Events with OpenAI-compatible chunks:

curl -N -X POST http://127.0.0.1:8080/generate \
  -H "Content-Type: application/json" \
  -d '{"input": "Write a haiku", "stream": true}'

Response:

data: {"id":"...","object":"chat.completion.chunk","choices":[{"delta":{"role":"assistant"}}]}
data: {"id":"...","object":"chat.completion.chunk","choices":[{"delta":{"content":"..."}}]}
data: {"id":"...","object":"chat.completion.chunk","choices":[{"delta":{},"finish_reason":"stop"}]}
data: [DONE]

How It Works

This package bundles a small native helper (fm-proxy) that communicates with Apple's Foundation Models framework over stdio. The helper is spawned on first request and stays alive to keep the model warm.

• No localhost server — npm package uses stdio, not HTTP
• No user setup — just npm install
• Fails gracefully — check compatibility.check() and fall back to cloud

Runtime Support

JS API (createClient()): | Environment | Supported | |-------------|-----------| | Node.js | ✅ | | Electron (main process) | ✅ | | VS Code extensions | ✅ | | Electron (renderer) | ❌ No child_process | | Browser | ❌ |

HTTP Server (fm-proxy --serve): | Environment | Supported | |-------------|-----------| | Any HTTP client | ✅ | | Browser (fetch) | ✅ | | Electron (renderer) | ✅ |

License

MIT