Runbook AI MCP Server

An MCP (Model Context Protocol) server that provides browser automation capabilities through a Chrome extension. It allows terminal-based agents like Claude Code to interact with any website through your live browser session.

Part of the Runbook AI ecosystem. Join the Discord community to provide your feedback and get involved in the development!

https://github.com/user-attachments/assets/a43fba64-bc40-4ef6-9840-e100203e2cf5

Why Runbook AI?

Most browser-based MCP tools (like chrome-devtools-mcp) blow up your LLM context window by sending the entire DOM after every browser action.

Runbook AI is different:

• Optimized Context: It generates a highly simplified version of the HTML. It strips the junk but keeps essential text and interaction elements. It’s condensed, fast, and won’t eat your tokens.
• The Ultimate Catch-all: If a site doesn't have a dedicated MCP server (like Expedia, LinkedIn, or internal tools), this fills the gap perfectly.
• Privacy First: It runs entirely in your browser. No remote calls except to your chosen LLM provider. No eval() or shady scripts (enforced by the Chrome extension sandbox).
• Efficient Navigation: The simplified HTML goes beyond the viewport, making scrolling and multi-page tasks much more efficient.

Installation

MCP Server

Add to your MCP settings configuration:

{
  "mcpServers": {
    "runbook-ai": {
      "command": "npx",
      "args": ["-y", "runbook-ai-mcp@latest"]
    }
  }
}

Chrome Extension

Install the Runbook AI extension from Chrome Web Store.

Enable MCP in the extension settings opened from extension side panel.

Set LLM API key, and model name, base URL. Use of Gemini 3 Flash (gemini-3-flash-preview) is recommended. Get your free API key from Google AI Studio.

MCP token (recommended)

Without a token, any program that can open a WebSocket to localhost:9003 can drive your browser. To lock it down, open the extension settings, click Generate next to MCP token, and save. From then on the extension runs only requests that carry the same token.

Pass the token to the MCP server with --token:

{
  "mcpServers": {
    "runbook-ai": {
      "command": "npx",
      "args": ["-y", "runbook-ai-mcp@latest", "--token", "<your token>"]
    }
  }
}

or for Claude Code (Copy setup command in the settings copies this line):

claude mcp add runbook-ai -- npx -y runbook-ai-mcp@latest --token <your token>

The server also reads RUNBOOK_AI_MCP_TOKEN when --token is absent; prefer it on shared machines, since command-line arguments are visible to other users in ps. A request with a missing or wrong token returns Error: invalid-mcp-token: .... An empty token in the extension settings turns the check off.

By default the extension has access to all websites. If you want to limit the access, go to Chrome Extension Details, and add individual sites to Site access setting.

Usage

Open Chrome and keep the extension side panel open.

Start the MCP server (it will automatically start when invoked by your MCP client).

Tool Schema

The server exposes a single tool:

browser-agent

Run a task in Chrome browser with AI and automation capabilities.

Parameters:

• prompt (string, required): The task prompt for the AI agent to execute
• maxIterations (number, optional): Maximum number of agent iterations for the task (default: 15). Each iteration is one agent action (navigate, click, type, etc.); raise this for long multi-page tasks. Token budgets scale with it.
• ephemeral (boolean, optional, default true): Each call runs in an isolated browser session — it starts on a fresh blank tab, cannot see tabs left by previous calls, and closes every tab

... View full README on GitHub