AI column

Automation@donghyeok99Staff

agent-browser: a browser your agent drives from the shell

Turns any web page into short text tagged @e1, @e2 and on, so your agent clicks by ref

open loads a login page, snapshot -i tags each element with a ref from @e1 to @e5, and click @e5 presses Continue.
open loads a login page, snapshot -i tags each element with a ref from @e1 to @e5, and click @e5 presses Continue.
Who
Made by Vercel Labs, for anyone who wants a coding agent to open a site, click through it and check the result.
When
The repository opened in January 2026. v0.42.0, out on November 10, 2026, added saved flows that replay a recorded task.
Where
A native Rust CLI for macOS, Linux and Windows. Any agent that runs shell commands can use it, and it also runs as an MCP server.

What was hard

Ask an agent to click through your web app and it ends up taking screenshots or reading the whole HTML, guessing where the button is. Launch a fresh browser for every click and the login is gone too.

What it does about it

agent-browser reads the page as an accessibility tree and returns short text with a ref such as @e1 on every element you can act on. The agent clicks or types by ref, then takes a new snapshot after the page changes to see what happened. The browser stays open between commands, so each step is one shell line. Per the README, a click on a target covered by something like a consent banner reports a failure rather than success.

agent-browser open example.comagent-browser snapshot -i       # interactive elements, each with a ref- link "Learn more" [ref=e1]agent-browser click @e1         # act on a ref from that snapshotagent-browser get url           # check the result
The README's own run
vercel-labs/agent-browser

What people made with it

abhimanyuaryan.github.ioAbhimanyu Aryan2026-03-24

Had an agent check a submenu on a local site: it opened the page, took a snapshot and reported the dropdown. The post says it replaced Playwright MCP in the author's workflow.

it navigated, took a snapshot, and reported back the dropdown contents
alexop.devalexop.dev2026-04-11

Built a QA workflow where Claude Code drives a real browser through agent-browser, explores an app like a user and returns a JSON report of the bugs it found.

Claude explores it like a user: clicking navigation, filling forms, checking for JS errors
Hacker Newsmiteshashar2026-05-05

Shared a video on Hacker News of agent-browser in Claude Code automating workflows in the accounting app Zoho Books.

Using Agent-Browser in Claude Code to Automate Zoho Books Workflows
likes 2 · comments 1 (2026-11-11 as of)Hacker News ↗

How it runs inside

  1. 1

    You → Claude Code · Codex

    Fill out the form on this siteThe installed skill is a thin stub: on form, click or screenshot requests it tells the agent to load the real guide.skills/agent-browser/SKILL.md
  2. 2

    Claude Code · Codex → agent-browser CLI

    agent-browser skills get coreThe CLI serves the core guide (snapshots, refs, waiting) from skill-data, matching the installed version.cli/src/skills.rs
  3. 3

    agent-browser CLI → Background daemon

    agent-browser open example.comThe first command starts a daemon for the session; later commands connect to the same daemon over its socket.cli/src/connection.rs
  4. 4

    Background daemon → Chrome for Testing

    Chrome DevTools ProtocolThe daemon keeps the browser open between commands, and closes a headless one after an hour with no commands.cli/src/native/daemon.rs
  5. 5

    Chrome for Testing → Claude Code · Codex

    snapshot -i: elements with [ref=e1]From the Accessibility.getFullAXTree result it numbers each actionable element e1, e2 and prints short text.cli/src/native/snapshot.rs
  6. 6

    Claude Code · Codex → Chrome for Testing

    click @e1It finds the element by ref, takes its centre and sends a click; a hidden, disabled or covered target fails instead.cli/src/native/interaction.rs
  7. 7

    Chrome for Testing → Checked result and screenshot

    get url · screenshotA new snapshot or the URL confirms the result; a screenshot is saved only when it is asked for.cli/src/native/screenshot.rs

What goes in, what comes out

agent-browser snapshot -i (example.com)One line, - link "Learn more" [ref=e1], ready for click @e1.
agent-browser find role button click --name "Submit"Finds the Submit button by role and name and clicks it, no ref needed.
Inspect the submenu on my local siteThe agent opened the page, took a snapshot and reported the dropdown contents (Abhimanyu Aryan's post).
agent-browser skills get dogfoodLoads the guide for exploratory testing and bug hunts that ends in a report.

Is it for you

Good for

  • Having an agent click through your web app in a real browser before you ship
  • Handing repeated web chores such as logins and forms to an agent
  • When a few shell commands suit you better than Playwright MCP
  • Several agents that each need a browser session of their own

Not for

  • Refs describe the page at snapshot time; after navigation or a DOM change, take a new snapshot.
  • The first run downloads Chrome for Testing with agent-browser install; on Linux, --with-deps adds the system libraries.
  • The unnamed default session is one browser shared by every agent on the machine, so give each task its own --session name.
  • Apache-2.0: companies may use and modify it, keeping the license and marking files they change.

Reviewed by Devalley’s editor before publishing. How the column is made

Sign in to post, comment and like.Sign in

Report

Reason

Read the community rules