Agent Browser Review: Give Your Coding Agent a Real Browser

Agents are blind on the web. They can fetch a page and read the text, but they cannot see how it renders, click the button that only appears after JavaScript runs, or check whether the checkout flow actually works. The moment a task needs a real browser, the agent has to ask a human to look.

Agent Browser fixes that. It is Vercel Labs' browser automation CLI for AI agents, with a matching skill that teaches the agent how to drive it. The repository holds nearly 48,000 GitHub stars, the skill has about 877,000 installs on skills.sh, and the whole thing is Apache-2.0 licensed. Your agent gets a headless browser it can open, click through, screenshot, and record.

What is Agent Browser?

Agent Browser is two things that work together. The CLI is a real browser automation tool built on Playwright: it opens pages, navigates, interacts with elements through a simple reference system, takes screenshots, and records video of a session. The skill is a thin discovery stub checked into the repository: it tells the agent to install the CLI, then load the workflow guide that matches the installed binary version, so the instructions never go stale.

That version-matched guide is the clever part. Browser automation changes fast, and a skill written against last month's CLI would teach the wrong flags. By fetching the guide at runtime, the skill always teaches the current behavior.

Key Features

1. Real page interaction, not just fetching

The agent navigates pages, clicks elements by reference, fills forms, and scrolls. It can reach the parts of a web app that only exist after JavaScript runs, which a plain HTTP fetch never sees.

2. Screenshots for visual verification

The agent can screenshot a page and look at the result. That closes the loop on frontend work: build the page, screenshot it, see the broken layout, fix it, and screenshot again, all without a human in the middle.

3. Session recording

Record a browser session as video, or capture a contact sheet of the run. The repository ships a contact-sheet example that shows the whole flow at a glance, which is useful for debugging and for proving a flow works.

4. Version-matched guides, loaded at runtime

The checked-in SKILL.md is a stub. The agent runs a command to fetch the full workflow guide for the exact CLI version installed. The instructions always match the binary, even as the CLI evolves.

5. Works across agents

The skill uses the open Agent Skills format, so it installs into Claude Code, Cursor, Codex, GitHub Copilot, Gemini CLI, and more through the skills CLI. The CLI itself is also available through npm and a shell installer.

How It Works

Install the skill with the skills CLI. The command below pulls the agent-browser skill from the Vercel Labs repository:

npx skills add https://github.com/vercel-labs/agent-browser --skill agent-browser

When the skill fires, the agent installs the CLI first, then loads the matching guide and starts driving the browser. You can also install the CLI directly and run it yourself: the npm package and the shell installer are both documented in the repository. After that, ask the agent to open a page, test a flow, or screenshot a design, and it works through the browser like a user would.

Who Should Use Agent Browser?

Frontend developers who want their agent to verify its own work are the core audience. It also fits QA work, where the agent walks a checkout or signup flow and reports what broke, and research tasks where the agent must interact with a page rather than read a static copy. If your agent only ever reads code, you do not need it.

Pricing

Agent Browser is free and open source under the Apache-2.0 license. The CLI, the skill, and the guides cost nothing. The browser automation runs locally on your machine, so there is no hosted service to pay for.

Pros and Cons

Pros: a real browser closes the verification gap for frontend and QA work; the version-matched guide design keeps instructions current; nearly 48,000 stars and close to a million installs; permissive Apache-2.0 license.

Cons: browser automation is a powerful capability, and independent security reviews have flagged items to check before installing, so review the skill and the CLI first; the skill is a stub, so the real guidance is fetched at runtime and needs network access; headless browsers can be slow and flaky on heavy pages; it needs a local install, not just the skill file.

Final Verdict

Agent Browser is the missing sense for coding agents: sight. Once your agent can open a page, click through it, and look at a screenshot, whole categories of verification stop needing a human. For frontend and QA work it is close to essential. Try it on GitHub: vercel-labs/agent-browser.