Data as of

Windows-MCP for Claude Desktop: Setup & Config

Claude Desktop: Let an AI assistant interact with Windows applications, windows, and input controls.

What you can do with Windows-MCP in Claude Desktop

Choose native Windows automation with optional DOM inspection and documented remote-access controls.

  • Automate application and window interactions on a Windows desktop.
  • Perform UI-driven QA tasks through an assistant.

Before you start: Windows-MCP in Claude Desktop

  • Windows
  • Python 3.13+
  • uv with uvx
  • An MCP-compatible client

This is the documented default stdio launch. English is preferred for App-Tool; disable that tool for other Windows languages. For network access, configure authentication and TLS rather than using an unprotected listener.

Windows-MCP Claude Desktop configuration

  1. Choose Claude Desktop configuration scope

    For Windows-MCP, open macOS: ~/Library/Application Support/Claude/claude_desktop_config.json Windows: %APPDATA%\Claude\claude_desktop_config.json. Merge under mcpServers via Settings → Developer → Edit Config when available.

  2. Add Windows-MCP to Claude Desktop

    Use the Windows-MCP entry below in mcpServers. Keep Windows-MCP's documented stdio connection fields and your other Claude Desktop entries.

  3. Check Windows-MCP in Claude Desktop

    Windows-MCP: Fully quit and reopen Claude Desktop; check developer diagnostics.

Config file · macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

Windows: %APPDATA%\Claude\claude_desktop_config.json

JSON
{
  "mcpServers": {
    "Windows-MCP": {
      "command": "uvx",
      "args": [
        "windows-mcp",
        "serve"
      ],
      "env": {}
    }
  }
}

Windows-MCP tools available to Claude Desktop

Windows-MCP's stored tool list documents 5 tools; compare it with Claude Desktop's connected tool list, because the stored list is not a live check.

Screenshot

When Claude Desktop requests Screenshot from Windows-MCP, review the action before approving it.

Capture screen imagery; this tool name appears in the README's tool-selection example.

Click

When Claude Desktop requests Click from Windows-MCP, review the action before approving it.

Perform a mouse click; this tool name appears in the README's tool-selection example.

Snapshot

When Claude Desktop requests Snapshot from Windows-MCP, review the action before approving it.

Inspect desktop state; this tool name appears in the README's tool-selection example.

PowerShell

When Claude Desktop requests PowerShell from Windows-MCP, review the action before approving it.

Registry

When Claude Desktop requests Registry from Windows-MCP, review the action before approving it.

Windows-MCP: full stored tool descriptions

Troubleshooting Windows-MCP in Claude Desktop

  • Fully quit and reopen Claude Desktop for Windows-MCP. If uvx windows-mcp (from uv, for Python tools) is missing, check the desktop process's PATH, which may differ from your terminal.
  • Windows-MCP uses stdio here: Claude Desktop must keep the configured process running; protocol output belongs on stdout and diagnostics on stderr.
  • Known limitation: Requires Windows-side execution even when the assistant runs in WSL.
  • Known limitation: Python 3.13+ is required by the documented setup.
  • Known limitation: Non-English Windows setups may need the application tool disabled.
  • General Claude Desktop troubleshooting lives in the /clients/claude-desktop guide.

Windows-MCP — details. Claude Desktop guide

Frequently asked questions

Claude Desktop loads Windows-MCP from claude_desktop_config.json: ~/Library/Application Support/Claude/ on macOS, or %APPDATA%\Claude\ on Windows. Open it via Settings → Developer → Edit Config when available.

Claude Desktop's Windows-MCP example uses stdio.

Windows-MCP's Claude Desktop configuration invokes uvx windows-mcp (from uv, for Python tools).

Windows-MCP's stored list contains 5 tools, including Screenshot, Click, Snapshot. Compare those names with Claude Desktop's connected tool list; the stored list is not a live check.

No. The README says it does not depend on traditional computer vision or a specially fine-tuned model.

The assistant can run in WSL, but the server must execute on Windows. A PowerShell bridge is documented.