Agent Browser

Agent Browser is Vercel Labs’ open-source browser-automation CLI (repo vercel-labs/agent-browser, site https://agent-browser.dev ) built for AI agents: compact text output to keep context small, implemented in native Rust. An agent opens a page, takes an accessibility-tree snapshot whose nodes carry refs, then clicks or fills forms with handles such as @e1 and @e2 instead of re-querying the DOM each turn.

The site lists 50+ commands for navigation, forms, screenshots, network, storage, files, tabs, frames, and debugging, plus built-in video recording, streaming, debugging, profiling, and diffing. Sessions, profiles, auth state, cookies, storage, proxy, and security controls are described as support for long-running agents. It works with any agent that can run shell commands; the homepage names Claude Code, Cursor, GitHub Copilot, OpenAI Codex, Google Gemini, and opencode.

Highlights:

• Install with npm install -g agent-browser on all platforms or brew install agent-browser on macOS, then run agent-browser install once to download Chrome. To try it without installing: npx agent-browser open example.com;
• Architecture is a Rust CLI plus a native daemon that talks to Chrome over the Chrome DevTools Protocol and stays up between commands;
• Native binaries cover macOS (ARM64 and x64), Linux (ARM64 and x64), and Windows (x64);
• It appeared on both GitHub daily and weekly trending in this window; radar recorded about 48.4k stars. It is not the same project as the existing FastToolHub term Agent Desktop.

Why it matters

Feeding a full DOM or a large JSON tree into an agent is expensive and makes it easy to click the wrong node. Agent Browser returns a compact text accessibility tree—about 200–400 tokens versus about 3,000–5,000 for a full DOM, according to the site—and stamps each element with a stable ref so later click and form commands point at the exact node from that snapshot. For Claude Code, Cursor, Codex, or any other agent that already runs shell commands, that is a path that does not require wrapping Playwright yourself.

Vercel Labs shipped it as a cross-platform native CLI rather than an IDE-only plugin, so it can sit inside different agent harnesses. Showing up on both GitHub’s daily and weekly trending lists in this window is a clear signal that agents need a browser tool whose output was designed for them.

Limits

It expects Chrome/Chromium in the environment where the agent runs, and the first use needs agent-browser install. By default the daemon saves configured restore state, closes its headless browser, and exits after an hour without commands or dashboard input. Without --restore or another restore key, shutdown discards transient browser state and open tabs. Headed browsers, Safari and iOS WebDriver sessions, and user-attached browsers are exempt from that default timeout; provider-owned cloud browsers are not. Tune it with --idle-timeout or AGENT_BROWSER_IDLE_TIMEOUT_MS, or set 0 to disable the timeout. First-class notes for React, Web Vitals, and Next.js + Vercel do not mean it can bypass every anti-automation or login wall. It is also not the desktop-OS product covered by Agent Desktop. The star count is a scan-time observation.

1. Install the CLI with npm install -g agent-browser or brew install agent-browser, then run agent-browser install to download Chrome. To try it once: npx agent-browser open example.com;
2. A typical loop is agent-browser open example.com, agent-browser snapshot -i to get a tree with [ref=e1] handles, agent-browser click @e2 (or other commands) to interact, agent-browser screenshot page.png if needed, and agent-browser close when finished;
3. Give it to any agent that can run shell commands (Claude Code, Cursor, Codex, Gemini, opencode, and similar) and have the model snapshot first, then use refs instead of guessing CSS selectors;
4. For longer jobs, read the site’s notes on sessions, profiles, --restore, and idle timeout. Docs and releases live at https://agent-browser.dev .