Actionbook provides developers with versioned, semantic JSON "action manuals" that describe key DOM elements and interaction flows, allowing AI agents to send only relevant page snippets to large language models. This reduces token usage and speeds up browser automation by up to tenfold while a browser extension executes real clicks, typing, and form submissions in the user's logged‑in session, supporting parallel tabs and human‑in‑the‑loop oversight.
Funding
Funding not disclosed
Founders
Product
Problem
Creating browser-based AI agents is time‑consuming and fragile: agents must parse full HTML, send large DOMs to large language models, and rely on hard‑coded selectors that break when websites change, leading to slow execution, high token costs, and frequent errors.
Solution
Actionbook supplies up‑to‑date “action manuals” that embed concise, semantic JSON definitions of relevant DOM elements and the exact steps needed to interact with a page. By providing these targeted selectors directly to the LLM, agents avoid full‑page parsing, reducing token usage and execution time by an order of magnitude. The manuals are versioned and automatically refreshed when site layouts change, making automation resilient to UI updates. Actionbook integrates with any LLM (OpenAI, Anthropic, Gemini, etc.) and any AI operator framework, delivering real‑browser actions—clicks, typing, form fills—through a browser extension that runs in the user’s logged‑in session. Developers can run multiple tabs in parallel and retain human‑in‑the‑loop oversight, ensuring control over critical actions.
Target Audience
Actionbook is aimed at developers and product teams building AI‑driven web automation, AI platform providers, and enterprises that need reliable, scalable browser agents for tasks such as data extraction, form submission, and multi‑site workflows.
Features
- Action manuals: pre‑computed, versioned JSON descriptors of page elements and interaction flows
- Token efficiency: delivers only relevant DOM snippets, achieving up to 100× token savings versus full HTML transmission
- Speed boost: agents execute tasks up to 10× faster by skipping exhaustive page parsing
- Resilient selectors: automatic manual updates keep automation functional despite UI changes
- Universal LLM compatibility: works with OpenAI, Anthropic, Gemini, and any custom LLM via standard APIs
- Browser extension runtime: performs real clicks, typing, and form submissions within the user’s authenticated session
- Parallel tab management: enables agents to operate across multiple browser tabs simultaneously
- Human‑in‑the‑loop UI: lets operators monitor, intervene, and approve actions before execution