Click accuracy is tuned for that resolution on a single screen, and running it in a virtual machine is the recommended shape rather than a footnote — this hands a model control of a real mouse and keyboard. Known gaps are stated plainly too: window minimize and restore are unsupported, and multiple screens may not behave.
A Windows control server for MCP: programmatic mouse, keyboard, window management, screen capture and clipboard.
- Mouse — movement, click and drag, scrolling and position tracking
- Keyboard — text input, key combinations, press and release, and hold
- Windows — list all windows, get the active one, focus, resize and reposition
- Screen — screenshots, screen size detection and active-window capture
- Clipboard operations
- Three automation providers you can mix per operation: keysender by default, powershell for simpler work, and autohotkey for AutoHotkey v2 scripting
- SSE transport, so the server can run on a VM and your client can connect to it across the network
Windows only — the keysender dependency is Windows-specific. Visual Studio Build Tools with the VC++ workload, Python and Node.js, then npm install -g mcp-control. Start it with mcp-control --sse; the default port is 3232, and --https with --cert and --key is mandatory for anything remote per the MCP spec. Providers are chosen with AUTOMATION_PROVIDER, or per operation with AUTOMATION_KEYBOARD_PROVIDER and friends.
One command — npm install -g mcp-control