actionbook

Actionbook is a browser tool for AI agents, with the goal of enabling agents to access data sources behind logins and paywalls and directly complete work, rather than just providing answers.

In terms of main capabilities, Actionbook offers the following features:

1. Read data after login: Agents can access data on websites where the user is already logged in, breaking through login walls.
2. Operate existing websites: Agents can operate everyday websites, including clicking, filling out forms, opening tabs, and more.
3. Deliver finished products: Output reports with sources, clean CSV files, or ready-to-send drafts, rather than simple text answers.
4. Save and reuse: When a task runs as expected, it can be saved as a Recipe and rerun later with new inputs.
5. Action Manuals: Tell the agent the exact steps through action manuals, reducing parsing and guesswork, and improving accuracy and speed.
6. Hybrid operation mode: Call APIs directly whenever possible, fall back to UI automation when necessary, and handle login state.
7. Concurrent operations: The stateless architecture supports parallel operation of multiple tabs, allowing dozens of tasks to be handled simultaneously.

In terms of usage, Actionbook runs as a remote MCP connector paired with a Chrome extension, with no CLI installation required. The specific steps are: first install the extension from the Chrome Web Store; then add Actionbook as a connector in an AI application, supporting Claude, ChatGPT, Perplexity, Hermes, OpenClaw, and other MCP-compatible applications; finally complete setup from the dashboard, connect the client, and confirm the extension is linked. After installation is complete, users only need to talk to the agent, and Actionbook will fetch action manuals in the background and operate the real Chrome browser. If explicit prompting is needed, specific instructions can be added to the prompt.

Actionbook provides a single actionbook MCP tool, whose subcommands cover action lookup (search, manual) and browser control (goto, snapshot, click, type, etc.). After connecting, the agent can discover the complete command set from the tool description.

Applicable scenarios include: needing to obtain data from websites after login, checking multiple websites in parallel, automating repetitive web tasks, and more. For example, in the video demonstration, the agent visited the websites of 192 First Round portfolio companies within 2 minutes and collected their taglines (videonot sped up).

In terms of limitations, the materials do not explicitly mention specific limitations. Note: The standalone @actionbookdev/cli package is deprecated; the remote connector plus Chrome extension approach should be used.