Skip to content
Get packs

Guide · Agent skill

How to use the Dogfood web app QA agent skill

Updated · written with AI from the project's own documentation and checked against our catalogue data

The dogfood skill is an agent skill that tells an AI agent how to explore a web application like a user would, looking for bugs, UX problems and console errors. It is for developers, QA staff and product teams who want a structured first pass at finding issues in a staging or local app. After this guide you will know how to install the skill, ask for a test in plain language, and read the report it produces.

Dogfood web app QAvercel-labs/agent-browser · Apache-2.0

Quick start

  1. 1Download and unzip. Download the dogfood ZIP from this page and unzip it into ~/.claude/skills/ or a project's .claude/skills/ folder.
  2. 2Or upload in the Claude apps. Go to Settings → Capabilities → Skills and upload the skill there.
  3. 3Install agent-browser. Make sure the agent-browser CLI from vercel-labs/agent-browser is installed and runs as the agent-browser command.
  4. 4Confirm a safe target. Pick a staging or local app you own or have permission to test, and gather login credentials if needed.
  5. 5Send a natural request. Ask for example: dogfood http://localhost:3000, adding a scope or login details if relevant.
  6. 6Review the report. Open report.md in the output folder and check the screenshots and videos referenced in each ISSUE block.

What the dogfood skill makes the agent do

An agent skill is a folder of instructions and supporting files that an AI agent loads when a request matches its description. The dogfood skill covers systematic exploration of a web app. The agent opens the app in a controlled browser, works through the main navigation, tests interactive elements such as buttons, forms and dropdowns, and checks edge cases like empty states and error handling.

For every issue it finds, the agent records reproduction evidence. Interactive bugs get a repro video and step-by-step screenshots. Visible-on-load problems such as typos or clipped text get a single annotated screenshot. Everything goes into a structured report, so the findings can be handed to the responsible team.

  • Severity levels: critical, high, medium and low.
  • Categories: visual, functional, ux, content, performance, console and accessibility.
  • Output: a report.md file plus screenshots and videos in an output folder.

Install the dogfood skill

Download the dogfood ZIP from this page and unzip it into one of two locations. For a personal install available across projects, unzip it into ~/.claude/skills/. For a single project, unzip it into that project's .claude/skills/ folder.

Alternatively, you can upload the skill in the Claude apps under Settings → Capabilities → Skills.

  • Personal: ~/.claude/skills/
  • Project only: .claude/skills/ inside your project
  • Claude apps: Settings → Capabilities → Skills, then upload

Install the agent-browser CLI

The skill requires the agent-browser CLI from vercel-labs/agent-browser. It must be installed and available as the agent-browser command. The skill is written to call the agent-browser binary directly rather than through npx, because the direct binary is faster.

Check that the command is available in your terminal before starting a session.

  • agent-browser must be on your PATH as the agent-browser command.
  • Have login credentials ready if the target app requires authentication.

Trigger the skill with a natural request

You do not need special syntax. The skill's description lists trigger phrases such as "dogfood", "QA", "exploratory test", "find issues", "bug hunt" and "test this app/site/platform". Include the URL and, if needed, the login details and a scope.

The only required input is the target URL. If you do not give a session name or output directory, the defaults are used: the session is named after the slugified domain, and output goes to ./dogfood-output/.

  • "Dogfood vercel.com" starts immediately with defaults.
  • "QA http://localhost:3000 and focus on the billing page" limits the scope.
  • "Sign in to user@example.com and bug hunt the dashboard" adds authentication.

Read the report the agent produces

The report is built from a template with a header table, a summary of severity counts, and one ISSUE block per finding. Each block lists severity, category, URL, the repro video path or N/A, a description, and numbered repro steps that reference screenshots.

The skill aims for 5 to 10 well-documented issues. The agent is told that depth of evidence matters more than total count.

  • Each issue has a short title such as ISSUE-001.
  • Static issues set Repro Video to N/A.
agent-browser --session {SESSION} open {TARGET_URL}

Safety notes before you run it

The skill runs scripts and uses the network. It logs in, clicks forms and can create, edit or delete data. Only test sites you own or have permission to test. Do not point it at production apps with real customer data.

The skill does not cover desktop or mobile native apps. For OTP or email login codes, the agent asks you for the code.

  • Use staging or local apps where creating or deleting records is safe.
  • Only test sites you own or have permission to test.
  • Skill covers web apps only, not native desktop or mobile software.

What people use it for

  • Finding bugs in a staging build before a release.
  • Building a structured bug list with reproducible evidence for a product team.
  • Having an agent log in to a local app and walk through core flows such as sign-up or checkout.
  • Checking a single area of an app, such as the billing page, by setting a scope.

Frequently asked questions

Is the dogfood skill free?

The source material for this page does not state licence pricing beyond the licence of the skill folder, which is Apache-2.0. The agent-browser CLI is a separate project from vercel-labs/agent-browser, so check its own terms before you install it.

What do I need to run the dogfood skill?

You need the agent-browser CLI installed and available as the agent-browser command, plus login credentials if the target app requires authentication. The skill also runs scripts and uses the network.

Where do I install the dogfood skill?

Unzip the downloaded ZIP into ~/.claude/skills/ for personal use or into a project's .claude/skills/ folder for one project. You can also upload it in the Claude apps under Settings → Capabilities → Skills.

Can the dogfood skill test production sites?

It is not recommended. The skill logs in, clicks forms and can create or delete data, so it is intended for sites you own or have permission to test, such as staging or local apps.

Does dogfood only work on websites?

It is for web applications. The source says it does not cover non-web software such as desktop or mobile native apps.

What is the output of a dogfood run?

A structured report.md built from the template, with a severity summary and one block per issue. Screenshots and repro videos are saved in the output folder's screenshots and videos subfolders.

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 dogfood if you have a staging or local web app you own and want a structured first pass with reproducible evidence for each bug. Pick something else if you need to test production systems with real customer data, or native desktop or mobile apps.

Download the skill