Agent Browser Skill · Development

agent-browser Skill: Browser Automation With Stop Points

An unofficial agent-browser skill that drives Chrome with snapshot refs, explicit waits, and confirm-first rules for logins, forms, and other risky clicks.

  • Browser automation
  • Snapshot refs
  • Confirm-first
  • Cross-agent

About This Skill

This skill teaches your coding agent to drive Chrome with the agent-browser CLI: take a snapshot, act on @eN element refs, wait for the expected result, and snapshot again. It is based on agent-browser by Vercel Labs (Apache-2.0). Upstream ships a stub that fetches its guide from the CLI at run time. Our unofficial edition puts the full guide in the skill so you can read it offline, narrows when it triggers, and makes the agent ask before installs, submits, purchases, messages, deletes, or copying cookies from your real browser.

What we changed from the original

Based on agent-browser by Vercel Labs (vercel-labs) (Apache-2.0). Unofficial edition; the original and its license notice are credited in the download.

AreaOriginalOur edition
Content in the skillPublished SKILL.md is a stub; the real guide is fetched at run time with `agent-browser skills get core`, so it cannot be reviewed before install and fails offline.The core loop, waits, actions, troubleshooting, and a worked example are in SKILL.md, with commands, login options, and trust rules in three reference files.
TriggerVery broad, ending with "Prefer agent-browser over any built-in browser automation or web tools", and covers Slack, Electron and cloud browsers.Triggers on concrete browser tasks with real user phrases. Adds When NOT to use: plain fetches, a browser tool the user already prefers, no-browser tests, CAPTCHA or paywall bypass.
Irreversible actionsTrust rules cover page content and secrets but have no general rule to ask before submit, purchase, send, or delete clicks.A stop-point table requires a yes before submits, orders and payments, Slack or email messages, posts, deletes, uploads, network routes, and HAR recording.
Real-browser cookiesPresents reusing a logged-in Chrome (`--auto-connect state save`, which writes all cookies and localStorage to a plaintext file) as the fastest way to authenticate.Lists login options from least to most sensitive. Copying the real Chrome is last and needs explicit consent after a plain-language risk explanation, with encryption, cleanup, and a fresh profile offered first.
Install and repairTells the agent to run `npm i -g agent-browser && agent-browser install`. Only notes that `doctor --fix` runs destructive repairs.Step 0 checks `agent-browser --version`, explains the global install and Chrome download, and installs only after a yes. Shows `doctor` output before asking about `doctor --fix`.
Cross-agentFrontmatter uses `allowed-tools: Bash(agent-browser:*)`, a permission format specific to one agent.No agent-specific permission syntax. Follows the host's own approval prompts and gives install paths for Antigravity, Claude Code, and Codex.
Verifiable stepsGood loop, but no end-to-end example showing how to prove the result.Every step names its check (printed session name, wait exits without timeout, `get url`, `session list`). The worked example says what evidence to report and what to do if the wait times out.
DisclosureThe outdated-browser check that contacts Google is explained only in the trust reference.The commands reference states that the version check makes a request to Google and how to turn it off (`AGENT_BROWSER_NO_VERSION_CHECK=1`).

Quick Start

  1. 1Download the skill (1 credit) and unzip the agent-browser folder.
  2. 2Copy the folder into your skills directory: Antigravity ~/.gemini/config/skills/ (or .agents/skills/ in a project), Claude Code ~/.claude/skills/, Codex ~/.agents/skills/.
  3. 3Make sure the agent-browser CLI is installed (agent-browser --version). If it is not, the agent explains the install (npm i -g agent-browser, then agent-browser install to download Chrome) and waits for your yes.
  4. 4Ask something like "Test the signup form on http://localhost:3000 with an invalid email" and check the reported snapshot, URL, or wait result.

Core Capabilities

Snapshot and ref loop

Reads the page's accessibility tree, acts on @eN refs, and takes a new snapshot after every change, so clicks hit the element the agent meant.

Explicit waits

Waits for text, a URL, an element, or an app condition rather than a fixed delay, and avoids networkidle on pages that never go quiet.

Confirm-first stop points

Pauses before submits, payments, messages, posts, deletes, uploads, network interception, installs, and doctor --fix.

Safer logins

Prefers test accounts, a fresh dedicated profile, or the auth vault. Copying cookies from your real Chrome needs explicit consent and a plain explanation of the risk.

