Computer Use MCP Server

io.github.domdomegg/computer-use-mcp

Lets Claude control your computer via screenshots, mouse, and keyboard automation.

What is the Computer Use MCP server?

Computer Use MCP is a Model Context Protocol server that lets Claude (or other MCP-compatible AI models) control your computer by taking screenshots and performing mouse and keyboard actions, similar to Anthropic's official computer use feature but easy to set up and run locally. It works by driving your OS input via nut.js, giving the model direct control to click, type, and navigate your desktop and applications.

Computer Use MCP gives an AI model direct control over your computer, implementing a tool nearly identical to Anthropic's official computer use guide, but packaged as a locally runnable MCP server. It captures screenshots so the model can see the screen, and issues mouse and keyboard commands (with extra nudging toward keyboard shortcuts) via nut.js. Because it grants full control of your machine to a model, it should be treated with caution - ideally run in a sandboxed user account, since models can make mistakes and are vulnerable to prompt injection.</overview> <parameter name="useCases"> <item>Automating repetitive desktop tasks (e.g., changing settings like the desktop background) via natural-language instructions to Claude</item> <item>Letting an AI agent navigate and interact with GUI applications by taking screenshots and simulating clicks/keystrokes</item> <item>Automating browser interactions more reliably using keyboard navigation (via the Rango extension) instead of coordinate-based clicking</item> <item>Prototyping and testing computer-use style AI agent workflows locally without Anthropic's hosted computer use environment</item> </useCases> <parameter name="tags">["computer-use", "automation", "mouse-keyboard", "screenshots", "nut.js", "claude", "desktop-control"]

How to install Computer Use

Copy-paste configuration for popular MCP clients.

transport: stdio

Config generated by PluginBench — verify against the source before use.

Claude Desktop

~/Library/Application Support/Claude/claude_desktop_config.json

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

Cursor

~/.cursor/mcp.json

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

Windsurf

~/.codeium/windsurf/mcp_config.json

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

VS Code

.vscode/mcp.json

{
  "servers": {
    "computer-use-mcp": {
      "type": "stdio",
      "command": "npx",
      "args": [
        "-y",
        "computer-use-mcp"
      ]
    }
  }
}

Claude Code

claude mcp add computer-use-mcp -- npx -y computer-use-mcp

Tools & capabilities

Tools this server exposes to the agent.

computer — A computer use tool (modeled on Anthropic's official computer use tool) that takes screenshots of the screen and performs mouse and keyboard actions to control the computer, with a preference for keyboard shortcuts where possible.

Computer Use MCP server FAQ

What does the Computer Use MCP server do?

It gives an AI model like Claude control of your computer by taking screenshots and performing mouse and keyboard actions, similar to Anthropic's official computer use feature but set up to run locally.

Is Computer Use MCP free to use?

Yes, it's an open-source npm package (computer-use-mcp) that you can install and run for free.

How do I install it in Claude Desktop, Cursor, or other MCP clients?

Use the install-mcp tool linked in the README to generate the correct configuration, or run it via npm (npx -y computer-use-mcp) or the provided .mcpb bundle for your MCP client.

What authentication does it require?

None described - it runs locally and directly controls your computer's input devices via nut.js, without requiring API keys or accounts.

Is it safe to give an AI model full control of my computer?

The README explicitly warns that models make frequent mistakes and are vulnerable to prompt injection, so this should be used with close supervision, ideally in a sandboxed user account.

How can I improve reliability when using this tool?

Use a strong computer-use-capable model like the latest Claude models, run at a small common resolution like 720p, and install the Rango browser extension for keyboard-based navigation instead of coordinate clicking.

README (reference)

Source of truth, from the repository.

computer-use-mcp

💻 A Model Context Protocol server for Claude to control your computer. This is very similar to computer use, but easy to set up and use locally.

Here's Claude Haiku 4.5 changing my desktop background (4x speed):

https://github.com/user-attachments/assets/cd0bc190-52c4-49db-b3bc-4b8a74544789

[!WARNING] At time of writing, models make frequent mistakes and are vulnerable to prompt injections. As this MCP server gives the model complete control of your computer, this could do a lot of damage. You should therefore treat this like giving a hyperactive toddler access to your computer - you probably want to supervise it closely, and consider only doing this in a sandboxed user account.

Installation

Follow the instructions on install-mcp, which generates the right config for your MCP client (Claude Code, Claude Desktop, Cursor, Cline, VS Code, and more).

Tips

This should just work out of the box.

However, to get best results:

Use a model good at computer use - I recommend the latest Claude models.
Use a small, common resolution - 720p works particularly well. On macOS, you can use displayoverride-mac to do this. If you can't use a different resolution, try zooming in to active windows.
Install and enable the Rango browser extension. This enables keyboard navigation for websites, which is far more reliable than Claude trying to click coordinates. You can bump up the font size setting in Rango to make the hints more visible.