Vision Input

Hey there! 👋

This is a Model Context Protocol (MCP) server that basically gives your AI superpowers on your computer. It lets the AI:

1. See what you see (by taking screenshots).
2. Use your mouse (move, click, drag, scroll).
3. Type for you (keyboard input).

Think of it as giving your AI agent a pair of eyes and hands.

getting started

You'll need Node.js v16+.

Platform Specifics

macOS: Since this tool controls your mouse and sees your screen, macOS will ask for permission. You'll need to allow:

• Screen Recording
• Accessibility

Windows:

• You may need to install build tools for robotjs: npm install --global --production windows-build-tools

Linux:

• You will likely need X11 libraries. For Debian/Ubuntu: sudo apt-get install libxtst-dev libpng++-dev

installation

It's pretty standard stuff:

1. Clone this repo.
2. Install the dependencies:

   npm install
   

3. Build it:

   npm run build
   

setup

Just add this to your MCP config (like claude_desktop_config.json or whatever you're using):

{
  "mcpServers": {
    "vision-input": {
      "command": "node",
      "args": [
        "/ABSOLUTE/PATH/TO/THIS/REPO/build/index.js"
      ]
    }
  }
}

(Make sure to change that path to where you actually put this folder!)

capabilities

Here's what it can do out of the box:

• screen_capture: Snaps a picture of your screen so the AI knows what's going on.
• calculate_screen_dimensions: Checks how big your monitor is.
• mouse_move / mouse_click / mouse_drag: Does exactly what it says on the tin.
• mouse_scroll: Scrolls up or down.
• keyboard_type: Types text out.
• keyboard_press: Hits specific keys (like cmd+c to copy).

trouble?

• Grey/Black Screen? macOS is probably blocking the screen recording. Check System Settings > Privacy & Security > Screen Recording.
• Can't Click? Check System Settings > Privacy & Security > Accessibility.

Enjoy building cool stuff! 🚀