macos-computer-use-mcp
# macos-computer-use-mcp
A standalone MCP server providing native macOS computer control — mouse, keyboard, screenshots, and app management — for any MCP-compatible agent.
## Demo
https://github.com/user-attachments/assets/130bee07-1f87-4291-8483-fc47ec51e493
> **Screenshot & Click in action** — The MCP server grants app access, captures a 2560×1664 native screenshot, clicks on-screen elements, and opens URLs — all orchestrated through standard MCP tool calls from a coding agent.
## Compatibility
Works with any client that supports the [Model Context Protocol](https://modelcontextprotocol.io/), including:
- **Claude Code** (`claude mcp add`)
- **OpenAI Codex** (`~/.codex/config.toml`)
- **Cursor** (`~/.cursor/mcp.json`)
- Any other MCP-compatible agent or IDE
## How It Works
The server exposes macOS system control as MCP tools. Under the hood it uses macOS native modules for low-level input simulation and system APIs:
- **`@ant/computer-use-input`** — Low-level mouse and keyboard event injection
- **`@ant/computer-use-swift`** — macOS native APIs for display management, app control, and screenshots
The MCP server process communicates over stdio, so any agent can spawn it as a subprocess and call its tools via the standard JSON-RPC protocol.
## Available Tools (24)
| Tool | Description |
|------|-------------|
| `request_access` | Request Accessibility permission for an app |
| `screenshot` | Capture the full screen |
| `zoom` | Zoom into a screen region |
| `left_click` | Left-click at coordinates |
| `right_click` | Right-click at coordinates |
| `middle_click` | Middle-click at coordinates |
| `double_click` | Double-click at coordinates |
| `triple_click` | Triple-click at coordinates |
| `type` | Type a string of text |
| `key` | Press a key or key combination |
| `cursor_position` | Get current mouse position |
| `mouse_move` | Move the cursor to coordinates |
| `scroll` | Scroll at coordinates |
| `drag` | Drag from one point to another |
| `left_click_drag` | Left-click and drag |
| `get_display_size` | Get screen dimensions |
| `list_displays` | List all connected displays |
| `get_frontmost_app` | Get the currently active application |
| `list_installed_apps` | List all installed applications |
| `open_app` | Open an application by name or bundle ID |
| `close_app` | Close an application |
| `focus_app` | Bring an application to the foreground |
| `get_screen_content` | Get accessibility tree for screen content |
| `wait` | Wait for a specified duration |
## Installation
```bash
curl -fsSL https://raw.githubusercontent.com/Zooeyii/macos-computer-use-mcp/main/install.sh | bash
```
Or manually:
```bash
git clone https://github.com/Zooeyii/macos-computer-use-mcp.git ~/.local/share/macos-computer-use-mcp
cd ~/.local/share/macos-computer-use-mcp
npm install
npm run build
```
## Configuration
### Claude Code
```bash
claude mcp add -s user computer-use-standalone node $HOME/.local/share/macos-computer-use-mcp/dist/cli.js
```
Or add to `~/.claude/mcp.json`:
```json
{
"computer-use-standalone": {
"type": "stdio",
"command": "node",
"args": ["$HOME/.local/share/macos-computer-use-mcp/dist/cli.js"]
}
}
```
### OpenAI Codex
Add to `~/.codex/config.toml`:
```toml
[mcp_servers.computer-use-standalone]
command = "node"
args = ["$HOME/.local/share/macos-computer-use-mcp/dist/cli.js"]
```
Or via CLI:
```bash
codex mcp add computer-use-standalone -- node $HOME/.local/share/macos-computer-use-mcp/dist/cli.js
```
### Cursor
Add to `~/.cursor/mcp.json`:
```json
{
"mcpServers": {
"computer-use-standalone": {
"command": "node",
"args": ["$HOME/.local/share/macos-computer-use-mcp/dist/cli.js"]
}
}
}
```
## Requirements
- **macOS** (Darwin) — macOS-only due to native module dependencies
- **Node.js 18+**
- **Accessibility Permission** — Required for mouse/keyboard control
- System Settings → Privacy & Security → Accessibility
- **Screen Recording Permission** — Required for screenshots
- System Settings → Privacy & Security → Screen Recording
## Architecture
```
MCP Client (Claude Code / Codex / Cursor / any agent)
│
│ stdio (JSON-RPC / MCP protocol)
│
▼
macos-computer-use-mcp (this server)
│
├── MCP Server
│ └── Tool handler
│
├── Tool Definitions
│ ├── Input tools (click, drag, scroll, type, key)
│ ├── Screen tools (screenshot, zoom, display info)
│ └── App tools (open, close, focus, list)
│
└── Executor
│
├── @ant/computer-use-input.node
│ └── Mouse / keyboard event injection
│
└── @ant/computer-use-swift
└── macOS native APIs
├── App management
├── Display control
└── Screenshot capture
```
## Project Structure
```
macos-computer-use-mcp/
├── src/
│ ├── cli.ts # MCP server entry point
│ ├── tools.ts # Tool definitions
│ └── executor.ts # Platform implementations
├── install.sh # One-line installer
├── package.json
├── tsconfig.json
├── tsup.config.ts
└── README.md
```
## Development
```bash
# Install dependencies
npm install
# Build
npm run build
# Run directly
node dist/cli.js
# Type-check only
npm run typecheck
```
## Disclaimer
This project is for educational and research purposes.
Native module interfaces are based on publicly observable runtime behavior.
Use at your own risk. Only run in trusted environments — computer use grants full control of your mouse, keyboard, and screen.
## License
MIT
TDQS
Scored across 24 tools
Each tool targets a distinct action (e.g., click types, keyboard, screenshots, clipboard, application control). The many click tools are clearly differentiated by modifiers (double, triple, right, drag, press/release), and composite actions like computer_batch occupy their own niche.
All tool names use lowercase with underscores for multi-word names (e.g., left_click_drag, list_granted_applications). Simple actions use short single words (key, type, wait). The pattern is uniform and predictable.
24 tools is on the higher end, but each tool covers a necessary primitive for computer control (clicks, keyboard, screenshots, clipboard, app management, batching). The count is justified given the breadth of actions an agent might need.
The tool set covers all major interaction primitives: mouse, keyboard, screenshots (full and zoomed), display switching, clipboard, application launching, permissions management, and batching. There are no obvious gaps for basic computer use tasks.