Windows Computer Use

Claude play-testing a live Unreal Engine project — entering Play-In-Editor and running up a ramp, driven entirely through this MCP. The full editor stays on screen, so you can see it really is the desktop app.

play(
    target="window:MyProject - Unreal Editor",
    script="""
        lmb 557 460
        wait 0.4
        look 150 0
        wait 0.15
        look 150 0
        wait 0.2
        down w
        wait 1.7
        tap space
        wait 0.7
        up w
        wait 0.4
    """,
)

A Model Context Protocol server that gives a Claude agent full control of the local Windows desktop — native screen capture, low-level input injection, video recording, and a play-test loop for driving games and apps.

Unlike Anthropic's sandboxed computer-use tool, this runs on the machine it controls: it reads the actual current displays (no resolution requests), is multi-monitor and per-monitor DPI aware, and injects input via SendInput scan codes so it works in games that ignore synthetic virtual-key events. Built for full Claude control — no security gating.

Why this exists

Anthropic's official computer use in the Claude Code CLI is a macOS-only research preview — Pro/Max only, interactive sessions only (not available with the -p flag). The cross-platform alternative is the Claude Desktop app. There is no official, non-Desktop computer-use for Windows: nothing you can drive headlessly from claude -p, from the API, or wire into an agent over MCP.

This server fills that gap. It's a standard MCP server, so it works on Windows in Claude Code (interactive and -p), in Claude Desktop, or from any MCP client / custom agent — with no plan gating — and it's tuned for what a Windows agent actually needs that the sandboxed cloud tool can't do: real multi-monitor capture, per-window GPU capture, game-grade input, and play-testing.

Tools

Coordinates & multi-monitor

Coordinates are physical pixels in virtual-desktop space (primary monitor's top-left is (0,0); monitors to the left/above are negative). You click in the image space of the last screenshot; the server maps that back to physical pixels (handling downscale, per-monitor offset, and DPI). Every screenshot returns a capture_id; act errors loudly if you click against a stale frame instead of mis-clicking. Use system displays to see the layout, then target a specific monitor with display:0 / display:primary|left|right.

Install

As a Claude Code / Desktop plugin (recommended)

/plugin marketplace add sshh12/claude-plugins
/plugin install windows-computer-use@shrivu-plugins

The plugin bootstraps a Python virtual environment and installs this package from GitHub on first run, then starts the MCP server automatically.

Standalone (project-local MCP)

Requires Python 3.10+ and (for video) ffmpeg on PATH.

pip install git+https://github.com/sshh12/windows-computer-use-mcp

Then add to your MCP client config (e.g. a project .mcp.json):

{
  "mcpServers": {
    "windows-computer-use": {
      "command": "python",
      "args": ["-m", "windows_computer_use"]
    }
  }
}

Development

python -m venv .venv
.venv\Scripts\python.exe -m pip install -e .
.venv\Scripts\python.exe tests\smoke_engine.py    # capture/input/display engine
.venv\Scripts\python.exe tests\smoke_server.py     # assembled MCP tool surface

MCP_OUTPUT_DIR overrides where screenshots/video are written (default: a client root → ~/Pictures/windows-computer-use → %TEMP%).

Debugging

Set WCU_DEBUG_HTML_DIR to a directory and the server writes a per-session session_<stamp>.html that pretty-prints every tool call — arguments, result text, and the returned screenshots inline — so you can replay exactly what the agent saw and did:

License

MIT