Skip to main content
Glama

UseMyMac

Claude Code on one machine driving a Mac through UseMyMac

Left: Claude Code, anywhere. Right: a Mac running usemymac up. The agent reaches it only through the printed URL.

Turn your Mac into a remote computer-use host. One command on the Mac, one paste into Claude Code or Codex, and the agent can see your screen, click, type, and open apps.

Built on CUA's computer-server for the macOS driver. UseMyMac adds the parts that make it safe to reach from outside: a bearer token, a tool allowlist that hides shell and filesystem access by default, a Cloudflare quick tunnel, keep-awake, and a permissions preflight.

Quick start

uvx usemymac up

Or without uv installed:

curl -fsSL https://raw.githubusercontent.com/yunfeng-enrich/usemymac/main/install.sh | sh

Either way UseMyMac fetches cloudflared on first run and starts the server. It prints one line to paste into your agent, plus the one-line commands:

claude mcp add --transport http usemymac https://<host>/<token>/mcp
codex mcp add usemymac --url https://<host>/<token>/mcp

The token lives in the URL so no client needs custom headers. Authorization: Bearer <token> against https://<host>/mcp works too.

Related MCP server: Automation MCP

Permissions

macOS grants Screen Recording and Accessibility per app, and the app is whichever terminal launched usemymac. The first run triggers both prompts. Screen Recording only takes effect after that terminal is fully quit and reopened. Without it, screenshots are black and window titles are empty. Check with:

usemymac check

What is exposed

By default: screenshot, mouse, keyboard, clipboard, windows, app launch, and the accessibility tree. Not exposed unless you ask:

  • --allow-shell adds computer_run_command

  • --allow-files adds file read, write, and delete tools

Other flags: --no-tunnel for LAN or SSH-forwarded use, --width/--height to scale screenshots and coordinates for the model, --port, --token.

Security

This is your real desktop on a public URL. The token is the only lock, and it is in the URL, so treat the URL like a password. Both the hostname and the token are new on every run, so a leaked URL stops working the moment you Ctrl-C. Keep the terminal visible while an agent is connected. Pass --token only if you need a fixed one.

Status

Spike. Works end to end on macOS 15 with Claude Code over a Cloudflare tunnel. Not yet: signed menu bar app, persistent tunnel hostnames, per-app allowlists, Lume VM sandbox mode.

Available Tools

34 tools
computer_activate_windowA

Bring a window to the foreground and focus it.

Inputs: window_id (required). Returns: object

computer_clickB

Click at the specified screen coordinates.

Inputs: x (required), y (required), button. Returns: object

computer_clipboard_getA

Get the current clipboard content.

Returns: Dictionary with clipboard content.

Inputs: no named parameters. Returns: object

computer_clipboard_setC

Set the clipboard content.

Inputs: text (required). Returns: object

computer_close_windowC

Close a window.

Inputs: window_id (required). Returns: object

computer_double_clickB

Double-click at the specified screen coordinates.

Inputs: x (required), y (required). Returns: object

computer_dragB

Drag from start coordinates to end coordinates.

Inputs: end_x (required), end_y (required), start_x (required), start_y (required), +2 more. Returns: object

computer_find_elementB

Find a UI element in the accessibility tree.

Inputs: role, title, value. Returns: object

computer_get_accessibility_treeA

Get the accessibility tree of the current window.

This provides detailed information about UI elements that can be useful for finding clickable elements or understanding the UI structure.

Returns: Dictionary with the accessibility tree structure.

Inputs: no named parameters. Returns: object

computer_get_active_windowA

Get the currently active window ID.

Returns: Dictionary with 'window_id' key.

Inputs: no named parameters. Returns: object

computer_get_app_windowsB

Get all windows belonging to an application.

Inputs: app (required). Returns: object

computer_get_cursor_positionA

Get the current cursor position.

Returns: Dictionary with 'x' and 'y' coordinates.

Inputs: no named parameters. Returns: object

computer_get_desktop_environmentA

Get information about the current desktop environment.

Returns: Dictionary with desktop environment details.

