windows-computer-use-mcp

Windows Computer Use for AI Agents — click, screenshot, type, drag, OCR, and verify native Windows UI via MCP. The missing hands for your agentic workflows.

Exhibit A: 100 Tauri/NSIS installers, one unattended run, $2 in LLM costs.
Install → screenshot → verify → report — zero human intervention. That's the gap this fills.

Built on pywinauto. Exposed as a clean FastMCP server.

Pair with virtualization-mcp for Windows Sandbox / VM isolation. Read docs/SAFETY.md before production use.

📖 INSTALL.md — desktop installer, uv setup, MCP client config

What It Can Do

• Window Management — find, activate, maximize, minimize, position, close
• Mouse Control — click, double-click, right-click, drag, scroll, hover, precise coordinates
• Keyboard Input — type text, send key combinations, shortcuts
• UI Element Interaction — click controls, read/set text, verify states, wait for elements
• Visual Intelligence — screenshots, OCR text extraction, image recognition
• Closed-Loop Tasks — multi-step automation with retry, refocus recovery, evidence trails
• Face Recognition — optional plugin for webcam-based verification workflows

Quick Start

git clone https://github.com/sandraschi/windows-computer-use-mcp.git
cd windows-computer-use-mcp
just bootstrap
just serve

MCP (stdio) — add to Cursor / Claude Desktop:

{
  "mcpServers": {
    "windows-computer-use": {
      "command": "uv",
      "args": ["--directory", "<PATH-TO-CLONE>", "run", "windows-computer-use-mcp"]
    }
  }
}

Optional web operator UI: .\start.ps1 → http://127.0.0.1:10788 (proxies API on 10789).

Demos: just demo — mouse dance, Notepad grid, typewriter (see examples/README.md).

Works With

• Claude Desktop (MCP native)
• DeepSeek via OpenCode/API
• Any FastMCP-compatible agent

Python Stack

See docs/py-stack.md for what each package does and how it plugs into the tool surface.

Optional invasive features (off by default)

Default install: window/element/mouse/keyboard/visual automation only. No ambient monitoring.

Keylogger is not stealth spyware. It is an explicit MCP tool: disabled unless you set the env flag, requires HITL approval to start, stores events in a bounded in-memory buffer (not hidden files), and the server stops the hook on shutdown. Use only on machines you own for debugging shortcut/focus issues — not for credential harvesting.

Face recognition is not a background plan-only sketch — the tool is implemented but opt-in like the keylogger. The operator UI Biometrics page is a control panel when you enable it; it does not run face matching unless you opt in and call the tool.

Documentation

An MCP prompt is also registered: desktop_automation_operator_protocol (see src/windows_computer_use_mcp/prompts.py).

Ports

Stdio MCP has no port (host-launched process).

Tech Stack

• MCP: Python 3.12+, FastMCP 3.2+, portmanteau tools (windows, elements, mouse, keyboard, visual, assert, dialog, shortcut, task, system)
• Input: win32_mouse (DPI-aware), optional Win32 keyboard (CUA_MCP_KEYBOARD=win32)
• Web: Vite + React operator dashboard, local LLM proxy (Ollama / LM Studio)
• Desktop: Tauri 2 operator app (optional installer)
• Quality: Ruff, Biome, pytest (environment-aware), just recipes

Related

Fleet standards: mcp-central-docs.

Browser vs desktop: This server drives Win32 / UI Automation. For HTML/DOM and websites, use a Playwright browser MCP alongside this one.

License

MIT — Copyright (c) 2026 Sandra Schipal.