# Computer Use [Health: Active]

**Category:** 🖥️ OS Automation  
**Repository:** https://github.com/domdomegg/computer-use-mcp.git  
**GitHub Stars:** 386  
**npm Downloads (last month):** 2821  
**Views:** 2  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/computer-use

## Description
Control your computer with screenshots, mouse, and keyboard automation.

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `npx` (confidence: high):

```json
"mcpServers": {
  "computer-use": {
    "command": "npx",
    "args": ["-y","computer-use-mcp"]
  }
}
```

## Documentation

## What Computer Use MCP server does

Computer Use exposes a local desktop to an AI model through an MCP interface. The model can inspect the screen with screenshots and interact with visible applications using mouse and keyboard automation. The project is designed to resemble Anthropic’s computer-use tool while being available as a locally installed MCP server.

This makes it suitable for tasks that require interacting with graphical applications rather than calling a structured API. Examples include changing desktop settings, navigating an application, or operating websites through visible controls. The model receives the current screen state and can then choose interaction steps based on what it sees.

The Computer Use MCP server grants the model broad control over the connected computer. That makes supervision important: the README warns that models can make mistakes and may be vulnerable to prompt injection. A separate sandboxed user account is worth considering for testing.

## How it works

The server implements a computer-use-style tool based closely on Anthropic’s documented approach, with additional guidance favoring keyboard shortcuts. It connects to the host computer through nut.js, which handles the underlying desktop interaction. The available interaction model is therefore based on screenshots, pointer actions, and keyboard input rather than application-specific APIs.

Results depend on the model’s ability to interpret screen content and choose reliable actions. The project recommends using a model that performs well at computer-use tasks. A 720p display resolution is suggested because smaller, common resolutions can make visual interaction easier. On macOS, the related displayoverride-mac project can be used to change the display resolution; if that is not possible, zooming the active window may help.

For browser tasks, the README recommends the Rango browser extension. Its keyboard-navigation hints can be more dependable than asking the model to click exact screen coordinates, and its font-size setting can be increased to make those hints easier to see.

## Setup and configuration

The README directs users to the install-mcp setup page. That flow generates the MCP client configuration needed to run the package with npx, using the computer-use-mcp package. It lists Claude Code, Claude Desktop, Cursor, Cline, and VS Code among the clients for which configuration can be generated.

The project describes the default setup as requiring no additional configuration. For development, contributors need Git and Node.js, then install dependencies with npm install, run tests with npm run test, and build with npm run build. These development steps are separate from the normal client installation path.

## Limitations and notes

Treat the Computer Use MCP server as a high-risk automation tool. Because the model can control the computer, an incorrect action or a prompt injection could affect files, settings, applications, or other data available to the account. Closely supervise sessions and avoid giving it unrestricted access to an important workstation.

It also does not claim that every model or display configuration will work equally well. Resolution, browser navigation, and model choice can all affect reliability.

_Full upstream README: https://allmcps.com/mcp/computer-use/readme_