Windows MCP is an open source server that gives Claude hands on your Windows PC: it can open apps, click, type, read the screen and run PowerShell. This article explains how it works, how to install it in Claude Desktop, and how to keep it safe.
You asked Claude to sort 300 files in your Downloads folder, and it gave you a neat plan that you still had to carry out by hand. Windows MCP removes that last step. It is an open source server that gives Claude a pair of hands on your Windows PC: it can read what is on the screen, click buttons, type into fields, open apps, move files and run PowerShell commands, all from a normal chat in Claude Desktop. This article explains what Windows MCP is, how the pieces fit together, how to install it, which jobs it handles well and which settings stop it from causing damage. It also clears up a naming mix-up, because Microsoft uses the phrase MCP on Windows for a separate feature built into the operating system itself.
The Model Context Protocol (MCP) is an open standard that Anthropic introduced in late 2024. It lets an AI app talk to outside tools through one shared format. Claude Desktop plays the client. Anything that speaks the protocol plays the server: a database connector, a calendar, a browser or, in this case, your whole desktop. The client asks the server which tools it offers, Claude picks the one that fits your request, and the server runs it and returns the result.
That split matters. Claude never touches your PC directly. It can only ask for actions the server has chosen to expose, and every one of them is a named tool you can inspect, limit or switch off.
Windows-MCP is an MIT licensed project by CursorTouch, published on GitHub. Its README describes it as a lightweight bridge between language models and the Windows operating system, built for file work, app control and UI interaction without needing a specialized vision model. It supports Windows 7, 8, 8.1, 10 and 11, and it is featured as a Desktop Extension in Claude Desktop.
💡 Two things share one name. Windows-MCP is a community server you install yourself. MCP on Windows is Microsoft's own framework, described in a later section. Most of this article is about the first one.
Claude does not watch your monitor. The server describes it. The Snapshot tool returns the full UI state with an ID for each element, so Claude can act on "element 14" instead of guessing at pixel positions. The Screenshot tool is the lighter option: a fast visual capture that includes the cursor and open windows. Based on what comes back, Claude picks the next tool, such as Click, Type, Scroll or Shortcut, and repeats the loop until the task is done. The project quotes a typical delay of 0.2 to 0.5 seconds between actions.
| Mode | What Claude receives | Best for |
|---|---|---|
| Screenshot (default) | A fast image of the screen with cursor and windows | Quick visual checks, lower token cost |
| Snapshot | The UI state with element IDs, plus a screenshot when use_vision=True | Precise clicks inside crowded apps |
Both modes support region capture and multi display selection, which keeps token use down. If screenshots are still too heavy, set WINDOWS_MCP_SCREENSHOT_SCALE to a value between 0.1 and 1.0 to shrink them.
The README lists about twenty tools. Grouped by job:
| Group | Tools | What they do |
|---|---|---|
| Mouse and typing | Click, Type, Scroll, Move, Shortcut, MultiSelect, MultiEdit | Click coordinates, enter text, scroll, drag, send shortcut combinations such as Ctrl+C, select or fill many items at once |
| Seeing the screen | Screenshot, Snapshot, DisplayInventory, ControlStatus | Capture the UI, read display layout and DPI, check who holds desktop control and how long it has been idle |
| Apps and system | App, PowerShell, Process, Registry, Notification, Clipboard | Launch apps and manage windows, run commands, list or end processes, edit the Registry, send toasts, read and write the clipboard |
| Files and web | FileSystem, Scrape | Read, write, copy, move, delete, list and search files; extract webpage content |
| Timing | Wait, WaitFor | Pause, or poll for text, windows, elements or focus |
pip install uv works)claude_desktop_config.json, normally inside %APPDATA%\Claude.windows-mcp entry inside your existing mcpServers section.{
"mcpServers": {
"windows-mcp": {
"command": "uvx",
"args": ["windows-mcp", "serve"]
}
}
}
💡 Microsoft Store version of Claude Desktop? It keeps its config in a different folder, and the project README says to use the full path to the executable instead, for example
C:\Users\<user>\.local\bin\uvx.exe. The packaged app may not see your normal PATH, so a bareuvxfails.
Prefer an extension? Windows-MCP also ships as a Desktop Extension. The project's steps are: install Anthropic's mcpb tool with npm, run npx @anthropic-ai/mcpb pack in the cloned project to build a .mcpb file, then install that file from the Extensions page in Claude Desktop settings.
Open a new chat and confirm the server shows as connected. Then start with something harmless:
"Take a screenshot and tell me which apps are open."
Once that works, raise the stakes a little:
"Open Notepad and type a three line shopping list. Do not save it."
Watch the screen while it runs. You will see the pointer move, windows open and text appear, which is the quickest way to build a feel for how fast and how literal Claude is when it drives a real desktop.
💡 You can always take control back. Move the mouse by a large amount, or press Ctrl+Alt+Shift+Backspace, and the desktop returns to you.
The FileSystem tool can list, search, copy, move and delete, so a messy Downloads folder is the classic first job. Ask for the plan before the action: "List what is in Downloads, propose sub folders by file type, and wait for my approval before moving anything." That single sentence turns a risky bulk operation into a reviewable one.
