Why Actionbook?
Without Actionbook
Building reliable browser agents is difficult and expensive:- Slow Execution: Agents waste time parsing full HTML pages to find elements
- High Token Costs: Sending entire DOM trees to LLMs consumes massive context windows
- Brittle Selectors: Updates to website UIs break hardcoded selectors and agent logic immediately
- Hallucinations: LLMs often guess incorrect actions when faced with complex, unstructured DOMs
With Actionbook
Actionbook places up-to-date action manuals with the relevant DOM selectors directly into your LLM’s context:- 10x Faster: Agents access pre-computed “Action manuals” to know exactly what to do without exploring
- 100x Token Savings: Instead of whole HTML page, agents receive only related DOM elements in concise, semantic JSON definitions
- Resilient Automation: Action manuals are maintained and versioned. If a site changes, the manual is updated, not your agent
- Universal Compatibility: Works with any LLM (OpenAI, Anthropic, Gemini) and any AI operator framework
How It Works
Agent Requests Action
Your AI agent calls Actionbook MCP tools like
search_actions or get_action_by_id to find the action it needs.Actionbook Retrieves Manual
Actionbook looks up the pre-verified action manual from its database, containing precise selectors and operation methods.
Integration Options
CLI
Default mode — launches browser automatically
Browser Automation
Isolated or extension — two modes for AI agents
MCP Server
IDE integrations for Cursor, Claude Code, VS Code, etc.