DEV Community

Mininglamp
Mininglamp

Posted on

Deploying a Local GUI Agent on Apple Silicon

We've been running Mano-P as a local GUI agent on Apple Silicon for a few weeks now. Here's the practical walkthrough for getting it up on your own Mac.

What It Does

Mano-P is a vision-driven GUI agent. It looks at your screen, decides what to click or type, and executes multi-step tasks autonomously. No OCR pipeline, no accessibility tree parsing. Pure screenshot-to-action. The model is Mano-CUA-4B-Thinking-1.1, quantized to MLX 8-bit, with a peak memory footprint of 4.3GB.

Setup

The install is a single Homebrew tap:

brew tap Mininglamp-AI/tap && brew install mano-cua
Enter fullscreen mode Exit fullscreen mode

Then three commands:

mano-cua check        # verify hardware compatibility
mano-cua install-sdk  # install runtime dependencies
mano-cua install-model # download the 4B model
Enter fullscreen mode Exit fullscreen mode

Hardware requirement: M4 chip or newer with 32GB+ RAM. Alternatively, any Mac with a USB 4.0+ compute stick works.

You'll need to grant Screen Recording and Accessibility permissions in System Settings. That's the entire setup.

Running It

mano-cua run "open Safari, search for today's weather" --local
Enter fullscreen mode Exit fullscreen mode

The --local flag keeps everything on-device. No API keys, no cloud calls, no screenshots leaving your machine.

Performance

On an M5 Pro, decode speed sits around 80 tok/s. In our 100-task real Mac GUI benchmark on a MacBook Pro M5 with 16GB, the Thinking-4B model hit 56% task completion. For comparison, Qwen3-VL-Plus running through cloud scored 39% on the same test suite. On OSWorld, specialized models reached 58.2%, currently #1 on that leaderboard.

Cider W8A8 quantization gives roughly 12.7% faster prefill compared to W8A16.

Trade-offs

The 4B model is small. It handles routine GUI tasks well but struggles with complex multi-app workflows. The 32GB RAM floor is real. And granting Screen Recording access means the agent can see everything on your display, so treat it like you'd treat any screen-sharing tool.

Source and License

Full client source is open under Apache 2.0 for audit. GitHub: https://github.com/Mininglamp-AI/Mano-P

If you try it out, we'd appreciate a star. Bug reports even more so.

Top comments (0)