windows-computer-use-mcp

# windows-computer-use-mcp

Windows **computer use agent** (CUA-MCP): structured native UI automation via Pywinauto/UIA — read docs/SAFETY.md. Not a browser agent; not covert surveillance.

## Related MCP servers

| Tool | Use for |
|------|---------|
| **[autohotkey-mcp](https://github.com/sandraschi/autohotkey-mcp)** | Low-level input recording/replay via AHK scripts. Use pywinauto for UIA/OCR; use AHK for raw mouse/keyboard macros. |
| **virtualization-mcp** | Sandbox/VM isolation — pair when you need disposable desktops. |

## Default portmanteau tools (always registered)

- **automation_windows**, **automation_elements**, **automation_mouse**, **automation_keyboard**
- **automation_visual**, **automation_assert**, **automation_dialog**, **automation_shortcut**, **automation_task**, **automation_system**
- **get_desktop_state**, **get_window_state**, **approve_automation**, **automation_safety**
- **analyze_winapp** — auto-crawl apps: UI tree walk + SOM screenshot + element map + markdown report. Operations: crawl, discover (shortcut probe), portfolio. Shows "CUA at work" HUD.

## Opt-in only (NOT in default install)

- **automation_face**: local webcam face profiles — `PYWINAUTO_MCP_ENABLE_FACE=1` + `face` extra. Implemented, not background surveillance.
- **global_keylogger**: session keyboard debug hook — `PYWINAUTO_MCP_ENABLE_KEYLOGGER=1`, HITL to start, in-memory buffer, stops on server exit. Shows "CUA at work" HUD while active.

## CUA HUD

All long-running operations (`analyze_winapp`, `global_keylogger`) show a "CUA at work" blinking red overlay with an emergency stop button. The target window is auto-refocused before each action to prevent focus stealers.

## Python stack (namesake)

- **pywinauto** — core UIA/Win32 automation
- **pywin32**, **pygetwindow**, **pyautogui**, **pynput**, **pyperclip**, **opencv-python-headless**, **pytesseract**
- **face_recognition** — optional extra only

## Infrastructure

- Web UI: http://127.0.0.1:10788 (dev) / Tauri operator (install)
- API: http://127.0.0.1:10789/api/v1/*
- MCP HTTP: http://127.0.0.1:10789/mcp

## Docs hub

- Contributor: docs/DEVELOPMENT.md, docs/TESTING.md
- Archived fleet imports: docs/development/archive/

## Metadata

- Package: windows-computer-use-mcp 0.7.x
- Framework: FastMCP 3.4+ (dual transport: stdio + HTTP)
- License: MIT