agent-browser Explained: Browser Automation for AI Agents

CODE LAB · PROJECT SPOTLIGHT

agent-browser Rust · Apache 2.0

Vercel Labs browser automation CLI for AI agents, with interactive snapshots and browser actions.

GitHub source ↗Releases ↗

Preview: vercel-labs/agent-browser on GitHub

agent-browser is an open-source browser automation tool from Vercel Labs designed for AI agents. It gives an agent a command-line way to open pages, inspect their structure and interact with elements, providing a bridge between a language model’s decisions and a browser session.

How the workflow works

An agent can open a page, ask for an interactive snapshot, and use the returned element references to take actions such as clicking. A snapshot gives a compact view of page elements; references like @e1 connect a later command to an element in that view. Pages change, so take a fresh snapshot after navigation or significant updates rather than assuming an old reference still points to the same target.

170fn normalize_snapshot_name(name: &str) -> String {171    name.chars()172        .filter_map(|ch| match ch {173            '\u{00A0}' => Some(' '),174            ch if ZERO_WIDTH_CHARS.contains(&ch) => None,175            ch => Some(ch),176        })177        .collect()178}

Source: Vercel Labs and contributors, Apache License 2.0. Excerpt linked to the reviewed revision.

Where it fits

Browser automation can help with testing and supervised research, but an agent should still verify the page before acting. Login states, changing interfaces and destructive actions require appropriate human controls. For a complementary coding workflow, see our T3 Code introduction.

Sources: Repository · Releases · README