Agent Browser

•
3
•

Automates headless browser tasks using accessibility-tree snapshots for stable and deterministic AI agent interactions.

Agent Browser is a high-performance automation skill for Claude Code that leverages Vercel Labs' agent-browser CLI to interact with web pages through accessibility tree snapshots. By providing deterministic element references (refs) instead of relying on visual coordinates, it allows Claude to navigate complex single-page applications, manage multiple isolated sessions, and handle multi-step workflows with extreme precision. It is particularly useful for tasks requiring session persistence, network request mocking, or high-speed data extraction where traditional visual-based browsing might fail or be too slow.

主な機能

01JSON-optimized output designed for seamless parsing by AI agents
023 GitHub stars
03Deterministic ref-based element selection using accessibility tree snapshots
04Isolated browser sessions for simultaneous multi-user workflow testing
05State persistence to save and load cookies or local storage for skipping login flows
06Advanced network control including request blocking, routing, and mocking

ユースケース

01Extracting data from authenticated web portals using saved state sessions
02Automating complex multi-step workflows in modern JavaScript-heavy web applications
03Conducting end-to-end testing with distinct admin and user browser contexts