fmx

A friendly CLI for Apple's on-device Foundation Model — chat, respond, and schema from your terminal, on macOS 26 today.

macOS 27 will ship a native fm command. fmx brings that same workflow to macOS 26 right now, built on Apple's apple-fm-sdk. It uses a distinct name (fmx) so it never collides with the future system fm.

$ fmx chat
fmx chat — Apple on-device Foundation Model
Type a message, or /help for commands. /exit to quit.

you ❯ explain async/await in one sentence
fmx ❯ async/await lets you write non-blocking code that reads top-to-bottom, pausing at
       await points so other work can run while you wait on I/O.

you ❯ /save mychat.json
Saved 4 entries to mychat.json.

Requirements

Everything runs on-device through Apple Intelligence — there's no cloud, no API key, and it only works on a compatible Mac:

• macOS 26 or later
• Apple silicon, M1 or later, with Apple Intelligence turned on (compatible devices)
• Python 3.10+

If the model isn't available, fmx tells you exactly why (Apple Intelligence off, device ineligible, or model still downloading).

Install

uv tool install fmx
# or
pipx install fmx

Both install the fmx command into your user bin directory (~/.local/bin). If your shell then reports zsh: command not found: fmx, that directory isn't on your PATH yet — fix it once and open a new terminal:

uv tool update-shell     # if you installed with uv
# or
pipx ensurepath          # if you installed with pipx

Verify with fmx --version.

Usage

fmx chat — interactive

fmx chat
fmx chat -i "You are a terse shell expert."   # set system instructions

Slash commands inside a chat:

fmx respond — one-shot, scriptable

fmx respond "Summarize this in 5 words: <text>"
echo "say hi" | fmx respond -                       # prompt from stdin
fmx respond "Write a haiku about Swift" --stream     # stream tokens
fmx respond "Describe this" --image photo.jpg        # multimodal (repeatable)
fmx respond "Be terse." -t 0.2 --max-tokens 100      # temperature / token cap

fmx schema — structured output

Build a JSON Schema, then constrain a response to it. Output is parseable JSON:

fmx schema object name:str age:int "bio:str:a short biography" tags:str[]

# pipe a schema straight into respond:
fmx schema object name:str age:int | fmx respond "Invent a person" --schema -

Field spec is name:type[:description]. Types: str, int, float, bool (append [] for an array, e.g. tags:str[]).

What about Private Cloud Compute / /model?

Not yet. apple-fm-sdk currently exposes only the on-device model — there's no Python API for Private Cloud Compute. When Apple ships PCC access (via the SDK, or the native fm in macOS 27), model switching will land behind /model; the code already isolates session creation in fmx/core.py for exactly this.

Roadmap

• /model → Private Cloud Compute, once the SDK exposes it
• Auto-defer to the native fm binary when running on macOS 27
• Homebrew formula
• Tool/function calling in chat

Development

git clone https://github.com/manjunathshiva/fmx
cd fmx
uv sync
uv run fmx chat
uv run pytest          # model tests auto-skip if Apple Intelligence is unavailable
uv run ruff check .

Contributing

Contributions are welcome! See CONTRIBUTING.md for the dev setup, the SDK-free-vs-on-device test split, and the review rules — every PR needs passing lint-and-test + CodeQL checks and one approving review. A pull-request template is filled in for you.

Found a security issue? Please don't open a public issue — follow the security policy instead.

License

MIT. Built on Apple's apple-fm-sdk (Apache-2.0). Not affiliated with Apple.