简体中文 · 日本語

---

Features

	Feature	Description
Vision	Screenshot & Display	Capture any display, enumerate monitors, zoom into regions
Input	Mouse & Keyboard	Click, drag, scroll, type, key combos, hold keys — with IME-safe clipboard routing
Apps	Application Control	Launch apps, detect frontmost app, list installed/running apps, tiered permission model
Clipboard	Read & Write	Full clipboard access for paste-based workflows
Batch	Action Batching	Chain multiple actions in a single MCP call for speed
Runtime	Zero-Config Bootstrap	Auto-creates Python virtualenv and installs dependencies on first run
Portable	Skill Packaging	Ships as a standalone skill — install once, works without the source repo
Public	No Private Dependencies	Built entirely on public packages: Node.js, Python, pyautogui, mss, Pillow, pyobjc

Quick Start

1. Clone & build

git clone https://github.com/wimi321/macos-computer-use-skill.git
cd macos-computer-use-skill
npm install && npm run build

2. Run the MCP server

node dist/cli.js

On first launch the server automatically creates a Python virtualenv in .runtime/venv and installs all runtime dependencies. No Claude desktop app, no private native modules.

3. Or install from ClawHub

clawhub install computer-use-macos

[!NOTE] macOS requires Accessibility and Screen Recording permissions for the host process. The server checks both on startup and reports status through MCP.

Architecture

flowchart LR
    A[AI Agent / MCP Client] --> B[MCP Server<br/>TypeScript + stdio]
    B --> C[Tool Layer<br/>28 MCP tools]
    B --> D[Python Bridge<br/>auto-bootstrapped venv]
    D --> E[pyautogui]
    D --> F[mss + Pillow]
    D --> G[pyobjc<br/>Cocoa + Quartz]
    E --> H[Mouse / Keyboard]
    F --> I[Screenshots]
    G --> J[Apps / Displays<br/>Clipboard / Windows]

Available Tools

Vision & Display

Tool	Description
screenshot	Capture the current display as a JPEG image
zoom	Crop and zoom into a region of the last screenshot
switch_display	Switch the active capture target to a different monitor

Input

Tool	Description
left_click	Left-click at a coordinate
double_click	Double-click at a coordinate
triple_click	Triple-click (select paragraph/line)
right_click	Right-click (context menu)
middle_click	Middle-click
left_click_drag	Click-and-drag between two points
left_mouse_down	Press and hold the left mouse button
left_mouse_up	Release the left mouse button
mouse_move	Move the cursor without clicking
scroll	Scroll in any direction at a coordinate
type	Type text (clipboard-routed on macOS to avoid IME corruption)
key	Press a key combo (e.g. cmd+c, ctrl+shift+t)
hold_key	Hold a key for a duration
cursor_position	Get the current cursor coordinates

Application & System

Tool	Description
open_application	Launch a macOS application by name
request_access	Request access to interact with an application
list_granted_applications	List apps the current session has permission to control
read_clipboard	Read the system clipboard
write_clipboard	Write to the system clipboard
wait	Pause for a specified duration

Batch & Teach Mode

Tool	Description
computer_batch	Execute multiple actions in a single call
request_teach_access	Request elevated access for teaching workflows
teach_step	Single-step action in teach mode
teach_batch	Batch actions in teach mode

MCP Configuration

Add to your MCP client config:

{
  "mcpServers": {
    "computer-use": {
      "command": "node",
      "args": ["/absolute/path/to/macos-computer-use-skill/dist/cli.js"],
      "env": {
        "CLAUDE_COMPUTER_USE_DEBUG": "0",
        "CLAUDE_COMPUTER_USE_COORDINATE_MODE": "pixels"
      }
    }
  }
}

See examples/mcp-config.json for a ready-to-use template.

Skill Install

This project ships as a self-contained skill at skill/computer-use-macos.

From ClawHub:

clawhub install computer-use-macos

From the repo:

bash skill/computer-use-macos/scripts/install.sh

The installer copies the full project to ~/.codex/skills/computer-use-macos/project — the skill keeps working even if the original clone is removed.

Environment Variables

Variable	Default	Description
CLAUDE_COMPUTER_USE_DEBUG	0	Enable verbose debug logging
CLAUDE_COMPUTER_USE_COORDINATE_MODE	pixels	Coordinate mode: pixels or normalized_0_100
CLAUDE_COMPUTER_USE_CLIPBOARD_PASTE	1	Prefer clipboard-based typing (IME-safe)
CLAUDE_COMPUTER_USE_MOUSE_ANIMATION	0	Animate mouse movement
CLAUDE_COMPUTER_USE_HIDE_BEFORE_ACTION	0	Hide overlay windows before actions

Requirements

Requirement	Version
macOS	12+ (Monterey or later)
Node.js	20+
Python	3.10+ (ships with macOS or via Homebrew)
Permissions	Accessibility + Screen Recording

Python dependencies (pyautogui, mss, Pillow, pyobjc) are installed automatically into an isolated virtualenv on first run.

Repository Layout

macos-computer-use-skill/
├── src/
│   ├── cli.ts                    # Entry point
│   ├── server.ts                 # MCP server setup
│   ├── session.ts                # Session context factory
│   ├── computer-use/
│   │   ├── executor.ts           # macOS executor (bridges to Python)
│   │   ├── pythonBridge.ts       # Venv bootstrap + Python IPC
│   │   ├── hostAdapter.ts        # Host adapter factory
│   │   └── ...
│   └── vendor/computer-use-mcp/
│       ├── mcpServer.ts          # MCP server factory
│       ├── toolCalls.ts          # Tool dispatch logic
│       ├── tools.ts              # MCP tool schemas
│       └── ...
├── runtime/
│   ├── mac_helper.py             # Python runtime (pyautogui + pyobjc)
│   └── requirements.txt
├── skill/
│   └── computer-use-macos/       # Portable skill package
├── examples/
│   ├── mcp-config.json
│   └── env.sh.example
├── assets/
│   └── hero.svg
├── package.json
└── tsconfig.json

Roadmap

• App icon extraction without private APIs
• Stronger nested helper-app filtering
• Automated MCP integration test suite
• Pre-built release artifacts for easier distribution

Contributing

Contributions are welcome. See CONTRIBUTING.md for guidelines.

License

MIT

Acknowledgments

This project extracts and adapts reusable TypeScript computer-use logic from the Claude Code workflow, replacing the private native runtime with a fully standalone, publicly installable macOS implementation. Built on top of the Model Context Protocol.