Macaw is a private AI model that runs entirely on your Mac and controls it — 97 verified tools for mail, files, calendar, music, system settings, and multi-step chains, in plain English. No cloud, no accounts, no payments.

It's live today — not a waitlist. Weights on Hugging Face, source right here, MIT code, fully local on Apple Silicon.

• Model weights: badtheorylabs/Macaw (BF16) · badtheorylabs/Macaw-4bit-MLX (~1.5 GB on-device build)
• Measured (Apple M2, 4-bit): 10/10 tool-call accuracy · 1.21 s mean request · ~40 tok/s decode
• Architecture: fine-tune of Liquid AI LFM2.5-2.6B
• Landing page: badtheorylabs.com/macaw

pip install mlx mlx-lm
python -m mlx_lm.server --model badtheorylabs/Macaw-4bit-MLX --port 8138

swift build          # in app/

The app runs in the menu bar, starts the server automatically (point MACAW_PYTHON at a Python with mlx-lm if it's not on your PATH), and exposes a floating prompt bar (⌥ Space). The model can then call the 97 macOS tools in runtime/tools.json.

Seven tools (screen reading, document reading, app control) shell out to helper scripts. Put them where the tools expect to find them:

mkdir -p ~/.macaw
cp runtime/*.py runtime/tools.json ~/.macaw/

Point MACAW_PYTHON at a Python that has mlx-lm and pyobjc if it isn't your default:

export MACAW_PYTHON=/path/to/python
export MACAW_HOME=~/.macaw        # only if you put them elsewhere

Screen reading also needs Screen Recording permission, and app control needs Accessibility — both in System Settings › Privacy & Security. macOS will prompt the first time each is used.

pip install mlx mlx-lm pyobjc-framework-Vision
python runtime/agent.py          # multi-turn agent loop
python runtime/loop.py           # autonomous loop, terminal-style

┌─────────────────────────────────────────────┐
│  Macaw app (Swift, menu bar)                │
│  ─ prompt bar → /v1/chat/completions        │
│  ─ tool schema from tools.json (top-12 by   │
│    retrieval)                               │
│  ─ tool name + args checked, then executed  │
│    (model never writes the AppleScript)     │
└──────────────────────┬──────────────────────┘
                       │ 127.0.0.1:8138
┌──────────────────────▼──────────────────────┐
│  mlx_lm server — Macaw-4bit-MLX on Metal    │
│  on-device, no network                      │
└─────────────────────────────────────────────┘

Prompt hygiene. The model was fine-tuned to emit a tool call immediately after the assistant turn — no <think> preamble. The shipped chat_template.jinja is patched so generation doesn't prefill <think>, which would otherwise cost ~20 s of reasoning per request. The app sends a system message ("You are Macaw, an on-device AI assistant running on this Mac.") so the model presents as Macaw rather than its base identity.

Safety. Only a tool name plus literal arguments cross from model to machine, validated against tools.json. The model never produces the AppleScript that runs; the app does. Untrusted text the model reads (mail, filenames, web pages) can never talk it into executing arbitrary script.

app/         macOS menu-bar app (Swift/SwiftUI)
runtime/     Python agent loop, tool executor, screen reader, benchmark
landing/     Launch page (static HTML)

app/ needs macOS 14+ and the model server. runtime/ needs Python 3.10+ with mlx, mlx-lm, and on macOS pyobjc-framework-Vision for screen reading.

runtime/bench.py measures end-to-end tool-call accuracy and latency against the live server:

• Code in this repo: MIT
• Model weights: derivatives of LFM2.5-2.6B under the LFM Open License v1.0. Redistribution must retain that license and Liquid AI attribution; commercial use is restricted for entities ≥ $10M annual revenue.

Bad Theory Labs · model repos on Hugging Face