Guide · Agent skill
How to use the Agent Browser agent skill: setup and examples
Updated · written with AI from the project's own documentation and checked against our catalogue data
The agent-browser skill lets an AI agent open a real browser, look at a web page or app, and report what it sees. It is written for checking behaviour: confirming a UI change, reproducing a bug with console and network evidence, or walking through a short flow. This guide explains what the skill does, how to install it, how to ask for it, and the safety rules it sets, so you can use it without guessing.
Agent Browserpaperclipai/paperclip · MIT Download the skill Quick start
- 1Download the ZIP. Download the agent-browser skill ZIP from this page.
- 2Unzip into a skills folder. Extract the agent-browser folder into ~/.claude/skills/ or a project's .claude/skills/ folder.
- 3Or upload in the Claude apps. In the Claude apps, open Settings, go to Capabilities, then Skills, and upload the skill.
- 4Check your browser setup. Make sure a browser automation library such as Puppeteer or Playwright and a runtime that can launch a browser are available.
- 5Describe the verification task. Name the URL, what should be true after loading, and the evidence you want, such as a screenshot or console errors.
- 6Review the evidence. Check the screenshots, filtered console output and any non-2xx network responses against the narration the agent gives.
What the agent-browser skill makes the agent do
An agent skill is a folder of instructions that an AI agent loads when a task matches it. The agent-browser skill, defined in a file called SKILL.md, tells the agent how to drive a controlled browser for verification work. Once installed, the agent can navigate to a page, take screenshots, read JavaScript console errors and network responses, and fill simple forms or click through a short flow.
The skill is aimed at pages that plain HTTP fetches cannot read well, such as single-page apps and dynamically rendered content. It is not meant for bulk data collection.
- Open a page in a controlled browser and take full-page or viewport screenshots
- Read console warnings and errors
- Inspect network requests, including non-2xx responses
- Walk through a short flow such as load, click and observe
- Record evidence and a short narration of each step
Install the agent-browser skill from the ZIP
Download the ZIP for the agent-browser skill from this page. Unzip it so the agent-browser folder containing SKILL.md sits in one of these locations: ~/.claude/skills/ for your user account, or .claude/skills/ inside a specific project if you want the skill only for that project.
If you use the Claude apps instead, open Settings, go to Capabilities, then Skills, and upload the skill there.
- Personal install: place the folder in ~/.claude/skills/
- Project install: place the folder in the project's .claude/skills/
- Claude apps: Settings → Capabilities → Skills, then upload
Set up a browser library and runtime
The skill's tags list Puppeteer and Playwright as browser automation libraries, and the skill needs a runtime that can launch a browser. The skill file does not specify the exact setup, so check that your environment has one of these libraries and a browser it can launch before you start a session.
The skill runs scripts and uses the network, so it should only be used in an environment where you are comfortable with both.
- A browser automation library such as Puppeteer or Playwright
- A runtime able to launch a browser
- Credentials only if a login-gated page is being tested, supplied explicitly for that run
How to trigger the skill with natural requests
You do not need a special command. Describe the verification task in plain language and name the URL. The agent should pick up the skill when the request matches its purpose: checking a page, reproducing a UI bug, or reading JavaScript-rendered content.
Good requests say what should be true after loading, what evidence you want, and which viewport matters. Ask for the console and network checks explicitly if you need them.
- "Open http://localhost:3000/pricing and screenshot the header to confirm the new banner shows."
- "A user reports the checkout button is unclickable on mobile. Reproduce it at 390×844 and capture console errors."
- "Load the dashboard, click Refresh once, and tell me if any requests return 4xx or 5xx."
What evidence the agent should record
For a verification task the skill expects a screenshot of each meaningful state, a console log filtered to warnings and errors, any non-2xx network response with its URL, status and a short body excerpt, and a short narration of what was navigated, observed and clicked. For a bug reproduction, it also asks for exact steps, viewport size, and whether the bug appears on first load or only after interaction.
Ask for this evidence up front so you can check the agent's claims rather than trusting a single screenshot.
- Screenshots of each state, with the viewport noted
- Filtered console warnings and errors
- Non-2xx responses with URL, status and body excerpt
- Narration: navigated to X, observed Y, clicked Z, observed W
Safety rules: scripts, network and login-gated pages
The skill runs scripts and uses the network, so treat it like any tool that touches live systems. It prefers static fetching when a page is reachable as plain HTML, and it avoids login-gated or sensitive sites unless you supply credentials explicitly for that run. Credentials should not be stored in the session log, screenshots or returned output.
The skill also sets politeness limits: throttle navigations on shared infrastructure, respect robots.txt for public sites, cancel navigations that exceed a reasonable timeout such as 30 seconds, and retry once with a longer timeout before escalating. It advises against banking, healthcare and government accounts, where automation can cause lockout or compliance problems.
- Prefer API tokens or magic links over UI login
- Never reuse credentials outside the session
- Do not use it for unattended large-scale scraping
- Skip sites whose terms prohibit automation
Tips for reliable results
Most failures come from timing or page structure rather than the browser itself. The skill recommends waiting for a specific selector or network-idle state instead of fixed sleeps, and taking a screenshot when a selector is not found so you can see the real page state.
Clicks that do nothing are often caused by an offscreen element, a modal covering it, or a shadow DOM. Cross-origin iframes you do not own cannot be inspected. Remember that a passing screenshot only proves the viewports you actually tested.
- Confirm the server is running and the port is correct for local dev servers
- Use a desktop viewport by default; switch to mobile only for mobile-specific tasks
- Some sites detect headless Chrome; the skill says to use non-headless mode only when authorised
- Close the browser cleanly when finished
What people use it for
- Confirming a UI change on a deployed page or local dev server
- Reproducing a reported UI bug with console and network evidence
- Reading JavaScript-rendered content with a person supervising the session
- Walking through a short click-and-observe flow to check acceptance criteria
Frequently asked questions
How do I install the agent-browser skill?
Download the ZIP from this page and unzip it so the agent-browser folder sits in ~/.claude/skills/ for personal use, or in a project's .claude/skills/ folder. In the Claude apps you can instead go to Settings, then Capabilities, then Skills, and upload it.
What does the agent-browser skill require?
It needs a browser automation library such as Puppeteer or Playwright and a runtime that can launch a browser. The skill file does not specify the exact setup. Credentials are only needed if a login-gated page is tested, and they should be supplied for that run only.
Is the agent-browser skill free?
The source is listed under the MIT licence. This page does not state pricing for any other service, so check the licence file in the repository for the exact terms.
Can I use agent-browser for web scraping?
Not for unattended large-scale scraping. The skill says that belongs to a dedicated scraper with rate limits and robots.txt handling. It is intended for supervised verification of how a page behaves.
Why does the skill prefer static fetching?
If a page is reachable as plain HTML, a static fetch with curl or an HTTP client is cheaper, faster and more reliable. The browser is for pages that a static fetch cannot read, such as JavaScript-rendered content.
What should I do if clicks do nothing or selectors are not found?
Take a screenshot to see the actual page state. The element may be offscreen, covered by a modal, or inside a shadow DOM, so the agent should scroll it into view or pierce the shadow root. Adjust the selector if the page has changed.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Use agent-browser if you are an engineer, QA tester or researcher who needs a supervised browser to confirm how a page behaves, with screenshots, console errors and network evidence. Pick a dedicated scraper for bulk collection, and avoid it for banking, healthcare, government or any login you are not authorised to automate.
Download the skill