jellythomas/agent-eyes ? reverse-engineered prompt

Build me a Python tool called Agent Eyes that acts as a computer use MCP server for desktop apps and browsers. I want it to use the computer’s native accessibility tree and real keyboard and mouse input so an AI agent can inspect and operate open apps without being tied to a specific browser or automation framework.

It should have a setup flow that checks the machine, installs or reuses the right launcher, and can configure common MCP clients like Claude Code, Claude Desktop, Codex, Cursor, VS Code, Cline, Roo Code, Windsurf, Zed, and Continue. It should also include a doctor command that reports whether everything is ready, plus a serve mode for the persistent MCP server.

Make the default behavior foreground first and native, so it looks at open tabs and existing windows before opening anything new. Only use background protocol automation when shadow mode is explicitly enabled. If you need current docs for any platform details, look them up online and keep the app working on macOS, Windows, and Linux.