ghostwright/ghost-os — reverse-engineered prompt

Build me a native macOS tool that lets AI agents actually use my computer, not just chat about it. I want it to connect to Claude Code, Cursor, VS Code, or any MCP client, then let the agent inspect open apps through macOS accessibility data and take actions like clicking, typing, scrolling, dragging, pressing keys, switching windows, and filling forms.

It should work across normal Mac apps like Finder, Slack, Messages, browsers, and Gmail. Please make it local first, so my screen and app data stay on my machine. If accessibility data is not enough, add a local vision fallback.

I also want reusable workflows. The agent should be able to watch me do a task once, turn that into an editable JSON recipe with parameters, then run it later when I ask, like sending an email or downloading a paper. Include a simple setup command that handles permissions, MCP config, recipes, and any local model setup. Look up current docs online if you need to.