Rocuronium

What Rocuronium does

Rocuronium MCP server gives an AI agent a way to observe and control macOS applications while a person remains at the computer. It can inspect application text, windows, controls, screenshots, and activity, then perform actions such as clicking, typing, keyboard shortcuts, scrolling, launching, and activating apps.

The default interaction path uses accessibility operations and posted per-process events rather than moving the physical cursor or changing the user's focus. Actions return an evidence verdict: confirmed, noEffect, or unverifiable. Verification can include reading text after a write, comparing pixels, checking windows, or detecting process exit.

How it works

The application runs as a menu bar daemon and maintains the macOS permissions needed for accessibility and capture. The command-line executable communicates with that daemon over a Unix-domain socket and can also start the MCP endpoint. MCP tools use the same verbs and arguments as the CLI.

When the accessibility tree is insufficient, Rocuronium MCP server can use local OCR and a local UI-element detector, followed by a local vision-language model for remaining cases. Returned UI rows identify the grounding tier used. Screenshots and text observations can be compared so an agent receives changed regions or elements rather than repeating the entire observation.

Multi-step plans execute inside the daemon. Guards check postconditions between steps, and failure behavior can abort, continue, pause for a person, or use a fallback. A plan uses one desktop for all of its steps.

Setup and configuration

Rocuronium MCP server requires macOS 26.5 or later on Apple silicon. Install it with Homebrew or download the signed, notarized DMG from the project’s release link. In System Settings, grant Accessibility access; Screen Recording is also needed for screenshots and OCR-based scrolling.

Start the MCP process with rocuronium mcp. A Claude Code configuration can invoke /Applications/Rocuronium.app/Contents/Resources/rocuronium with mcp as its argument. The MCP interface permits only status and stop for the session capability; creating a login session remains a human-controlled operation.

Tools and capabilities

Available command groups include:

• Observation: status, diag, apps, windows, find, read, map, wait, screenshot, and activity.
• Input and application control: type, click, key, shortcut, menu, scroll, statusitem, launch, and activate.
• Window and desktop operations: resize, window, space, display, and park.
• Orchestration and support: plan, busy, demo, request-capture, guide, and version.

Virtual display leases can isolate work from the physical screen, and windows can be parked there temporarily. The optional agent desktop targets a separate macOS user session when that session is available.

Limitations and notes

Hardware cursor operations such as move and drag are not silent: they require explicit opt-in and are refused when a person is present, the screen is locked, or another window covers the target. Cursor-taking actions display a presence overlay, yield when the mouse is touched, and can be halted with Ctrl+Option+Shift+Escape. Resuming requires the menu bar control.

The daemon needs the stated macOS version and Apple silicon. Some applications expose little or no usable accessibility data, so OCR or vision-based grounding may be required. MCP session creation is intentionally unavailable, and the agent cannot clear the emergency halt through a socket command.