Browser Agent

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 it opened when it finishes. Pass false to continue from the tabs of a previous call and leave the final page open (e.g. multi-call workflows that build on the same page).
• effort (string, optional, default normal): How much exploration the agent invests — quick (one fast pass over loaded content, missing optional details reported as "not specified", tighter iteration budget), normal (exploration matched to what the ask requires), or thorough (follow all pagination, open detail pages, check candidates one by one, larger iteration budget). Accuracy rules apply at every level.
• outputDir (string, optional): Directory on the machine running the MCP server where files produced by the agent are written (created if missing; same-named files are overwritten). Default: a fresh per-call directory runbook-ai-mcp/task-<timestamp>-<id> under $RUNBOOK_AI_FILES_DIR if set, else the OS temp dir.

Files: Anything the agent saves during the run — data it writes to a file (e.g. a scraped list or an API payload it captured with saveToFile), downloads, extracted datasets, screenshots — is written to disk and listed at the end of the result text as absolute paths with MIME type and size. File content is never inlined into the result, so bulk data stays out of your context; read the files with your own tools. Image files (≤ 4 MB) are additionally returned as inline MCP image content so screenshots are visible directly.

To get bulk data as a file, say so in the prompt, e.g. "... collect all orders from the account page and save them to orders.json".

Example:

{
  "name": "browser-agent",
  "arguments": {
    "prompt": "Go to google.com and search for 'MCP protocol'"
  }
}

Development

# Install dependencies
npm install

# Build
npm run build

# Run in development mode
npm run dev

# Run tests
npm test

Architecture

1. MCP Server: Communicates with MCP clients via stdio
2. WebSocket Server: Listens for Chrome extension connections on port 9003
3. Chrome Extension: Executes browser automation tasks

When a tool is invoked:

1. MCP client sends request to MCP server via stdio
2. MCP server forwards request to Chrome extension via WebSocket
3. Extension executes the task and returns result
4. Result is sent back to MCP client; files the agent produced are written to disk (outputDir) and their paths appended to the result

Contributing

Contributions are welcome! Feel free to send out a PR.