Windows Computer Use

A Windows desktop-control backend for Claude Code, delivered as an MCP server.

Takes screenshots, clicks, types, scrolls, and manages windows on your Windows desktop -- giving Claude Code the same kind of computer interaction that Anthropic's first-party Computer Use provides on Mac.

What this is

An MCP server that exposes three tools:

• computer -- Anthropic-compatible action model: screenshot, click, move, type, key, scroll, drag
• list_windows -- Enumerate visible windows with stable HWND-based IDs, PIDs, titles, and bounds
• focus_window -- Bring a window to the foreground by its window ID

Claude Code connects over stdio and uses these tools in a perception-action loop: screenshot the desktop, analyze what's on screen, take an action, screenshot again.

What this is not

This is not a full clone of Anthropic's first-party Mac experience. It is:

• Local and installable via npm
• Standard-user only (cannot interact with elevated/admin windows)
• Subject to Windows focus and UAC limitations
• Desktop screenshot based (no occluded/minimized window capture)

For the target audience -- Windows users who want Claude Code to see and control their desktop -- this is still valuable.

Install

cd /path/to/windows-computer-use
npm install
npm run build

No local C++ toolchain required. koffi and sharp ship prebuilt native binaries.

Configure Claude Code

Add to your Claude Code settings.json (user or project scope):

{
  "mcpServers": {
    "computer-use": {
      "command": "node",
      "args": ["/path/to/windows-computer-use/build/index.js"]
    }
  }
}

Restart Claude Code. The tools will appear automatically.

Tools

computer

The primary tool. Dispatch actions via the action field:

Action	Required Params	Description
screenshot	--	Capture the full virtual desktop as a scaled JPEG
screenshot	window_id	Capture a specific window's visible rect
click	coordinate	Click at [x, y] in scaled space
move	coordinate	Move cursor to [x, y]
type	text	Type a text string
key	key	Press a key combo (e.g., ctrl+s, alt+f4)
scroll	coordinate, direction	Scroll up/down/left/right
drag	from_coordinate, to_coordinate	Click-drag between points

Optional params: button (left/right/middle), count (click count), amount (scroll notches).

Coordinates are in the scaled screenshot space. Take a screenshot first, then use coordinates from that image.

list_windows

Returns JSON array:

[
  {
    "window_id": "0x00030014",
    "pid": 19148,
    "title": "My App - Visual Studio Code",
    "bounds": { "left": -7, "top": -7, "right": 2568, "bottom": 1400 },
    "visible": true
  }
]

focus_window

Bring a window to the foreground:

{ "window_id": "0x00030014" }

May fail silently for elevated windows due to Windows UAC restrictions.

Architecture

Claude Code (stdio) --> MCP Server --> koffi FFI --> Win32 APIs
                           |
                       sharp (resize + JPEG encode)
                           |
                       audit.jsonl (redacted)

• Win32 FFI: koffi calling user32.dll (SendInput, EnumWindows, SetForegroundWindow) and gdi32.dll (BitBlt, GetDIBits)
• Image pipeline: Raw BGRA pixels -> sharp resize to fit Anthropic's 1568px/1.15MP constraints -> JPEG quality 80 -> base64
• DPI: PER_MONITOR_AWARE_V2 set at startup
• Coordinates: Virtual desktop aware (multi-monitor from day one)

Security

• Bounds validation: Coordinates outside the scaled virtual desktop are rejected
• Rate limiting: Token-bucket at 10 actions/second (configurable in config.ts)
• Redacted audit log: Every action logged to audit.jsonl with timestamp, tool name, action type, target window, and coordinates. Screenshot payloads and typed text are never logged.
• Standard user: Runs without elevation. Cannot interact with admin windows.
• Localhost only: stdio transport, no network exposure

Known Limitations

• Screenshots use GDI BitBlt -- works for most content but may show black for some hardware-accelerated overlays
• Window screenshots are crops of the visible desktop, not private window captures. Occluded or minimized windows will show whatever is visually on top.
• Aggressive scaling on ultra-wide or multi-monitor setups (the API constraint is 1568px on the longest edge)
• No semantic element discovery (no UI Automation). All interaction is coordinate-based.
• HWND-based window_id values are stable within a session but not across restarts

Dependencies

• @modelcontextprotocol/sdk -- MCP server protocol
• koffi -- Win32 FFI bridge (no C++ toolchain needed)
• sharp -- Image processing (resize, JPEG encode)

License

MIT