Control

MCP server that gives Claude Code full mouse, keyboard, and app control on macOS. Tell Claude what to do and watch it operate your computer.

Pair with mcp-screen for vision + control.

Setup

Add to ~/.claude/settings.json:

{
  "mcpServers": {
    "control": {
      "command": "npx",
      "args": ["mcp-control"]
    }
  }
}

Restart Claude Code. Done.

Usage

Just tell Claude what to do:

• "Open Safari and go to github.com"
• "Click the search box and type my query"
• "Fill in this form - name is Andre, email is..."
• "Switch to Slack and send a message"
• "Scroll down on this page"

Tools

Mouse

Tool	Description
mouse_move	Move cursor to (x, y)
mouse_click	Click at (x, y) with left/right, single/double/triple
mouse_drag	Drag from one point to another
mouse_scroll	Scroll up/down/left/right
get_cursor_position	Get current cursor coordinates

Keyboard

Tool	Description
type_text	Type a string into the focused field
press_key	Press a key combo (e.g. cmd+c, tab, enter)

App Control

Tool	Description
get_frontmost_app	Get focused app name and bundle ID
focus_app	Bring an app to the front
list_windows	List all visible windows with positions and sizes

Batch

Tool	Description
run_actions	Execute a sequence of actions in one call - click, type, press keys, switch apps

The run_actions tool is the key for speed. Instead of one tool call per action (each with Claude think time), it batches everything into a single round-trip.

How It Works

Uses macOS Core Graphics events via JXA (JavaScript for Automation) and AppleScript. Zero external dependencies - just built-in macOS tools.

• Mouse control: CGEvent API (sub-millisecond)
• Keyboard: AppleScript keystroke (handles Unicode)
• Key combos: CGEvent with modifier flags
• App control: AppleScript

Permissions

Your terminal needs Accessibility permission:

System Settings > Privacy & Security > Accessibility - enable your terminal app.

Requirements

• macOS
• Node.js >= 18

Author

@voidmode

License

MIT