Key takeaways

  • Give an AI agent a real browser it can drive one command at a time, straight from the shell.
  • Pages come back as compact text with a reference for each element, so an agent can act on a reference and read the page again.
  • The browser stays open between commands, so each step is one short shell call.
  • Failed clicks are reported as failures, not success, when a target is hidden, disabled or covered by a banner.
  • The catches: references go stale after page changes, it is labelled a Vercel Labs product, and we did not test it.

What it is

agent-browser is browser automation for AI agents: a native Rust command-line tool that lets an agent open pages, read them as an accessibility tree and act on elements. It is released under the Apache-2.0 licence, and the code is on GitHub at vercel-labs/agent-browser.

The problem it solves

An AI agent that needs to use a website has to see it first, and raw page markup is bulky and noisy. Developers building agents feel this most, because the agent needs a way to look at a page, choose an element and check what happened.

agent-browser answers with a loop. A snapshot returns the page as compact text with a reference for each element. The agent acts on a reference, reads the page again to check the result, and repeats.

What you get

  • Simple steps: the browser stays open between commands, so each step is one short shell call.
  • Honest failures: a click fails, instead of reporting success, when the target is hidden, disabled, outside the viewport or covered by something such as a consent banner.
  • Flexible browsers: Chrome for Testing is the default, and a cloud browser provider can be used instead.
  • Standard plumbing: it controls the browser over the Chrome DevTools Protocol, and an MCP server over stdio is available for MCP clients.
  • Guidance for agents: the tool ships skills that match its version.

The catches

  • Stale references: after navigation or a page change, the agent must take a new snapshot before using references again, so a careless agent can act on an outdated page.
  • Blocked clicks: a covered or hidden element means a failed click, and the agent has to clear the obstruction and retry.
  • Maturity: the README badge labels it a Vercel Labs product, which tells you little about long-term support.
  • Safety: the README points to security documentation covering an auth vault, domain allowlists, action policies and content boundaries. We have not read it, so check it before giving an agent access to logged-in sites.
  • Requirements: it relies on a downloaded browser, Chrome for Testing by default.

What we did not test

We did not run agent-browser, drive a browser with it, try its cloud providers or MCP server, or read its security documentation. Every claim here comes from a literal quote in the project's README or the GitHub API, and the feature descriptions are the project's own.

Verdict

Editors' view: this is aimed at developers building agents that must work with real websites and want a small, scriptable tool rather than a heavy framework. If your agent never touches the web, you can skip it. If you try it, point it at a harmless test page first and keep it away from accounts you care about.

Repository
vercel-labs/agent-browser
Rust · Apache-2.0
View on GitHub ↗

Sources

Need help adopting this, or have a software project in mind?

Tell us what you are building. The initial consultation is free. Email [email protected] or follow @zolnaltd.

Start a conversation →