MCP Server

The Actionbook MCP server lets any MCP-compatible AI app operate websites for you. Your agent drives your own Chrome — with the accounts you’re already signed into — so it can research, fill forms, and run repetitive web tasks without logging in again or scraping raw HTML. It’s a remote server, so there’s nothing to install locally besides the browser extension.


Endpoint	`https://edge.actionbook.app/mcp`
Transport	Streamable HTTP
Auth	OAuth (recommended), or an API key as a Bearer token
Tools	One tool: `actionbook`

The fastest path is the guided setup at actionbook.app/onboarding. It walks you through all three steps below for 20+ AI apps.

Set up

Install the Actionbook extension

Install the extension for Chrome or Edge and sign in. The extension keeps a secure connection to Actionbook, which is how your agent’s browser commands reach your browser.

Connect Actionbook to your AI app

Add https://edge.actionbook.app/mcp as a remote MCP server in your AI app. See Connect your AI app for per-app steps.

Run your first task

Ask your agent something that needs a live page, then watch it happen in Chrome:

Use Actionbook to open Hacker News and tell me the title and a one-sentence summary of the top story right now.

Connect your AI app

Most apps support OAuth: you paste the server URL, then approve access in the Actionbook tab that opens. No API key needed.

Claude (claude.ai / Claude Desktop)

Open Settings → Connectors → Add custom connector.
Name it Actionbook and paste https://edge.actionbook.app/mcp.
Click Add, then Connect, and approve in the Actionbook tab.

ChatGPT

Turn on Developer Mode in ChatGPT.
Click Create app, name it Actionbook, and paste https://edge.actionbook.app/mcp into MCP Server URL.
Leave Authentication as OAuth, tick I understand and want to continue, then click Create.

Grok

Open Grok Connectors, click New connector, then Custom.
Name it Actionbook, paste https://edge.actionbook.app/mcp, and click Add Connector.
Approve access in the Actionbook tab.

Claude Code, Codex, Cursor, Antigravity CLI

These agents can install MCP servers themselves. Paste this prompt and approve the OAuth request in your browser:

Please install the Actionbook MCP server for me, then confirm it connected.

It's a remote (streamable HTTP) MCP server:
MCP Server URL: https://edge.actionbook.app/mcp

Actionbook uses OAuth — approve it in the browser when prompted.

Apps without OAuth (API key)

For apps that only accept a static header (for example Windsurf, Manus, Genspark, Trae), create an API key in the Actionbook dashboard and send it as a Bearer token. Most apps use a config like this:

{
  "mcpServers": {
    "actionbook": {
      "url": "https://edge.actionbook.app/mcp",
      "headers": {
        "Authorization": "Bearer ak_xxxxxxxxxxxxxxxx"
      }
    }
  }
}

Config keys differ between apps (mcpServers vs servers, url vs serverUrl). The guided setup shows the exact format for each one.

Treat your API key like a password. Anyone with it can drive the browser connected to your account.

What your agent can do

The server exposes a single actionbook tool that takes CLI-style commands.

Look up verified actions

Actionbook keeps a library of verified selectors and step-by-step manuals for popular sites. Agents check it first so they don’t have to guess at the page every turn.

Command	What it does
`search "<query>"`	Finds verified actions that match a task, such as `search "post a tweet on x.com"`
`manual --site <slug>`	Returns the exact selectors, parameters, and steps for a site’s actions

Control your browser

browser <verb> runs in your own Chrome:

Verbs	What they do
`goto`, `back`, `forward`, `reload`	Navigate
`tabs`, `start`, `new-tab`, `close`	Manage tabs
`snapshot`, `screenshot`	Read the page. `snapshot` returns element refs like `@e1` that later commands can use.
`eval`, `run-code`, `vision`	Act on the page with JavaScript, Playwright-style scripts, or screenshot coordinates

A typical run: search for a verified action (or snapshot the page), act on it, then snapshot again to check the result.

Troubleshooting

Browser commands fail or time out

Make sure the Actionbook extension is installed, you’re signed in, and Chrome is open. search and manual work without a browser, but browser commands need the extension.

My app can't find the Actionbook tool

Some apps turn connectors on per chat. In Perplexity, for example, click + → Connectors → Actionbook in the chat input. Restart desktop or CLI apps after adding the server.

Authorization failed

Remove the server and add it again to restart the OAuth flow. For API-key setups, check that the header is exactly Authorization: Bearer <key> and that the key hasn’t been revoked.

Getting Started

Guides

Set up

Connect your AI app

What your agent can do

Look up verified actions

Control your browser

Troubleshooting

​Set up

​Connect your AI app

​What your agent can do

​Look up verified actions

​Control your browser

​Troubleshooting

Set up

Connect your AI app

What your agent can do

Look up verified actions

Control your browser

Troubleshooting