apple-fm

apple-fm

Apple Intelligence from your command line and your code. apple-fm gives you the on-device Foundation Models on macOS 26+ / Apple Silicon — free, private, and fully offline. No API key, no network, nothing leaves your Mac.

Apple ships these models as a Swift-only framework with no command-line front-end. apple-fm provides one — a fast, lightweight CLI and a TypeScript library, so you can use the on-device model from your terminal or your code.

Why apple-fm

• On-device & private. Runs on the model Apple Intelligence already installed. No key, no cloud, no telemetry — your prompts never leave the machine.
• One binary, three shapes. probe, one-shot generate (freeform, guided, or streamed), and an interactive chat — all over a single NDJSON protocol.
• Guaranteed structured output. Hand it a JSON Schema and the output is guaranteed to conform — native guided generation built on Apple's DynamicGenerationSchema, not best-effort prompting. Stream it, too.
• Long conversations stay coherent — and cheap. chat holds one on-device session across turns, reusing the model's cache instead of replaying the transcript, and automatically summarizes older turns as the small context window fills.
• Lightweight, nothing to audit. Zero runtime dependencies — apple-fm just talks to the on-device model and gets out of your way.
• Future-proof. The helper resolves the current on-device model at runtime, so OS and model updates are picked up without a rebuild.

Install

npm install -g apple-fm     # CLI
npm install apple-fm        # library

Requires macOS 26+ on Apple Silicon with Apple Intelligence enabled. Every release is Developer-ID signed and notarized, so it runs without security prompts.

Platform support

apple-fm installs on any OS — so it can be a dependency of a cross-platform project — but the on-device model only runs on macOS 26+ Apple Silicon. Everywhere else it degrades gracefully instead of crashing: probe() returns { available: false, reason: 'unsupportedPlatform' }, and generate / chat throw a clear [unsupportedPlatform] error. Gate on isPlatformSupported() or probe():

import { isPlatformSupported, probe, generate } from 'apple-fm';

if ((await probe()).available) {
  await generate({ prompt: '…' });
} else {
  // fall back (cloud model, cached result, skip the feature, …)
}

CLI

apple-fm probe                       # is the on-device model available?
apple-fm generate "Summarize: …"     # one-shot text
cat notes.md | apple-fm generate     # read the prompt from stdin
apple-fm generate "…" --stream       # stream tokens as they arrive
apple-fm generate "…" --schema shape.json   # structured/guided JSON output
apple-fm chat                        # interactive chat (streamed, auto-compacted)

Run apple-fm --help for the full flag list.

Check availability

Before generating, confirm Apple Intelligence is ready on this machine.

Structured output

Pass a JSON Schema with --schema and apple-fm returns JSON guaranteed to conform to it — native guided generation, not prompt-and-hope — ready to pipe into the rest of your tooling.

Interactive chat

chat is a multi-turn REPL that streams replies and compacts the transcript automatically near the context window. Built-in slash commands: /reset, /system, /clear, /compact, /help, /quit.

Library

import { probe, generate, ChatSession } from 'apple-fm';

if ((await probe()).available) {
  // One-shot
  const summary = await generate({ prompt: 'Summarize: …', system: 'Be terse.' });

  // Streaming
  await generate({ prompt: '…', stream: true }, {}, (chunk) => process.stdout.write(chunk));

  // Guaranteed structured output — pass a JSON Schema, get conforming JSON back
  const json = await generate({ prompt: 'A classic sci-fi novel', schema: novelSchema });

  // Multi-turn chat: one persistent on-device session, auto-compacted
  const chat = new ChatSession({ system: 'You are a helpful assistant.' });
  const reply = await chat.send('Hello');
}

The full API (probe, generate, ChatSession, protocol helpers, and types) is documented in docs/ai/code-summary.md.

How it works

apple-fm talks to the same on-device model that powers Apple Intelligence — nothing is sent to the cloud. For chat it keeps one model session alive across turns so replies stay fast, and automatically summarizes older messages as the context window fills, so long conversations stay coherent and cheap. And because it resolves the current on-device model at runtime, OS and model updates are picked up automatically — no reinstall.

Documentation

• Overview

Disclaimer

apple-fm is an independent, unofficial project and is not affiliated with, endorsed by, or sponsored by Apple Inc. It simply provides command-line and programmatic access to Apple's on-device Foundation Models on machines that already support them (macOS 26+ on Apple Silicon with Apple Intelligence enabled). Apple, Apple Intelligence, Apple Silicon, Foundation Models, and macOS are trademarks of Apple Inc.

License

MIT © Brian Westphal