windows-computer-use-mcp
A tool for agents, and an agent itself.
| You | It |
|---|---|
| Use it as an MCP server | Claude, Cursor, DeepSeek call automation_click, automation_screenshot, automation_ocr — 22 tools |
| Use it as an autonomous agent | Give it a goal: automation_mission(run="install app, verify UI, screenshot result") — it plans, executes, retries, and reports |
| Use it as a webapp | start.ps1 opens a React dashboard at http://127.0.0.1:10788 with HITL, crawler, logging |
| Use it as a desktop app | The NSIS installer bundles everything into one binary — no Python, no uv, no git needed |
Exhibit A: 100 Tauri/NSIS installers, one unattended run, $2 in LLM costs. Install, screenshot, verify, report — zero human intervention. That is what agentic Windows automation looks like at scale.
Built on pywinauto. Read docs/SAFETY.md before production use.
- Quick Start
- Features
- Documentation
- Ports
- License
Quick Start
| Method | Command / Config |
|---|---|
| MCP stdio (Cursor, Claude Desktop) | { "mcpServers": { "windows-computer-use": { "command": "uv", "args": ["--directory", "<PATH>", "run", "windows-computer-use-mcp"] } } } |
| HTTP streamable (any MCP HTTP client) | { "mcpServers": { "windows-computer-use": { "url": "http://127.0.0.1:10789/mcp" } } } |
| Web operator UI | .\start.ps1 → http://127.0.0.1:10788 |
| Desktop app (NSIS installer) | Download from Releases — zero deps |
See INSTALL.md for detailed setup. Run just demo for examples.
Features
- Window Management — find, activate, maximize, minimize, position, close
- Mouse & Keyboard — click, drag, type, hotkeys, app shortcuts
- UI Elements — inspect, click, read text, verify state via UIA / Win32
- Visual Intelligence — screenshots, OCR, template matching
- Autonomous Missions — give it a goal, it plans and executes with retry + verification
- Macro Recording — record any UI sequence, replay, verify outcomes
- Multi-App Workflows — chain actions across Notepad, Calc, Paint, or any Windows app
- Telemetry — every action logged to SQLite; query failure patterns by tool
- Adaptive Location — auto-cascades through title/auto_id/control_id/class/OCR to find elements
- Face Recognition — optional, off by default
Documentation
| Doc | Content |
|---|---|
| INSTALL.md | Setup: desktop app, uv, MCP config |
| docs/README.md | Full documentation hub |
| docs/py-stack.md | Python dependency deep dive |
| docs/composing-with-playwright.md | Browser automation with Playwright MCP |
| docs/ocr.md | OCR system — Tesseract setup, limitations, competition |
| docs/cua-nsis-certification.md | Dogfooding: using the tool to test its own NSIS installer |
| docs/SAFETY.md | HITL, kill switch, opt-in features |
| docs/TOOLS.md | Portmanteau tool reference |
| tests/README.md | Test suite guide and e2e setup |
| examples/README.md | Runnable demos |
| mcpb/README.md | MCPB bundle packaging |
| web_sota/README.md | Operator UI build/dev guide |
| CHANGELOG.md | Release history |
Ports
| Port | Service |
|---|---|
| 10788 | Frontend — Vite operator UI |
| 10789 | Backend — FastAPI + FastMCP HTTP |
| stdio | MCP transport (port-free) |
Related
| Repo | What it does |
|---|---|
| autohotkey-mcp | Raw input recording/replay via AHK |
| browser-mcp | Playwright browser control — for webapps, HTML DOM, websites |
| virtualization-mcp | Sandbox / VM isolation |
| windows-operations-mcp | Registry, services, accounts |
Browser vs desktop: This server drives Win32 / UI Automation. For HTML/DOM and websites, pair with browser-mcp (Playwright). Both MCPs can run side by side — use one profile that loads both and let the LLM pick the right tool for the target.
Fleet standards: mcp-central-docs.
License
MIT — Copyright (c) 2026 Sandra Schipal.