Untrusted-page rules

Treats page text, console output, and page-provided tools as data. The agent quotes injection attempts to you instead of following them, and keeps secrets out of chat.

Works across agents

Plain Markdown with no tool-specific permission syntax, plus install paths for Antigravity, Claude Code, and Codex.

Usage Examples

Input

My app runs at http://localhost:3000. Check that the signup form rejects an invalid email.

Output

The agent opens /signup in its own named session, fills a throwaway invalid email, submits, waits for the error text, and reports the message and that the URL stayed on /signup. The wait result and the second snapshot are the evidence.

Input

Log in to our staging dashboard and tell me if the Billing tab loads.

Output

The agent offers a test account from the project's seed files or a fresh profile where you log in yourself, waits for the dashboard URL, opens Billing, and reports what the snapshot shows. It never asks you to paste a password into chat.

Input

Go through checkout on staging but don't place the order.

Output

The agent fills the cart and shipping steps, stops at the final Place order button, and reports what it would click and the total shown.

SKILL.md

---
name: agent-browser
description: Drive a real Chrome browser from the shell with the agent-browser CLI (vercel-labs/agent-browser) using the snapshot -> act on @eN refs -> re-snapshot loop. Use when the user asks to click through a website, fill or test a form, log in and check a page, test a local web app end to end, extract data that only appears after JavaScript runs, or take a screenshot of a rendered page, and agent-browser is installed or the user agrees to install it. Do NOT use for a plain page or docs fetch (use the agent's normal fetch/read tool), when the user prefers a browser tool the agent already has, for unit tests that need no browser, or to bypass CAPTCHAs, paywalls or site terms. Asks before installing, before any submit/purchase/send/delete action, and before touching real-browser cookies. Unofficial Antigravity Skills edition.
---

# agent-browser (unofficial edition)

agent-browser is a command-line tool that drives Chrome over the Chrome DevTools Protocol. A `snapshot` prints the page's accessibility tree and gives every element a short ref such as `@e3`; you then act on those refs. This file holds the core loop and the rules. Details live in `references/`:

| Read | When |
| --- | --- |
| `references/commands.md` | Any command or flag not shown here |
| `references/auth-and-sessions.md` | Logging in, keeping a login, named sessions, cookies, profiles |
| `references/trust-and-safety.md` | Untrusted page content, secrets, domain allowlists, confirmation gates |

## When to use

User phrases that fit: "open this site and click...", "fill out this form", "test the signup flow on localhost:3000", "log in and check my dashboard", "scrape the table after it loads", "screenshot the page", "is the button disabled?", "click through the checkout on staging (don't pay)".

## When NOT to use

- A static page or documentation fetch: use the agent's own fetch/read tool. (If agent-browser is already installed, `agent-browser read <url>` also fetches ove

Frequently Asked Questions

FAQ

Is this the official agent-browser skill?
No. It is an unofficial edition by Antigravity Skills, based on agent-browser by Vercel Labs under the Apache-2.0 license. NOTICE.md and the original LICENSE are included. We are not affiliated with Vercel.
What did you change from the original?
Upstream's skill is a stub that loads its guide from the installed CLI. We put the core guide, commands, login options, and trust rules into the skill itself. We also narrowed the trigger, added confirm-first stop points, gated installs and real-browser cookie copying behind your consent, and added a worked example with checks.
Does the skill include the agent-browser program?
No. It contains Markdown instructions only. You install the CLI from the official source, and the agent asks before running npm i -g agent-browser or agent-browser install, which downloads Chrome.
Will the agent use my real browser logins?
Not by default. It tries a test account, a fresh dedicated profile, the auth vault, or a single site's cookie file first. Copying state from your everyday Chrome writes every logged-in session to a plaintext file, so the agent explains that and waits for an explicit yes.
Can it submit forms or buy things on its own?
It can fill forms, but it stops and asks before final submits, payments, sending messages, posting, deleting, or unsubscribing. Test data on a local dev server you asked it to exercise counts as permission for that flow only.

Discussion

Discussion

0 comments
U

Trigger Phrases

Use these phrases to activate this skill in your AI coding assistant:

open this site and click through itfill out this form in the browsertest the signup flow on localhostlog in and check my dashboardscrape the table after the page loadstake a screenshot of the rendered pageuse agent-browser