Rocuronium — Claude Code MCP server for macOS app control
Gives an agent its own macOS login session, desktop and cursor: it reads the UI, clicks, types and scrolls while you keep using your machine.
Open on GitHubWhat Rocuronium does
Rocuronium spins up a separate macOS login session for the agent, with its own desktop and cursor. Your own session stays untouched. When an app exposes little accessibility data, local vision models step in to locate controls; on-device OCR and window screenshots cover the rest.
Every interaction comes back with a verdict: confirmed, noEffect or unverifiable. That matters because macOS often reports a successful API call while the screen shows nothing new. The check looks at pixel, focus, window or selection changes.
Requirements are Apple silicon, macOS 26.5+ and Accessibility permission; Screen Recording is needed for screenshots and OCR. Creating a session asks for human approval, and the emergency stop sits on ⌃⌥⇧⎋ — resuming only works from the app's popover. Install with brew install kageroumado/tap/rocuronium. Agent setup and MCP configuration live in .claude/skills/rocuronium/SKILL.md, and rocuronium guide prints the same.
What it helps with
- drive a macOS app without hijacking the human cursor
- tell a real click apart from an API call that changed nothing
- locate a control in an app with a thin accessibility tree
- run a sequence of actions with postconditions and fallback
- halt the agent when it grabs the wrong window
Example prompts for the agent
- Open Safari in your own session and find the privacy settings
- Click the Save button and show what changed on screen
- Scroll the list to the bottom and confirm new items appeared
How to install Rocuronium
Install
The MCP registry has no launch command for it: see the repository README
Step by step, with commands for each platform: How to add an MCP server to Claude Code, Codex, Cursor, VS Code, Gemini CLI and more
Questions about Rocuronium
How do I install Rocuronium in Claude Code?
Use Homebrew: brew install kageroumado/tap/rocuronium. Then wire up MCP following .claude/skills/rocuronium/SKILL.md, or run rocuronium guide.
What does Rocuronium do?
It hands the agent a separate macOS session with its own desktop and cursor. The agent reads the UI, clicks, types and scrolls, and each action returns a confirmation or a signal that nothing changed.
What does Rocuronium need to run?
Apple silicon, macOS 26.5+ and Accessibility permission. Screenshots and OCR also require Screen Recording.
