agent-browser

agent-browser lets AI agents interact with websites through compact text snapshots and references to specific page elements. It's for developers connecting coding assistants or other agents to a browser without sending them the full page HTML.

• Page interaction: Agents can follow links, fill forms, work with tabs and frames, and interact with local HTML or PDF files through optional Chromium file access.
• Compact snapshots: Accessibility-tree snapshots give agents a smaller page representation and references that identify the elements to act on.
• Inspection and testing: Screenshots, video recording, streaming, network controls, profiling and page comparisons help you examine browser behavior.
• Persistent sessions: Sessions, profiles and saved authentication state let agents continue work across browser interactions.
• Optional safeguards: An encrypted local credential vault keeps passwords out of the model's context, while domain restrictions, action policies and content boundary markers add controls for agent access.

The Apache 2.0 tool runs locally on macOS, Linux and Windows. It works with Claude Code, Cursor, GitHub Copilot, OpenAI Codex, Google Gemini, opencode and other agents that can execute shell commands.

Chrome is the default browser; Lightpanda, Obscura and Safari for iOS are also supported. The native Rust daemon doesn't require Node.js or Playwright. Browser-provider plugins can use cloud browsers instead of a local browser. Security controls are opt-in, and the domain allowlist doesn't support every browser or connection mode.