How to use the agent-browser skill for browser automation

What the agent-browser skill lets the agent do

An agent skill is a folder of instructions and tools that an AI agent can load when a task calls for them. This one is built around the agent-browser command-line tool. Once it is in place, the agent can navigate to a URL, read the interactive elements on the page, fill fields, select options, check boxes, click, wait for pages to load, scroll, and capture screenshots or PDFs.

The skill also keeps a background browser running between commands, so the agent can work through a multi-step flow without starting over each time. It can save cookies and localStorage so later runs stay signed in. Because the skill is a documented workflow, the agent follows the listed commands rather than guessing at browser syntax.

• Open pages and read interactive elements with references such as @e1 and @e2
• Fill forms, select dropdown options, check boxes and click buttons
• Wait for network idle, a URL pattern or a specific element
• Take screenshots, full-page captures or PDFs
• Extract text or JSON from elements on a page
• Save and reuse login state

Install the agent-browser skill

Download the ZIP for the skill from this page. You can install it in one of two ways. For a personal setup, unzip it into ~/.claude/skills/. For a single project, unzip it into the project's .claude/skills/ folder. If you use the Claude apps instead, open Settings, go to Capabilities, then Skills, and upload the ZIP there.

The skill calls a command-line tool, so you also need the agent-browser CLI and a browser runtime for it. The skill can invoke the CLI as npx agent-browser or as agent-browser if it is on your PATH. The source for this skill does not include a separate install command for the CLI, so follow the CLI's own documentation for that part.

• Personal install: unzip into ~/.claude/skills/
• Project install: unzip into your project's .claude/skills/
• Claude apps: Settings, then Capabilities, then Skills, then upload the ZIP
• Confirm that agent-browser runs from npx or from your PATH

Trigger the skill with natural requests

You do not need special syntax. Describe the browser task in plain language and the agent will load the skill when the request matches its description. The skill's trigger list includes requests such as open a website, fill out a form, click a button, take a screenshot, scrape data from a page, test this web app, login to a site and automate browser actions.

Be specific about the site and the outcome. A request that names the page, the fields and the result you want gives the agent what it needs to plan the steps.

• Open https://example.com/signup and fill in the form with my test details, then tell me what the confirmation page says
• Take a screenshot of the pricing page on this site
• Scrape the product names and prices from this listing page into a text file
• Test the checkout flow in my local web app and report any step that fails

Follow the core workflow: navigate, snapshot, interact

Every browser task in this skill follows the same pattern. The agent opens the page, takes a snapshot to get element references, interacts using those references, and then takes a fresh snapshot after anything changes the page. References look like @e1 and @e2, and they are only reliable until the page changes, which is why the workflow re-snapshots after navigation.

• Navigate with agent-browser open <url>
• Snapshot with agent-browser snapshot -i to list elements with refs
• Interact with click, fill, select or check using the refs
• Re-snapshot after navigation or DOM changes

Save login sessions and handle credentials safely

For sites that need a login, the skill can save state so later runs stay signed in. With the --session-name option, cookies and localStorage are saved automatically when the browser closes and loaded the next time you use that session name. Saved sessions are stored under ~/.agent-browser/sessions/. You can also save state to a file with agent-browser state save and load it with agent-browser state load.

To protect that state at rest, set the AGENT_BROWSER_ENCRYPTION_KEY environment variable before you start a session. Credentials for login flows are supplied through environment variables, so they do not need to appear in the command text. Use a test or low-stakes account while you learn the skill.

• Use --session-name to auto-save and restore state
• Saved sessions live in ~/.agent-browser/sessions/
• Use state list, state show, state clear and state clean to manage them
• Set AGENT_BROWSER_ENCRYPTION_KEY to encrypt saved state
• Pass usernames and passwords through environment variables such as $USERNAME and $PASSWORD

Get data out of pages and fix common problems

For data extraction, the agent can read a single element with get text, dump the page text, or add --json to snapshot and get commands for parsing. Some get subcommands do not accept @refs. According to the skill's local notes, get html, get box, get styles, get value, get attr and get count need a CSS selector instead of a ref, and a ref there can fail silently.

If a command behaves unexpectedly, check the help for the CLI version you have installed, since the local notes say command details may differ between releases. Re-snapshot if the page has changed, and verify the result yourself rather than trusting the agent's summary.

• Use get text with @refs or CSS selectors
• Use get html, get box, get value or get attr with CSS selectors, not @refs
• Add --json to snapshot or get for structured output
• Use --headed to watch the browser while debugging
• Use screenshot --annotate to see numbered element labels