Step-by-step: let Claude Code drive a real browser for you with agent-browser, from login to export图文教程：用 agent-browser 让 Claude Code 替你操作真实浏览器，从登录到导出一遍跑通

On Windows, download the agent-browser executable below, put it in any folder and add that folder to your PATH; or just run it on the fly with npx agent-browser.

In Claude Code, say: "Install agent-browser from https://github.com/vercel-labs/agent-browser" and confirm the skill installation when prompted.

After installation, ask it to run agent-browser --version; a version number means everything is wired up.

State the goal and grant permission: "open https://example.com, log in with the account saved in my environment variables, then screenshot the page".

It decomposes the task into open, fill, click and snapshot commands and runs them step by step — every action is visible in your terminal.

For sites with CAPTCHA or two-factor login, have it pause at that step, complete it manually, then let it continue.

Ask for a snapshot extraction: "pull the title, price and link from every row of this list page into JSON".

For pagination, just say "keep extracting through page 5" — it loops and merges the results.

Finally, have it save the JSON to a file, or emit a Markdown table you can paste straight into your docs.

If results look wrong, ask for an element snapshot and check whether the selector picked what you expected.

Use the built-in recording to replay the whole run and spot the failing action.

For long jobs, hand the agent small chunks and confirm each before continuing — far more reliable than one giant instruction.

This article is compiled from public sources and ships with the matching resource. Product features and pricing are subject to the official page. Resources are for learning and exchange only — please respect the original license.