๐ช Windows MCP Server
August 28, 2026 ยท View on GitHub
Windows automation that actually works. Uses the Windows UI Automation API to find buttons by name, not pixels. Tested with real AI models through a dedicated manual workflow.
Why This Exists
Screenshot-based automation doesn't work reliably. Vision models guess wrong, coordinates break when windows move or DPI changes, and you burn through thousands of tokens on retry loops. We tried it (check the commit history) โ it failed too often to be useful.
Windows MCP Server asks Windows directly: "What buttons exist in this window?" Windows knows. It's deterministic.
How It Works
# 1. Find the window
window_management(action='find', title='Notepad') โ handle='123456'
# 2. Click elements by name
ui_click(windowHandle='123456', nameContains='Save')
# 3. Type into fields
ui_type(windowHandle='123456', controlType='Edit', text='Hello World')
# 4. Fallback for games/canvas โ screenshot + mouse
screenshot_control(target='window', windowHandle='123456') โ element coordinates
mouse_control(action='click', x=450, y=300)
Same command works every time. Any machine. Any DPI. Any theme.
Browsers follow the same semantic flow: launch msedge.exe or chrome.exe, then use ui_find, ui_click, and ui_type on links, buttons, and fields exposed through UIA names and ARIA labels.
Key Features
- ๐ง Semantic UI โ Find elements by name, not coordinates. Works regardless of DPI, theme, or window position.
- ๏ฟฝ Multi-Monitor โ Full support for multiple displays with per-monitor DPI scaling.
- ๐งช LLM-Tested โ 130+ tests with a real AI model (GPT-5.5 via GitHub Copilot), run intentionally through a dedicated manual workflow.
- ๐ป Broad App Support โ Tested against classic Windows apps, modern Windows 11 apps, and Electron apps (VS Code, Teams, Slack). Chromium browser pages follow the same ARIA-driven pattern, but browser chrome remains best-effort.
- ๐ Full Fallback โ Screenshot + mouse + keyboard for games and custom controls.
- ๐ช Token Optimized โ Short property names, JPEG screenshots, and auto-scaling substantially reduce token usage compared to standard JSON.
Installation
VS Code Extension โ Install from Marketplace. Works with GitHub Copilot automatically.
Plugin (GitHub Copilot CLI / Claude Code) โ Install the shared plugin bundle in plugin/:
copilot plugin install sbroenne/mcp-windows:plugin
For local Claude Code development:
claude --plugin-dir .\plugin
On first use, the plugin downloads the current standalone release into plugin\bin\.
Standalone โ Download from Releases. Add to your MCP config:
{ "servers": { "windows": { "command": "path\\to\\Sbroenne.WindowsMcp.exe" } } }
Limiting the exposed tools (optional) โ for a least-privilege setup, pass an allowlist or denylist on the server command line (or via environment variables). CLI flags take precedence over the env vars.
{ "servers": { "windows": {
"command": "path\\to\\Sbroenne.WindowsMcp.exe",
"args": ["--tools", "ui_snapshot,ui_find,ui_click,ui_type,ui_read"]
} } }
--tools <a,b,c>/WINDOWS_MCP_TOOLSโ expose only these tools.--exclude-tools <x,y>/WINDOWS_MCP_EXCLUDE_TOOLSโ expose everything except these.
Excluded tools never appear in tools/list and cannot be invoked.
Tools
| Tool | Purpose |
|---|---|
ui_snapshot | Capture a compact element tree; mode=auto returns smaller updates after the first view |
ui_find | Discover elements in a window (with timeout/retry) |
ui_click | Click buttons, checkboxes, menu items by name |
ui_type | Type into text fields |
ui_select | Pick a value in a combo box, list, or tab |
ui_read | Read text from elements (with OCR fallback) |
ui_read_table | Extract a grid/table/list-view into structured rows + headers |
ui_wait | Wait for an element to appear, disappear, or reach a state |
ui_batch | Run several UI steps (find/click/type/select/wait/read/snapshot/key/mouse/polyline) in one call |
ui_macro | Record & replay a ui_batch sequence by name (save/run/list/get/delete) |
file_save | Save files via Save As dialog |
file_open | Open an existing file via the Open dialog |
clipboard | Read/write the Windows text clipboard (get/set/clear) |
process | List or kill running processes, task-manager style (list/kill) |
screenshot_control | Get element metadata (image optional) |
window_management | Find, activate, move, resize windows |
mouse_control | Coordinate-based clicks (fallback for games) |
keyboard_control | Hotkeys and key sequences |
app | Launch applications |
Full reference: FEATURES.md
Use the default mode=full for one inspection. For repeated views of the same window or a known
subtree through the long-running MCP server, use mode=auto from the first view; full is not
remembered.
