Skip to content

Windows automation that actually works

Let your AI assistant control any Windows app, browser or website by name — not pixels. Works with GitHub Copilot, Claude Desktop, Cursor and any MCP client.

Install for VS Code Other MCP clients

VS Code Marketplace installs GitHub stars GitHub downloads

Control Windows apps by name, not pixels

Windows MCP Server gives your AI assistant direct access to Windows applications through the Windows UI Automation API — the same API screen readers use to read buttons, menus and text fields.

Your AI says "click Save" and the server finds the Save button by name. No screenshots. No pixel parsing. No coordinate guessing. It's deterministic — the same command works every time, regardless of DPI, theme or resolution.

Also automating Excel?

Check out Excel MCP Server — the sister project, built the same way.

Also building PowerPoint decks?

Check out PowerPoint MCP Server — another sister project, built the same way.

Key features

  • Semantic UI access

    Find elements by name, type or ID — not coordinates. Works regardless of DPI, theme, resolution or window position.

  • Browser automation

    Automate Edge and Chrome page content by name — links, buttons, forms and ARIA labels. Works with signed-in sessions via the dedicated windows_mcp_browser_automation prompt.

  • Multi-monitor

    Full support for multiple displays with per-monitor DPI scaling. Move windows between monitors and capture any screen.

  • Broad app support

    Tested against classic Windows apps, modern Windows 11 apps, Electron apps (VS Code, Teams, Slack) and Chromium browsers. Same commands everywhere.

  • Full fallback

    Screenshot + mouse + keyboard for games and custom controls. Annotated screenshots return element metadata — image omitted by default to save tokens.

  • LLM-tested quality

    Every tool tested with real AI models before release. 54 automated tests across 7 scenarios. 100% pass rate required.

See all tools and operations

See it in action

Ask your AI assistant in plain language — it drives Windows for you:

🖱️ Control any app

You: "Click the Save button in Notepad, then type my email in the login field and toggle Dark Mode in Settings."

The AI finds each control by name through the UI Automation API and acts on it — no screenshots or coordinates required.

🌐 Automate the browser

You: "Open my banking site in Edge and click Transfer, then fill out the form on this page in Chrome."

The AI drives page content by name — links, buttons, forms and ARIA labels — working with your signed-in sessions.

🪟 Manage windows

You: "Move this window to my second monitor, then read the error message from that dialog."

The AI finds, activates, moves and resizes windows and reads on-screen text (with OCR fallback).

Why this approach

Most automation tools take screenshots and ask vision models to find buttons in the pixels. That approach is slow, expensive and breaks when windows move or themes change.

Windows MCP Server asks Windows directly: "What buttons exist?" Windows knows. It's deterministic — the same command works every time.

For repeated observations, automatic snapshots compare a simplified semantic view and return a compact diff only when it is safely smaller. A five-run benchmark measured 84-96% median byte/token savings for Electron, Word, and Excel changes. The same Playwright-style view reduced realistic Chrome navigation by 13.1% in bytes and 13.4% in approximate tokens, even though 18 of 20 responses were complete simplified views. A later strict Chrome run measured the conservative display cleanup on the same captures: another 10.6% fewer bytes and 13.6% fewer tokens.

Read the reproducible benchmark

Tested with real AI models

Tool descriptions that seem clear to humans often confuse AI. Parameters get misunderstood. Actions get skipped.

We test every tool with a real AI model (GPT-5.5 via GitHub Copilot) using pytest-skill-engineering. 130+ automated tests. The tests are intentionally manual-only and isolated from PR, CI, and release workflows. If the AI can't use a tool correctly, we fix the tool — not the prompt.

Run the LLM Integration Tests workflow

Quick start

Install directly from the VS Code Marketplace:

Install from VS Code Marketplace

Download from GitHub Releases and add to your MCP config:

{ "servers": { "windows": { "command": "path/to/Sbroenne.WindowsMcp.exe" } } }

Caution

This MCP server controls your Windows desktop. Use responsibly.

GitHub star history

GitHub stars over time for mcp-windows

Updated daily from GitHub's exact public star count. The historical seed and persisted snapshots contain aggregate dates and counts only.