Control iTerm Terminal with LLMs

iTerm MCP Server is an MCP server that gives an AI assistant access to your iTerm session, letting it read only the terminal output it's interested in - typically the last few lines, even for long-running commands - for efficient token use. You share iTerm with the model directly: you can ask questions about what's currently on screen, or delegate a task to the model and watch as it performs each step in the same terminal. It's built with minimal dependencies, runnable via npx, and designed to be easy to add to Claude Desktop and other MCP clients.

Use this when you want an AI assistant to run and monitor commands in your actual iTerm session, interact with REPLs, send control characters like ctrl-c or ctrl-z, or watch it perform a delegated multi-step terminal task while you can interrupt at any point.

Not needed if you're not on macOS with iTerm2, or for tasks where you're unwilling to let a model execute unreviewed commands in a live terminal - iterm-mcp has no built-in restrictions and makes no attempt to evaluate the safety of the commands it runs, so the user is responsible for monitoring activity and aborting when appropriate, especially for multi-step tasks that could go off track; it's a good idea to start with smaller, focused tasks until you're familiar with how the model behaves.

Three tools: write_to_terminal (writes to the active iTerm terminal, typically to run a command, returning the number of output lines produced), read_terminal_output (reads a requested number of lines from the active terminal), and send_control_character (sends a control character, such as ctrl-c, to the active terminal).

Requires iTerm2 running and Node 18 or greater. Add to Claude Desktop's config (~/Library/Application Support/Claude/claude_desktop_config.json on macOS, %APPDATA%/Claude/claude_desktop_config.json on Windows):

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

Can also be installed automatically via Smithery: npx -y @smithery/cli install iterm-mcp --client claude. For development, install dependencies with yarn install, build with yarn run build, or use yarn run watch for auto-rebuild; since MCP servers communicate over stdio, debugging is done via the MCP Inspector (yarn run inspector, yarn debug <command>), which provides a browser-based debugging URL.

Developers who want to share their iTerm session with an AI assistant to run commands, interact with REPLs, and delegate multi-step terminal tasks, while staying able to watch and interrupt if the model goes off track.