agent-browser is a free, open-source tool from Vercel Labs that lets an AI coding agent drive a real web browser, one shell command at a time. It is a native Rust CLI under the Apache-2.0 licence and runs on macOS, Linux and Windows. It gained about 3,600 GitHub stars in the past week and now has about 49,700. Version 0.42.0 came out on 10 November.
How it works: the agent runs snapshot and gets the page as compact text from the accessibility tree, with a short ref such as @e1 for each element. It then clicks or types on a ref, takes a new snapshot to check the result, and repeats. A background daemon keeps the browser, tabs, cookies and storage open between commands, so the agent does not start a new browser for every step. By default it controls Chrome over the Chrome DevTools Protocol; it can also attach to a Chrome you already run, use other engines, Mobile Safari in the iOS Simulator, or cloud browsers.
What it is good for: letting an agent test the login, checkout or form it just built, read JavaScript-heavy pages, take screenshots or recordings for bug reports, inspect network requests and run accessibility checks with built-in axe-core rules, according to ScriptByAI's review. It works with any agent that can run shell commands, and it can also run as an MCP server (agent-browser mcp) or install as an agent skill. The new 0.42.0 adds experimental "saved flows" that record a browser task and replay it, and raises the minimum to Node 22.
Limits: refs only last until the page changes, so the agent has to take a new snapshot after navigation; a click fails when a cookie banner or another element covers the target. Standard automation is centred on Chrome/Chromium; iOS Safari needs macOS, Xcode and Appium. The built-in chat command needs a Vercel AI Gateway API key, but the core CLI and MCP mode do not. Saved session files can contain login tokens and should be protected. The project has had little Hacker News discussion so far; its traction shows mainly in GitHub stars.
