Press release · Rocuronium
kageroumado Releases Rocuronium, Letting AI Agents Operate a Mac Without Taking the User's Cursor
Rocuronium is a macOS menu bar daemon, CLI and MCP server for AI computer use. It supports background input and returns an evidence verdict for each action: confirmed, noEffect or unverifiable. Free and open source.
Published · Updated . Features and availability describe the product at the update date.
On September 9, 2026, kageroumado released Rocuronium 1.0, a free and open-source macOS menu bar daemon that lets an AI agent see and operate a Mac while its owner keeps working. Rocuronium supports background reading, clicking and typing through accessibility APIs and per-process events. Actions that require the real cursor use a visible consent flow. It is controlled through a command-line tool and a Model Context Protocol (MCP) server.
Most actions are delivered as ghost input, through accessibility APIs and per-process posted events that stay invisible to the person at the keyboard. Every action returns an evidence verdict: confirmed, noEffect or unverifiable. Rocuronium reads text back after writing, compares pixels before and after, counts windows and detects process exits, because macOS accessibility calls frequently report success without changing anything.
Perception is incremental. The read and screenshot commands return an observation token, and passing it back returns only what changed, which costs agents a fraction of a full screenshot. For apps without a usable accessibility tree, a local UI-element detector with OCR, and an optional local vision-language model, locate controls on screen.
When an action needs the real cursor, a visible escort, an on-screen narration of intent and a consent prompt keep the user informed, and the key combination Control-Option-Shift-Escape halts the engine within about 8 milliseconds. Windows can also be parked on a virtual display, operated there with full hardware input, and returned. Multi-step sequence plans run inside the daemon with postcondition checks between steps.
Key features
- Background clicking, typing and reading without moving the cursor
- Evidence verdicts on every action: confirmed, noEffect, unverifiable
- Diff-based reading and screenshots via observation tokens
- Local OCR and vision grounding for apps with no accessibility tree
- MCP server exposing every command as a typed tool
- Hardware kill switch, presence overlay and virtual display isolation
How it compares
| App | Price | License | Your cursor and screen | Runs on |
|---|---|---|---|---|
| Rocuronium | Free | MIT | The agent gets its own login session, desktop and cursor | macOS 26.5 or later; Apple silicon |
| Peekaboo | Free | MIT | Background input when it can resolve the process; some key chords need foreground consent | macOS 15 or later |
| Cua Driver | Free; optional Cua Spaces plans from US$15/month | MIT | Uses apps without taking over your cursor or keyboard | macOS, Windows and Linux |
| Claude computer use | Pro, Max or Ultra plan; research preview | Proprietary | Uses your screen and hides other apps while it works | macOS (CLI); macOS and Windows (Desktop) |
| Codex Computer Use | Included in ChatGPT plans, in specific regions | Proprietary | macOS: runs in the background; Windows: takes the pointer and foreground | macOS and Windows (ChatGPT desktop app) |
Availability
Rocuronium is available now as a free download from kagerou.glass/rocuronium and GitHub. It requires macOS 26.5 or later, including macOS 27, on a Mac with Apple silicon, with Accessibility permission (and Screen Recording for screenshots and OCR). The source code is published under the MIT License.
About kageroumado
I'm Mado Kagerou, publishing as kageroumado: an independent researcher and expert Swift engineer with more than ten years of software development experience. As a founder and technical lead, I take projects from architecture through release. My research interests center on computational pharmacology, computational neuroscience and AI safety. My engineering specialties include Apple platform internals, reverse engineering, GPU rendering and local model training.
Research, engineering work and leadership experience
Media contact
kageroumado · [email protected] · kagerou.glass