The first response has kind=full; later responses have kind=diff when a short change list saves
space, otherwise they safely fall back to kind=full. Use mode=reset to start a new comparison,
and use parentElementId only to revisit a subtree returned by an earlier snapshot or find. Separate
wincli commands do not share remembered views.
The reproducible benchmark measured 84-96% median
byte/token savings in Electron, Word, and Excel. A Playwright-style semantic view improved realistic
Chrome navigation to 13.1% fewer bytes and 13.4% fewer approximate tokens even though 18 of 20
responses were complete simplified views. A later strict Chrome run found that conservative display
cleanup alone removed another 10.6% of bytes and 13.6% of tokens from automatic responses.
Snapshot response compatibility: complete snapshots still return the compact tree, but no
longer repeat the same hierarchy in the former full-detail elements field. Consumers that read
that undocumented duplicate should migrate to tree (or use ui_find for a flat result).
Command-line interface (wincli)
The server ships with a twin command-line entry point, wincli, in
src/Sbroenne.WindowsMcp.Cli. It exposes the exact same
capabilities as the MCP server โ every command calls the same underlying tool, so the JSON output is
byte-for-byte identical to the MCP tools (verified by a parity test).
wincli is the token-efficient path for coding agents: instead of loading every MCP tool schema
into context, an agent with shell access discovers the whole surface through --help / tools /
guidance and issues one command per action. Because window handles are OS-global, each call is
stateless โ there is no server session to keep alive.
wincli window find --title Notepad # -> window handle
wincli ui snapshot --window 12345 --mode full # complete accessible element tree
wincli ui click --window 12345 --name Submit --with-snapshot
wincli clipboard set --text "hello" # write the clipboard
wincli macro run --name login --window 12345 # replay a saved workflow
wincli guidance # full automation guide
Exit codes: 0 success, 1 tool error, 2 usage error. See the
CLI README for the full command reference.
โ ๏ธ Caution
This MCP server controls your Windows desktop. Use responsibly.
Known Limitations
UAC & Elevated Processes โ Windows security prevents any non-elevated process from interacting with UAC prompts or elevated (Administrator) windows. This is a fundamental Windows security boundary, not an MCP limitation.
| Scenario | What Happens | Workaround |
|---|---|---|
winget install triggers UAC | AI cannot click the UAC prompt | Run terminal as Administrator first |
| App running as Administrator | UI automation tools return ElevatedWindowActive error | Run MCP server elevated, or use the app non-elevated |
| UAC prompt appears | AI cannot interact with secure desktop | User must manually approve |
See FEATURES.md for details.
Testing
dotnet test # All tests
dotnet test --filter "FullyQualifiedName~Unit" # Unit only
Framework coverage: Tests run against WinForms, WinUI 3, Electron apps, and real Chromium browser app windows by default. The Chromium smoke stack now exercises both Edge and Chrome (when installed) across a deterministic local page and a required public-web slice (demo.playwright.dev/todomvc) using the same semantic-first UI Automation model with isolated browser state and browser-window-only cleanup.
dotnet test .\tests\Sbroenne.WindowsMcp.Tests\Sbroenne.WindowsMcp.Tests.csproj --filter "FullyQualifiedName~ChromiumBrowser"
LLM tests: 130+ tests with a real AI model (GPT-5.5 via GitHub Copilot). They are intentionally manual-only and never run as part of PR, CI, or release workflows. Run them from the dedicated LLM Integration Tests workflow in GitHub Actions.
cd tests/Sbroenne.WindowsMcp.LLM.Tests
uv run pytest -v
Requires GitHub authentication (GITHUB_TOKEN or an existing gh login) and a Windows desktop session. See LLM Tests README.
Related Projects
- pytest-skill-engineering โ LLM agent testing framework (powers our integration tests)
- Excel MCP Server โ AI-powered Excel automation
- OBS Studio MCP Server โ AI-powered streaming control
Documentation
| Document | Description |
|---|---|
| FEATURES.md | Complete tool reference โ all actions, parameters, examples |
| Incremental snapshot benchmark | Five-run Electron, Chrome, Word, and Excel measurements |
| CONTRIBUTING.md | Build instructions, coding guidelines, PR process |
| LLM Tests README | How to run LLM integration tests |
| Release Setup | Azure OIDC and GitHub Actions configuration |
License
MIT โ see LICENSE
Contributing
See CONTRIBUTING.md