Skip to content
Get packs

Windows-MCP

An MCP server that lets AI agents control a Windows PC: open apps, click, type, and read window state.

GitHub Download v0.8.8
Stars
11k
Forks
1.2k
Licence
MIT
Language
Python
Last update
4 days ago
Type
AI model / agent

What it is

Windows-MCP is an open-source MCP server that connects AI agents and language models to the Windows operating system. Once connected, an agent can navigate files, open and control applications, interact with UI elements, and run QA-style tests. It is designed to work with any LLM and does not rely on traditional computer vision or specific fine-tuned models, according to the README.

  • Have an AI assistant open a desktop application and perform a sequence of clicks and keystrokes
  • Capture the state of open windows and UI elements so an agent can decide its next step
  • Use DOM mode to automate web page content in Chrome, Edge, or Firefox
  • Run the server as a background task that starts at login

Great for

  • Developers experimenting with AI agents that need to operate Windows apps
  • Users of MCP-compatible clients such as Claude Desktop who want their assistant to control local Windows software
  • QA testers looking to automate UI interactions on Windows 10 or 11
  • Builders who want a readable, MIT-licensed base to extend with custom tools

Not ideal for

  • Anyone on Windows 8 or earlier, since the README says only Windows 10 and 11 are supported
  • Users on Python versions below 3.14, because the current package requires Python 3.14+
  • People who want unsupervised control of a machine holding sensitive data, since an agent driving the desktop can take real actions
Step-by-step guide: how to use Windows-MCPInstall, first run, examples, tips and FAQs

How to get it

  1. 1

    Install with a package manager

    You'll need: Python 3.14+ · uv (Python package manager)

    Install uv

    pip install uv

    Run the server directly

    uvx windows-mcp serve

    Install as a background task

    windows-mcp install
  2. 2

    Run it

    • Install Python 3.14+ and uv.
    • Run uvx windows-mcp serve to start the server (first run may take a minute or two and may time out; restart if so).
    • Add the server to your MCP client config, for example the Claude Desktop JSON shown in the README.
    • Fully restart the MCP client and confirm windows-mcp appears in its tools list.
  3. 3

    Get the most out of it

    • Use the --transport sse or --transport streamable-http options if your client connects over HTTP instead of stdio
    • Use windows-mcp install to run it as a per-user Scheduled Task that starts at every login
    • Enable use_dom=True in the State-Tool when automating web pages for cleaner output
    • Check ~/.windows-mcp/server.log and server.error.log when troubleshooting

Downloads

Source code (.zip)main branch · straight from GitHubRelease v0.8.84 files incl. windows_mcp-0.8.8-py3-none-any.whl.publish.attestation

Strengths

  • Works with any LLM, and the README says no fine-tuned models or traditional computer vision are required
  • Simple setup via uvx windows-mcp serve, with a PyPI package and an MIT license
  • Supports stdio, SSE, and streamable HTTP transports

Trade-offs

  • The README states Windows 10 and 11 only and Python 3.14+, which limits compatibility
  • Latency of 0.2 to 0.5 seconds per action, which varies with system load and LLM speed
  • Granting an agent desktop control has real security implications the README does not deeply address

How it compares

ProjectStarsTypeLevelUpdatedBest for
Windows-MCP (this)11kAI model / agentintermediate4 days agoDevelopers experimenting with AI agents that need to operate Windows apps
PowerToys140kAppbeginner4 days agoWindows users who want more control over window layout and multi-monitor setups
DevToys32.2kAppbeginner1 months agoDevelopers who often decode, encode, or format small pieces of data
popcorntime10.7kAppintermediate2 months agoPeople who want one app to browse movies and TV shows across platforms
invisible_dots32.4kAI model / agentintermediate6 days agoDevelopers who want a self-hosted agent and are comfortable with a command-line setup
UFO10.2kAI model / agentadvanced6 days agoDevelopers experimenting with LLM-driven Windows desktop automation

Stats synced from GitHub. The assessment above was drafted with AI from the project's README on 2026-10-29 and may miss recent changes — check the README before relying on a command. MIT licence: free to use, modify and distribute, including commercially, as long as you keep the copyright notice.