
Browser-automation command line...
Putting a full page into the chat fills the context quickly. This tool first returns a compact accessibility tree where each element has a reference id, and later clicks and screenshots point at that id. A resident daemon stays up between commands. After an idle period it closes a headless browser; headed browsers and browsers you attached yourself are exempt. Without a restore setting, shutdown drops temporary tabs. It fits developers who want an assistant to navigate, fill forms, and capture screenshots while keeping the output short enough to stay in context.
agent-browser is a browser-automation command line from Vercel Labs for coding assistants that can run shell commands. A Rust client and a resident daemon drive the browser through the debugging protocol and return the page as an accessibility tree with reference ids, so the assistant clicks by id instead of putting the whole page into the conversation. Sessions, profiles, auth state, and proxies can be kept. Recording, debugging, and network controls are included. The first run needs a browser binary.
Difficulty: Intermediate
Element refs from a snapshot
A snapshot returns an accessibility tree with reference ids. Later clicks use that id and do not query the whole document again.
A daemon that drives the browser
A Rust daemon manages the browser over the debugging protocol and stays running between commands, so each command does not cold-start the browser.
Sessions and auth state
Sessions, profiles, auth state, storage, and proxies let a multi-step task keep the same login.
Recording, debugging, and network control
Recording, streaming, a debugger, a profiler, and page diffing are built in, along with network-control commands.
Open a page and click
The assistant opens a URL, takes a snapshot, clicks a link by its reference id, and saves a screenshot.
Keep a login across steps
For a sequence of dashboard actions, auth state stays in the session so each command does not log in again.
Compare a page before and after
Before and after a redesign, page diffing and a recording show whether the click path still completes as expected.
The snapshot command returns a compact accessibility tree and each element has a reference id. Later clicks use that id, so the assistant does not query the whole page again or paste the full document into the chat. The id comes from that snapshot, so the assistant snapshots before it acts. A screenshot command can save the page as it looked then.
The daemon stays up between commands. By default, after a period with no commands or dashboard input, it closes the headless browser and exits. Headed browsers, Safari and iOS sessions, and browsers you attached yourself are exempt from that timeout. Without a restore setting, shutdown drops temporary tabs. The timeout can be changed with a flag or an environment variable, including a value that disables it.
The site says any assistant that can run shell commands can use it, and it names several coding assistants. Install options include a global package, a system package manager, or a one-off run. The first time you also prepare a browser binary. It is not a plugin for a single assistant. Session, proxy, and security settings live in the command-line configuration.
Real reviews and feedback from users