HavenSkills

Claude Browser Automation Skills

Test browser workflows you own and turn authorized page interactions into repeatable checks.

Skills
6 skills
Full write-ups
2

How to choose

Choose a CLI for quick reproducible interactions, a test skill for assertions, or testing guidance for a flaky suite.

Filter

All skills in this category

agent-browser

Navigate pages, inspect snapshots and repeat common browser actions using a CLI.

by vercel-labs Skill guide
npx skills add vercel-labs/agent-browser --skill agent-browser
Author
vercel-labs
Repo
GitHub
Write-up
Read →

playwright-cli

Record browser actions and inspect results with Playwright CLI tooling.

by microsoft
npx skills add microsoft/playwright-cli --skill playwright-cli
Author
microsoft
Repo
GitHub

webapp-testing

Test a running web application with Playwright and capture screenshots or logs.

by anthropics Skill guide
npx skills add anthropics/skills --skill webapp-testing
Author
anthropics
Repo
GitHub
Write-up
Read →

playwright-best-practices

Improve the reliability of Playwright test suites, selectors and assertions.

by currents-dev
npx skills add currents-dev/playwright-best-practices-skill --skill playwright-best-practices
Author
currents-dev
Repo
GitHub

browser-act

Automate repeatable form and navigation tasks on websites you are authorized to use.

by browser-act
npx skills add browser-act/skills --skill browser-act
Author
browser-act
Repo
GitHub

browser-use

Guide a browser agent through browsing and data entry on permitted sites.

by browser-use
npx skills add browser-use/browser-use --skill browser-use
Author
browser-use
Repo
GitHub

What browser automation skills are for

A browser skill lets Claude drive a real browser instead of guessing what a page looks like. That distinction matters: a model reading HTML can tell you a button exists, but not that a cookie banner covers it, that its label only appears on hover, or that the form resets when you tab away. A browser skill closes the loop — act, then look at what actually happened.

Three jobs cover almost every use: filling a form that has no API, reading a dashboard that only renders client-side, and proving that a change you shipped still works after login.

Which of the six to install

The six skills on this page split into three kinds, and picking the wrong kind is the usual reason someone concludes “browser automation doesn't work”.

For a repeatable test of your own app, use webapp-testing from anthropics. It writes Playwright assertions you can run in CI, and it expects a running web application — not a production site you do not own. Pair it with playwright-best-practices from currents-dev when your suite is already flaky: that one is about selectors and waiting strategy, not about driving a browser for you.

For one-off inspection of a live page, use a CLI. Agent Browser from Vercel Labs navigates and prints a snapshot of the page; playwright-cli (microsoft) records the actions you take by hand so you can replay them. Record the flow manually once, then let the agent turn the recording into a script — that order matters, because a script written from a description is a guess.

For form and navigation tasks you are authorised to run, browser-act and browser-use sit at that level: multi-step flows on permitted sites, with the agent deciding the next action. They are the most useful and the least predictable of the six, so give them a stopping condition.

Where these skills break in practice

  • Screenshots mistaken for state. The agent reports success because the page rendered, while the record was never saved — the submit never fired, or the request failed silently. Check the saved row, not the screen.
  • Selectors that only exist after hydration. The element is in the DOM for a second with placeholder content, the click lands, nothing happens. Wait for a stable condition — an aria-label, a row count — not a fixed number of seconds.
  • Login walls. A fresh session has no cookies, so the run lands on the marketing page and confidently reports what it found there. Say which profile it should reuse.
  • Strict-mode breakage. Two elements match the same selector and the run dies halfway, usually because a mobile and a desktop header are both in the DOM.
  • Infinite scroll with no bound. Nothing tells the agent to stop. Bound it: N pages, or stop at a specific date.

The check that catches most of it

  1. Run the flow twice. The second run is where idempotency problems show up — duplicated rows, “already exists” errors.
  2. Open the app yourself afterwards and look at the record, not the log.
  3. Ask for a screenshot of every step. A run reported as successful with no screenshots is a run you cannot audit.
  4. Log out and run it again. Silent failure at the login wall is the most common way these reports go wrong.

How to evaluate the result

Use a test account and record the page state, locator and assertion for each step.

How to check the result

Record the page state, locator and assertion for every step, and re-run the check before calling it stable.

Frequently asked questions

How do I install one of these browser automation skills?

Use the command beside the skill, select a supported agent, then reload the agent and confirm it can see the skill in your project.

Which one should I use for testing my own app?

For your own application, webapp-testing writes assertions you can run in CI. agent-browser and browser-act drive a live page for inspection. If the goal is a repeatable test, use the testing skill, not the CLI.

Can these skills scrape or automate any website?

No. Driving a site you do not own or have written permission to use runs into terms of service, rate limits and legal exposure that no skill removes. Test your own staging environment instead.

How do I get a repeatable script from a manual browser session?

Install playwright-cli when you want a script built from actions you have already done by hand, and record the flow manually once first. The recorded script and the manual flow should match before you trust it.

What should I change before running a test against my live site?

Use a test account with no real customer records. Deleting a row or sending an email during a test run is not recoverable, and production data is the most common cause of an automation incident.

Do I need a browser skill if the site has an API?

Use the API for operations it supports. Use browser automation when you need to verify what a user sees or interacts with on the page.

Will it log into my accounts?

A browser skill needs an authorized test session to work behind a login. Use a dedicated test account with sample records.

What happens when the browser encounters a CAPTCHA?

Stop the run at the CAPTCHA and review the step manually. In a test environment you own, use the CAPTCHA provider's test configuration.

How do I know the automation actually worked?

Check the resulting record in the application, such as the saved row or created ticket. A screenshot shows the page rendered, and the record confirms that the action persisted.