dsh-computer-use-win MCP
MCP plugin for DeepSeek Harness to drive a Windows desktop: reads controls, clicks, types, takes screenshots, runs OCR.
- 522 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- The agent gets control of the mouse, keyboard, windows and screenshots on your Windows machine, including private data in open programs.
- The sign-in and open-browser modes expose your logged-in sites; actions are performed as you.
Install
- Works in
Manual install
dsh plugin --profile desktop add dsh-computer-use-winThe README's first command, for DSH Desktop. Then exit DSH from the tray and reopen it. For other MCP clients use `npx -y dsh-computer-use-win`.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- The agent gets control of the mouse, keyboard, windows and screenshots on your Windows machine, including private data in open programs.
- The sign-in and open-browser modes expose your logged-in sites; actions are performed as you.
This is third-party code. Review the repository files before installing.
What it does
dsh-computer-use-win lets a DeepSeek Harness (DSH) agent see and operate real Windows programs and web pages. The plugin finds buttons and fields, reads on-screen text, clicks, types and manages windows. In total there are 25 working tools and one management entry. Web pages are handled by default in a separate DSH browser in the background without touching your mouse. For classic programs it uses UI Automation, screenshots and OCR. When the mouse and keyboard are in use, a visible prompt appears on screen and Esc stops the actions. The plugin also runs as a plain stdio MCP server for any client via `npx -y dsh-computer-use-win`.
Who it is for. DeepSeek Harness users on Windows who want an agent to handle routine work in desktop programs that have no API.
Good fit when
- You must fill in fields in an old desktop program with no API.
- Text is typed into Notepad and then read back to confirm the write.
- The agent locates a button by on-screen text when the accessibility tree does not show it.
- One task needs a chain of 1–20 actions run through `batch`.
Not a fit when
- The task can be solved through an API, a file or the command line.
- You are on macOS or Linux; the plugin is Windows only.
- The screen shows banking or work data that must not reach a model.
Example request
Type "Meeting minutes" into Notepad and read the window text back to confirm it was written.MCP
- Transport
- stdio
- Authentication
- not required
Environment variables
`WCU_TOOL_MODE`- Toolset when run outside DSH: `lite` for the core tools, `lazy` to load them on demand.
Limitations
Windows only; DSH Web and a standalone MCP client need Node.js 22 or newer. After installing in DSH Desktop you must fully exit from the tray and reopen it, otherwise the plugin does not load. Screenshots and recognized text can hold sensitive data, and whether they reach a remote model depends on your DSH and model settings. The keep-sign-ins and already-open-browser modes give the agent access to your logged-in sites.
How to disable. Remove the plugin with `dsh plugin` for your profile or delete the `wincu` entry from your MCP client configuration. Esc stops the current run of actions.
FAQ
How do I stop the agent?
Press Esc or your own hotkey: the current run and its queue stop with no automatic retry. Before the mouse is first used, DSH's stop button works.
How does `sent` differ from `observed`?
`sent` means the input reached the program with the result unchecked. `observed` is returned when `wait_for` saw the expected window, control or text.
Can it run without DSH?
Yes, as an MCP server: `{"command": "npx", "args": ["-y", "dsh-computer-use-win"]}` with Node.js 22 or newer.
What can the agent do with this plugin?
Find windows and controls, read them, click, type, scroll, drag, take screenshots, recognize text and launch programs.
How do I reduce the number of loaded tools?
Set `WCU_TOOL_MODE=lite` for the 12 core tools or `lazy` to load them on demand.
Discussion
No questions or reviews yet.
Related
MCP serversA DeepSeek Harness plugin-runner: starts and stops the rlm-tools-bsl MCP server for 1C code analysis alongside DSH and registers it in the MCP manager itself
Clawd Cursor: desktop control through MCP
Clawd Cursor
An MCP server that lets an agent drive windows and apps: it reads the screen as structure and acts by element name
The agent moves the mouse, types, drives windows and captures PNG screenshots of a Windows desktop through AutoIt.
MCP server and wincli CLI: agents drive Windows apps by control name through UI Automation instead of screen coordinates
An MCP server for driving Windows: the agent opens apps, clicks and types, reads window and UI state, and works with files and PowerShell
computer-use-linux is an MCP server with an on-screen agent activity indicator that lets user clicks reach the application.
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| dsh-computer-use-win (this card) | no | not required | 522 | |
| computer-use-linux | yes | not required | 816 | drives a Linux desktop; dsh-computer-use-win works on Windows through UI Automation and PowerShell |
| computer-use-mcp | no | not required | 394 | offers a general set of screen actions; dsh-computer-use-win adds a web lane, Esc stop and result confirmation |
README in short
The agent works in a loop: observe, locate, act, confirm. Web and desktop programs share one toolset.