An MCP Server That Lets AI Actually Use Windows

What it does

Windows-MCP is an MCP server that translates LLM requests into native Windows UI actions—clicking, typing, window management, and file navigation. It sits between your AI client and the operating system, effectively turning the entire Windows desktop into a callable tool. The project supports Windows 7 through 11 and is built to work with any LLM, vision models optional.

The interesting bit

Instead of relying on computer vision or fine-tuned models to locate buttons, it uses the Windows UI Automation API to read interface state and interact with elements programmatically. There is even a special DOM mode that strips away browser chrome and exposes only web page content to the agent, with a fallback for Firefox via IAccessible2.

Key highlights

• Works with any LLM; no dependency on vision models or CV pipelines
• Typical action latency of 0.2–0.5 seconds between clicks/keystrokes
• Native browser automation via DOM mode for Chrome, Edge, and Firefox
• Can install itself as a per-user Windows Scheduled Task for always-on operation
• Supports bridging from WSL so Linux-side AI clients can still control Windows apps

Caveats

• The README claims 2M+ users via the Claude Desktop Extensions directory, though it is unclear whether this reflects active installations or directory impressions
• English must be set as the Windows default language; otherwise the App-Tool should be disabled
• First-time installation may timeout while dependencies resolve

Verdict

Useful if you need an LLM to drive real Windows applications for testing or automation. Not for you if you are on macOS or uncomfortable with AI agents executing live UI actions.