fast-agentic-browser, or fab, is a command-line tool that lets AI agents describe browser tasks in plain English and receive JSON results. It can navigate sites, sign in with saved credentials, fill forms, scrape paginated data, inspect detail pages, and stop before a sensitive action such as placing an order. In a blind evaluation of 40 tasks across 20 unseen applications using the same mercury-2.5 planner, fab reported a 92.5% pass rate versus 73.1% for an LLM paired with chrome-devtools-mcp; median task time was 16.6 seconds versus 40.7 seconds, and cost was $0.0044 versus $0.0133. The project says its scraping requests are converted into small programs, while an LLM maps page layouts to fields and code handles pagination, loops, conditions, limits, and parallel detail-page retrieval. A separate 20-case generated-site scrape evaluation covered 1,746 records and varied page structures; it was exact in most runs, with occasional field-mapping errors, and 1,000 products took about 20 seconds. Every command emits JSON Lines containing start, record, and end or typed-error events, and agents can declare a JSON Schema for the returned shape. Credentials from 1Password, Bitwarden, or macOS Keychain are typed only on their saved sites, remain masked in output, and require one-time approval per site for cards, identities, or keys. The tool supports shared browser sessions, concurrent tabs, crash-resistant tasks, multiple browsers, headless mode, and a running-browser connection. Installation uses Cargo and requires an OPENROUTER_API_KEY for the high-level workflow; the repository also includes benchmark and demo tooling.
