agent-browser: the agent opens the site, clicks and checks itself
Testing a web app properly takes a real browser, not just reading HTML. agent-browser is a fast Rust CLI that drives Chrome over CDP: the agent gets the page as an accessibility tree with compact refs like @e2, then clicks, fills fields and takes screenshots through those refs.
open, click, fill, type, select, upload and drag act on refs from a snapshot or on ordinary CSS selectors
snapshot returns the accessibility tree with @eN refs, and a click fails loudly when the target is hidden, disabled or covered
screenshot with --annotate and --if-changed, saving a page as PDF, and running JavaScript with eval
read pulls page text without launching Chrome: it asks for Markdown, looks for the nearest llms.txt, then falls back to HTML
Extra skills via agent-browser skills get cover Electron apps, Slack, exploratory testing, Vercel Sandbox and AgentCore
When to use it
You want the agent to walk a user flow in your web app and show the exact step where it breaks
You need data from a page whose content only appears after it renders in a browser
You want exploratory QA where the agent roams the interface hunting for bugs
Ask your agent
Open localhost:3000, sign up with a test user and screenshot the result
Take a snapshot of /pricing and check that every button has a label
Use agent-browser read on this docs URL and find the auth section
Questions
How do I install agent-browser?
Add the skill with npx skills add vercel-labs/agent-browser@agent-browser and the CLI with npm install -g agent-browser. Then run agent-browser install to fetch Chrome.
Does agent-browser work with Codex or Cursor?
Yes. SKILL.md names Cursor, Claude Code, Codex, Continue and Windsurf; it is a plain CLI any agent with a terminal can call.
Do I need Playwright or Node.js?
No. It talks to Chrome directly over CDP; Node.js 24+, pnpm and Rust are only needed to build from source.