Plenty of desktop software has no API. A purchasing tool from 2012 or an internal form with 40 fields still needs a person to type. MultiEdit fills several input fields in one go, Clipboard moves values between windows, and Snapshot gives Claude reliable targets. Treat the first few runs as supervised: watch the screen and check the totals yourself.
The project lists QA testing among its uses. Click through an installer, flip a Windows setting, confirm an app opens, then ask Claude to describe what changed. This job fits a virtual machine best, where a mistake costs you a snapshot restore instead of a bad afternoon.
Here are four starter prompts and the tools they are likely to trigger:
| Task | Starter prompt | Tools involved |
|---|---|---|
| Tidy a folder | "Propose a folder structure for Downloads and wait for my approval." | FileSystem, Snapshot |
| Fill a form | "Copy the five rows from this spreadsheet into the form." | Snapshot, MultiEdit, Clipboard |
| Check an app | "Open the app, go to Settings and tell me the version." | App, Click, Screenshot |
| Wait on a job | "Wait until the export window says finished, then notify me." | WaitFor, Notification |
💡 Ask for a pause before anything destructive. A line such as "stop and ask me before deleting, overwriting or closing anything" costs nothing and catches most surprises.
The project's own security policy is blunt. The server is not sandboxed, it runs with all the permissions of the Windows user that launched it, and many actions cannot be undone. Files removed through PowerShell or the interface may be gone for good, and screenshots can capture sensitive information that then travels to the model.
The README says you can disable risky tools with the --exclude-tools option or the WINDOWS_MCP_EXCLUDE_TOOLS environment variable, so check it for the exact syntax. For a first week, leaving out Registry and PowerShell removes the two most powerful ways to break something. Telemetry is another setting worth knowing: it is on by default and anonymous according to the project, and you can switch it off by adding "env": {"ANONYMIZED_TELEMETRY": "false"} to the server entry.
If you ever expose the server over a network instead of the local Claude Desktop link, the project asks for more protection: an authentication token, an IP allowlist and TLS.
The security policy recommends a handful of habits:
It also advises against using Windows MCP on production servers, in compliance regulated environments, on machines with irreplaceable data or on shared multi user computers.
💡 Rule of thumb: if you would not leave a colleague alone with your open laptop for ten minutes, do not give Claude that same access on a machine that holds unsaved or irreplaceable work.
Microsoft announced a public preview of native MCP support in Windows in May 2025. As described at the time, an on-device registry lists agent connectors, which are MCP servers published by app developers. Built-in connectors for the file manager and System Settings were part of the preview. Agent access is switched off by default, the registry only admits servers that meet baseline criteria such as code signing and fixed tool definitions, a system proxy applies policy and keeps an audit trail, and users get an approval prompt, similar to a UAC dialog, before sensitive tool calls.
| Feature | Windows-MCP (community) | MCP on Windows (Microsoft) |
|---|---|---|
| Built by | CursorTouch, open source, MIT license | Microsoft, part of Windows |
| Install | Config file or Desktop Extension | On-device registry managed by Windows |
| Reach | The whole desktop through UI automation, PowerShell, files and Registry | Only the connectors that apps register |
| Safety model | Your user permissions, no sandbox | Off by default, signed servers, proxy and approval prompts |
Both are moving targets, so check the project README and Microsoft's current documentation before you build a workflow on either. For now, the community server is the one that lets Claude drive any app on screen, while Microsoft's route is narrower and more tightly controlled.
The README is honest about the limits:
There is also a structural cost. Every step is a round trip through the model, so a long job runs slower, and uses more tokens, than a script would. When a task repeats every day, ask Claude to write a PowerShell script once and run that script from then on.
| Symptom | Likely cause | Fix |
|---|---|---|
| Server never shows as connected | uvx not found, Python older than 3.13, or a Store build without full paths | Install uv, update Python, use the full path to uvx.exe |
| Tools missing after restart | Typo in the config JSON, or Claude Desktop was not fully quit | Validate the JSON, quit from the tray, reopen |
| First start seems frozen | Dependencies are still installing | Wait one to two minutes |
| Clicks land in the wrong place | Display scaling or several monitors | Run DisplayInventory and prefer Snapshot element IDs over raw coordinates |
| App tool cannot open a program | Non English Windows language | Switch the display language, or open the program by path through PowerShell |
Windows MCP gives Claude a pair of hands. Picasso IA gives you a studio. Once Claude can drive your PC, the slow part of many creative jobs is no longer the clicking, it is producing the pictures and clips themselves.
Here is where to start:
PicassoIA also offers an MCP connection of its own, so Claude can request images and videos from inside a conversation. With both servers connected, one chat can ask for an image and then use Windows MCP to save or place it where you need it. Check the PicassoIA site for how to connect your account and what your plan includes.
💡 Your next step: open Picasso IA, pick one model, write a prompt about a scene from your own desk, and generate your first image today. Then browse every model at picassoia.com/en/all-models and try a short video from that same image.
