Vercel Labs released agent-browser, a shell CLI for AI agents, 46K stars

A coding agent that must click a button in a web page usually uses a vendor API or Playwright. Vercel Labs released agent-browser this week to remove that step. It is a Rust command-line tool for AI agents. The repository has 46,186 stars. It was ranked 22nd on Trendshift's daily list on 6 November 2026. The licence is Apache 2.0.

The pitch, from the repository description: "Browser automation CLI for AI agents". Every action is a shell command. The agent runs a command, reads the output as text, and runs the next command. The browser session stays open between commands.

What the shell gives an agent

The README lists four main command groups. Navigation and reading include opening a page, taking an accessibility-tree snapshot, taking a screenshot, and saving a PDF. Interaction includes click, fill, type, scroll, drag and file upload. Network and state include cookies, storage and request control. A set of debug tools reports the browser console and React component tree.

The snapshot command is the point. A plain HTML page is thousands of lines, most of them markup an agent does not need. The accessibility-tree output lists only the elements a screen reader would see. Each element has a short reference label like @e1. The agent clicks @e1 or types into @e3. The output stays small enough to fit a model's context.

Which browsers it runs

The tool supports Chrome for Testing, mobile Safari and smaller browsers named Lightpanda and Obscura. A team can pick the browser that fits the test. Mobile Safari matters for a product sold through an iPhone app that renders inside a web view. The choice is one flag. The same shell command works on all of them.

The tool also integrates with Model Context Protocol clients and the agent-skills system. A coding agent like Claude Code or Codex can call agent-browser as a tool. The agent does not see the HTML. It sees the accessibility tree.

Who it is for

Three teams. One, a team that runs a coding agent against a web app it runs. The agent can check the app's real pages before merging a change. Two, a team that runs agents against a third-party site, like a booking tool or a supplier portal. Three, a team that runs end-to-end tests against its own product. The agent writes the test and runs it in the same session.

The honest limit is that a shell CLI is not always the right shape. A test team that already runs a Playwright suite has to decide whether to swap or run both. The accessibility-tree approach is cheaper on tokens than full HTML. A page that uses non-standard components can be missing from the tree. Teams should check a few of their own pages before switching.

Source

• Primary source: vercel-labs/agent-browser on GitHub.