How to use the agent-browser agent skill: setup and examples

What the agent-browser skill does

Agent skills are folders of instructions that an AI agent loads when a task matches them. This one covers browser work: navigating pages, filling forms, clicking buttons, taking screenshots, extracting data, and testing web apps.

The tool it relies on is a Rust-based CLI. It controls Chrome or Chromium over CDP, and it presents pages as accessibility-tree snapshots with short element references such as @eN, so the agent can refer to a specific button or field reliably.

• Navigate pages and interact with elements by reference
• Fill forms and click buttons
• Take screenshots and extract data
• Test and debug web apps, including exploratory QA
• Support for sessions, state persistence and video recording

Install the agent-browser skill

Download the skill ZIP from this page and unzip it. The skill is a folder containing SKILL.md. Put it in one of these locations:

For a personal install across projects, place the unzipped folder in ~/.claude/skills/. For a single project, place it in that project's .claude/skills/ folder. In the Claude apps, you can instead upload the ZIP under Settings → Capabilities → Skills.

• Personal: ~/.claude/skills/
• Per project: .claude/skills/ inside the project
• Claude apps: Settings → Capabilities → Skills, then upload the ZIP

Install the CLI the skill needs

The skill is only a stub. It needs the agent-browser CLI, which requires Node.js with npm. Install the CLI globally, then install the browser it controls.

Run these commands in a terminal. The skill's own instructions also tell the agent to load its workflow content from the installed CLI, so the instructions match the installed version.

How to trigger the skill with natural requests

You do not need special syntax. Describe the browser task in plain language and the agent should pick up the skill when the request matches its description. Typical trigger phrases include "open a website", "fill out a form", "click a button", "take a screenshot", "scrape data from a page", "test this web app" and "login to a site".

The skill also covers specialized areas. For a desktop Electron app such as VS Code, Slack, Discord, Figma, Notion or Spotify, or for Slack tasks like checking unreads or sending messages, the agent loads a sub-skill for that work. Be specific about the site or app and the outcome you want.

• "Open example.com and take a screenshot of the pricing table."
• "Fill out the contact form on my staging site with the test data in test.json."
• "Test this web app's signup flow and report any bugs you find."

Examples of useful tasks

The skill fits tasks where a page has no API or where you need to exercise an interface as a user would. Developers use it to test or debug a web app in a real browser. Analysts use it to pull data from sites that have no export. Anyone with repetitive web chores, such as filling out the same form weekly, can hand those steps to the agent.

Keep requests concrete: the URL, the steps, what counts as success and where to save output.

• Reproduce a bug on your staging app and capture a screenshot of the failure
• Pull a table of listings from a site that offers no API
• Fill a repetitive internal form on a site you own
• Run an exploratory QA pass and list what looks broken

Safety notes before you run it

This skill runs scripts and uses the network. The browser runs with real session access, so anything you are logged into can be reached by the agent. Review what the agent plans to do before it submits forms, logs in, or handles payments.

Only automate accounts and sites you have permission to use. Logins and scraping can breach a site's terms of service. Also note that launching a browser is slower and heavier than a normal fetch or search, so use a simpler tool for simple lookups.

• Use it on sites and accounts you own or are allowed to automate
• Review sensitive logins and payments step by step
• Avoid it for simple lookups that a fetch or search tool can answer

Tips for better results

Start with a short, specific request and check the first few actions. If the agent gets stuck, ask it to load the core workflow from the CLI with the command below, which includes workflows, common patterns and troubleshooting. Add the --full flag for the complete command reference.

• Name the exact page or element you mean
• Ask for a screenshot at key points so you can verify progress
• Use a staging environment for testing forms and flows