Windows
by CursorTouch
Windows MCP server that lets an agent operate a live Windows desktop — clicking, typing, launching apps, editing the registry, and reading the screen through an accessibility tree instead of pixels.
About
Windows-MCP hands an AI agent a real desktop session instead of a sandboxed API. It exposes low-level input control — click, type, scroll, drag, keyboard shortcuts — alongside higher-level state tools that return the active windows, the interactive UI element tree, and display/DPI metadata, so a model can decide what to do next without relying on screenshot guessing or a fine-tuned vision model.
Beyond UI automation, the server reaches into the OS itself: launching and switching applications, running PowerShell commands, reading and writing the clipboard, listing or killing processes, editing the Windows Registry, and doing full filesystem read/write/search operations. That combination makes it useful for QA testing existing Windows apps and for general desktop-automation tasks an agent would otherwise need remote-desktop screen scraping to attempt.
Key features
- Accessibility-tree snapshots with interactive element IDs instead of raw pixel coordinates
- Direct input control: click, type, scroll, drag, and keyboard shortcuts
- App lifecycle control — launch, resize, move, switch, and list running processes
- PowerShell execution and full Registry read/write/delete access
- Filesystem operations (read, write, copy, move, delete, search) and clipboard access
- Windows toast notifications and webpage content scraping
Use cases
- QA testing a Windows desktop application end to end without a human at the keyboard
- Letting an agent perform routine desktop chores (file organization, app switching, form filling) across multiple applications
- Driving legacy Windows software that has no API by controlling its UI directly
- Registry and process inspection/cleanup tasks delegated to an agent
Available tools
Snapshot
Captures the full UI state — window list, interactive element IDs, and accessibility tree — for the model to reason over.
Click
Clicks at specified screen coordinates or on a resolved UI element.
PowerShell
Executes an arbitrary PowerShell command and returns its output.
FileSystem
Reads, writes, copies, moves, deletes, lists, or searches files and folders.
Registry
Reads, writes, deletes, or lists Windows Registry keys and values.
App
Launches, resizes, moves, or switches focus to a given application.
Frequently asked questions
Does it need a vision model to see the screen?
No — it reads the live accessibility tree and window state directly, falling back to screenshots only when an app exposes no accessible UI elements.
What can it touch on the machine?
Everything the process running it can touch: UI input, the filesystem, the clipboard, running processes, PowerShell, and the Registry, so it should run under an account scoped to what you actually want automated.