mac-computer-use lets Claude Code, Cursor and Codex click, type and read back on macOS 14

Reading a checklist to an AI agent is easier than watching it click through your web app for the seventh time. mac-computer-use is a new open-source Agent Skill for macOS that gives a coding agent the tools to observe your windows, act on them, and verify what actually happened, all from the same package no matter which agent runtime you use.

The author, To3akaRin, published it under the MIT licence on GitHub. As of 2026-09-27 the repository shows 1,127 stars in the last 30 days and 16 forks, with the most recent push on 2026-09-24. It follows the open Agent Skills specification, so it works from Claude Code, Codex, Kimi Code, Cursor, OpenClaw, WorkBuddy, Qwen Work and ZCode without a runtime-specific plugin.

What the skill gives your agent

The package exposes native macOS controls and an embedded Chromium bridge through a set of local commands. The author's own capability table lists reading Accessibility (AX) controls, listing windows and processes, sending native Unicode text, clicking, hovering, scrolling and keyboard shortcuts, taking window captures with ScreenCaptureKit, and running batches of Chrome DevTools Protocol commands against pages that already expose a debug port.

Two rules are baked into the tools rather than left to the model. Every write supports a --dry-run preview that checks the target and parameters before it moves the mouse or types. Foreground actions require 2 seconds of user inactivity, wait at most 15 seconds, use a session lock, and abort the moment you touch the keyboard. An unconfirmed write returns unknown instead of being replayed through a second channel.

What was verified

The author states that recorded acceptance covers dedicated native test apps and an isolated browser on macOS 15.7.7 with Apple Silicon, Swift 6.1.2, and Node.js versions 22.23.1 and 24.18.0. The same page says compatibility with every third-party app is not established, and that Intel Macs and macOS 14 still need physical-device coverage. Screenshot capture and browser control need Accessibility and Screen Recording permissions on the host terminal or agent app.

One published example: from a small set of reference photos, an agent produced a rotatable 3D exhibition space in Blender in about 12 minutes without the user touching the mouse. The author reports 3,388 triangles across 117 meshes, a single GLB file of around 218 KiB, and an editable Blender source in the repository.

The tools run locally, need no hosted backend or API key of their own, and carry zero runtime npm dependencies for the CDP client. The agent's model and its own runtime still bring their own requirements.

The bigger point for a team using several coding agents is that a desktop control layer is now one dependency and one permission grant, shared across every runtime that reads the skill package. Test recording, screenshot-based bug reports and long clicking-through-a-UI tasks stop being tied to one vendor's computer-use offering, and running them locally means the pixels never leave the machine.

Source

• Repository and README: To3akaRin/mac-computer-use on GitHub
• Open specification the skill follows: Agent Skills specification at agentskills.io