ChatGPT is a powerful tool but since it is a Web-based tool, it lacks the capability to access local files or run commands.

Maccha is an HTTP proxy intended to be run on macOS locally, providing:

• read access to all local files
• write access to files under ~/maccha
• any execution of local commands, with the above access privileges

Internally, maccha uses sandbox_exec (1) for sandboxes.

The web interface is a dead copy of llama.cpp and its license is subject to that of llama.cpp.

See screenshot: screenshot.

1. Git clone this repository, e.g., to ~/projects.
2. To create the working directory for maccha, run mkdir ~/maccha.
3. Run maccha.js, with OPENAI_APY_KEY set (e.g., OPENAI_API_KEY=... node ~/projects/maccha/maccha.js).
4. Open http://127.0.0.1:11434.
5. Try asking things like "What time is it in Tokyo?", "take a screenshot of the active window!"

To faccilitate AI assitance over the local enviroment, following commands are provided. They can be either called via chat or directly in the console:

• maccha-activate-app: activates a macOS application and optionally simulates a keystroke
• maccha-filter: an AI agent that processes data and returns the result (can be used to filter data)
• maccha-ocr: an OCR program that converts images to text
• maccha-windowcapture: captures the active window and saves it as an image

The chat agent understands that these commands are available.

By default, maccha connects to OpenAI API, specifying gpt-4.1-mini as the model to be used.

This behavior can be changed by adding /model:<model-name> to the end of the chat message. If the model name starts with gpt, then OpenAI is used as the backend, with the model name specifying the actual model. Otherwise, model name is interpreted as host:port specifying where a local LLM server is being run.