Inputs: no named parameters. Returns: object

computer_get_desktop_stateA

Capture the whole desktop through the active Cua Driver session.

This tool is available only with CUA_BACKEND=cua-driver. In an auto session, explicitly escalate the capture scope only after the window-focused ladder is exhausted.

Inputs: screenshot_out_file. Returns: object

computer_get_screen_sizeA

Get the screen dimensions.

Returns: Dictionary with 'width' and 'height' keys in pixels.

Inputs: no named parameters. Returns: object

computer_get_window_nameA

Get the title/name of a window.

Inputs: window_id (required). Returns: object

computer_get_window_positionB

Get the position of a window.

Inputs: window_id (required). Returns: object

computer_get_window_sizeB

Get the size of a window.

Inputs: window_id (required). Returns: object

computer_hotkeyB

Press a key combination (hotkey).

Inputs: keys (required). Returns: object

computer_key_downC

Press and hold a key.

Inputs: key (required). Returns: object

computer_key_upA

Release a held key.

Inputs: key (required). Returns: object

computer_launch_appB

Launch an application.

Inputs: app (required), args. Returns: object

computer_maximize_windowC

Maximize a window.

Inputs: window_id (required). Returns: object

computer_minimize_windowC

Minimize a window.

Inputs: window_id (required). Returns: object

computer_mouse_downA

Press and hold a mouse button.

Inputs: x, y, button. Returns: object

computer_mouse_upC

Release a mouse button.

Inputs: x, y, button. Returns: object

computer_moveA

Move the cursor to the specified screen coordinates.

Inputs: x (required), y (required). Returns: object

computer_openA

Open a file or URL with the default application.

Inputs: target (required). Returns: object

computer_press_keyB

Press a single key.

Inputs: key (required). Returns: object

computer_screenshotA

Capture a screenshot of the current screen.

Returns the current screen state as an image. Always call this first to see what's on screen before performing any actions.

Inputs: no named parameters.

computer_scrollC

Scroll at the specified position.

Inputs: x (required), y (required), scroll_x, scroll_y. Returns: object

computer_set_window_positionB

Set the position of a window.

Inputs: x (required), y (required), window_id (required). Returns: object

computer_set_window_sizeC

Set the size of a window.

Inputs: width (required), height (required), window_id (required). Returns: object

computer_typeB

Type text at the current cursor position.

Inputs: text (required). Returns: object

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    B3.4/5.0

    Scored across 34 tools

    Disambiguation4/5

    Most tools are clearly distinct (screenshot, click, type, window management), but computer_get_desktop_state overlaps with computer_screenshot, and computer_get_active_window/computer_get_window_name could be confused. The mouse_down/up and key_down/up pairs are distinct but similar in purpose.

    Naming Consistency5/5

    All tools follow a consistent computer_verb_noun pattern (e.g., computer_get_cursor_position, computer_set_window_size, computer_activate_window). The naming is uniform and predictable across the entire set.

    Tool Count4/5

    34 tools is on the higher end, but the scope is broad (screen, mouse, keyboard, clipboard, windows, accessibility), so each tool covers a distinct action. It feels slightly heavy but justified for a full computer control server.

    Completeness5/5

    The tool surface covers the full lifecycle of computer interaction: observation (screenshot, accessibility tree), input (mouse, keyboard, clipboard), window management (get/set/activate/minimize/maximize/close), and app launching. No obvious dead ends for an agent automating a Mac.

    Maintenance

    ActivityStale
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables AI services like Claude and Cursor to remotely control a Mac by executing shell commands, managing files, and running AppleScript for UI automation. Access is secured through OAuth 2.0 authentication and encrypted tunnels to protect remote interactions.
      2
      MIT
    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables AI assistants to automate macOS desktop tasks including mouse control, keyboard input, screenshots, window management, and UI interaction.
      417
      MIT
    • A
      license
      Not graded
      quality
      D
      maintenance
      Provides native macOS desktop automation for AI agents, enabling screen capture, mouse/keyboard control, window management, and iOS/Android simulator control in both foreground and background modes without focus stealing.
      5
      MIT

    Appeared in Searches