agent-browser

Lets Claude drive Chrome from the terminal—click, fill forms, take screenshots. Powerful for live UI checks, high blast radius if misused.

What it is for

Live UI verification, form flows, and screenshots without leaving the terminal loop. Claude can drive Chrome when a static code read is not enough—confirming a checkout path, capturing a regression, or checking that a deployed page actually behaves.

Rank #10, score 82/100 (Nov 21, 2026), with a high-risk warning. ~1.5M install reads—reach, not a free pass. Our ranking policy keeps high-risk capabilities out of the top three even when public adoption is large.

Why this score

Eighty-two is not a claim that the skill is “bad.” It is a claim that blast radius matters on an editorial shortlist. Browser control can touch authenticated sessions, downloads, and clicks that change real data. Usefulness is high for the right job; default recommendation for casual browsers of this site is “not yet.”

Who should install

Yes if you need automated browser checks and accept the risk—with a disposable profile and a non-production account for first runs. No if you are browsing skills casually—start with find-skills, grill-me, or a low-risk task skill.

Install vs skip examples

  • Install: You must verify a live staging flow Claude cannot see from source alone.
  • Install: Screenshot-based regression checks in a controlled environment.
  • Skip: You only needed package discovery or UI taste.
  • Skip: Your only browser profile is logged into production admin.

What SKILL.md typically covers

Read every permission and network assumption. Browser control can touch authenticated sessions if you point it at the wrong profile. Know how the skill launches Chrome, where profiles live, and what it can click or download. Do not skim this file.

Install

npx skills add vercel-labs/agent-browser --skill agent-browser

Mandatory: safe Claude skills checklist. Install guide. Prefer pins. Confirm vercel-labs/agent-browser.

How to use it with eyes open

  1. Create or select a disposable browser profile with no production cookies.
  2. Use a non-production account and staging URLs first.
  3. State the exact flow to verify; avoid open-ended “explore the site.”
  4. Watch the first runs; do not leave unsupervised clicking on sensitive systems.
  5. Remove the skill from everyday sessions when you are not doing live verification.

Common mistakes: pointing it at a daily-driver profile; treating install counts as safety; using browser control as a substitute for frontend-design taste work or for reading SKILL.md on other packages.

vs nearby

  • frontend-design — design taste vs live browser control. Different jobs; different risk.
  • find-skills — discovery vs driving Chrome.
  • skill-creator — authoring vs runtime browser access.
  • vercel-react-best-practices — engineering perf rules vs live verification. You may use both, but browser control is not a Lighthouse replacement by itself.

Risks and caveats

Highest blast radius on our shortlist: credentials, clicks, downloads, payment flows, admin panels. Use a disposable profile and a non-production account for first runs. If your security policy forbids agent-driven browsers, do not install it—no editorial score overrides policy.

Source and date

Vercel Labs, skill agent-browser, Nov 21, 2026.

Hub: Claude skills for research. Ranking: best Claude skills. Related: Claude Code skills, skill safety checklist.

FAQ

What is agent-browser?

A Vercel Labs skill that lets Claude control a browser from the terminal for clicks, forms, and screenshots.

How do I install agent-browser?

npx skills add vercel-labs/agent-browser --skill agent-browser

Why is it out of your top three?

High-risk capabilities keep a warning chip and cannot take the top three slots on our editorial ranking, even with large install counts.

Is agent-browser required for research?

No. Start with find-skills and grill-me. Add browser control only when a live page check is required and you accept the risk.

Related