agent-browser — AI Agent Browser Automation CLI (2026)

Main

• The core ergonomic win is snapshot refs (like @e2) so agents can click/fill by stable accessibility-derived handles instead of brittle CSS selectors.
• Installation stays pragmatic: npm/Homebrew/Cargo paths are documented, and agent-browser install can fetch Chrome from the official Chrome-for-Testing channel when needed.
• The command surface is intentionally CLI-first (open/click/fill/find/screenshot/pdf/stream), which makes it easy to wrap into scripts, task runners, or agent toolchains.
• For traditional automation, it still supports CSS selectors and role-based find commands; you can start with refs and drop to selectors when you need precision.

FAQ

• Do I need Playwright in my project?: Not necessarily. The README positions agent-browser as a standalone CLI; you install it and drive Chrome via the CLI commands and refs.
• Why run agent-browser install?: It can download Chrome from the official Chrome for Testing channel (first run) and helps standardize automation environments across machines.
• What should I verify first?: Open a URL, run snapshot, then click @e... and take a screenshot so you validate the ref-based interaction loop end-to-end.