Skills and scripts that let a coding agent do website QA the way a senior tester would: plan the run, test in a real browser, keep evidence, rate what it finds on a fixed scale and give a verdict by rule.
It is for teams who already build with an AI coding agent and want its "I checked it, looks fine" replaced by a report they can act on.
| Skill | Use it for |
|---|---|
qa-standards |
Any QA job. The test charter, severity S1-S4, priority P1-P4, defect format, evidence rules and verdict rules. |
qa-site-audit |
A read-only audit of a whole site: layout at five widths, accessibility, performance, SEO, content, cookie consent and tracking, leftover debug code. |
qa-release |
A pre-release programme for a web product: every route and role in a coverage matrix, a defect register and a release verdict. |
The site audit comes with headless check scripts you can also run on their own. See docs/scripts.md, and examples/sample-report for what a finished report looks like.
The method is the same everywhere. An edition says who drives the browser.
| Edition | Who drives the browser | Good for | Verdict |
|---|---|---|---|
| Local | Your own coding agent, on your machine | Quick checks while you build, including localhost | Preflight only |
| Hosted QA bot | A bot on a hosted machine | Hands-off checks of public or preview URLs | Full |
| Dedicated QA Mac | A separate Mac with a real desktop browser | Release sign-off and acceptance | Full |
Not sure? Start with Local. docs/choosing-an-edition.md has the longer comparison.
Paste this to your coding agent:
Install the Agentic QA Toolkit from https://github.com/code23/agentic-qa-toolkit for me.
Clone it, read AGENTS.md in the repository and follow its install steps.
I want the Local edition. Show me the dry run before you apply anything.
Change "Local" to "Hosted QA bot" or "Dedicated QA Mac" if that is your setup. The agent finds its own skills folder, shows you what it will copy, installs after you agree, and checks the result. AGENTS.md is the script it follows; it is short enough to read first.
git clone https://github.com/code23/agentic-qa-toolkit.git
cd agentic-qa-toolkit
./install.sh --edition local --target ~/.claude/skills # shows what it would do
./install.sh --edition local --target ~/.claude/skills --apply # installsUse your own agent's skills folder as the target.
Ask your agent:
Use qa-standards to QA the sign-up flow on https://staging.example.com
You get a charter, a walk through the flow, findings in a fixed format with the screenshot for each one shown beside it, and a verdict line. Reports export to a self-contained HTML file or PDF with the images inside. To try the audit scripts with no site of your own, run them against the demo site, which has faults planted on purpose.
Prefer a download? Each release on the repository's Releases page has one zip per edition, with checksums.
- Read-only by default. The agent looks, clicks through and opens, but does not submit, pay, send, invite or delete. Forms are typed into with test data to check validation and never submitted.
- It stops at a payment form, a phone code, an identity check or a CAPTCHA.
- Test accounts only. Never your own or a customer's.
- Nothing is filed or sent without your approval.
- Only test sites you own or have permission to test. See docs/responsible-use.md.
- An AI coding agent that supports skills (folders with a
SKILL.md). - For the scripts: Python 3.9+ with the packages in the two
requirements.txtfiles underskills/, Node 22.12+ withpuppeteer-core25, and Chrome or Chromium. - macOS or Linux. The Dedicated QA Mac edition needs a Mac for the worker.
skills/ the three skills, the same in every edition
editions/ one folder per edition: its README, environment note and extra tools
demo-site/ a made-up site with planted faults, for trying the scripts and for tests
examples/ a sample report
docs/ guides
tests/ offline tests for every script, the installer and the package build
tools/ package build, link check and the one-command check script
install.sh installs one edition into your agent's skills folder
AGENTS.md what an AI agent should do when asked to install or change this toolkit
- Choosing an edition
- How a run works
- The check scripts
- Branding your reports
- Frequently asked questions
- Responsible use
- Maintaining and releasing
How many tokens does it use? In our runs, a light QA of one to three flows takes 15-30 minutes and about 130,000 fresh tokens on the executor; a deep site audit takes 2-6 hours and about 1.3 million. Most usage is cached context, which costs far less. The full figures.
Do I need a Mac mini? No. Only the dedicated QA Mac edition needs a second machine, and any Mac that runs the ChatGPT desktop app will do. The local edition runs on your own computer.
Does it work on Linux or Windows? Linux, yes, except the dedicated QA Mac worker. Windows is not tested.
Does it work with only Claude Code, or only Codex? Yes, either can be the QA lead. The dedicated QA Mac worker uses the ChatGPT app, so it needs a ChatGPT account.
Do I need an API key? No. A normal subscription to your coding agent is enough, and the check scripts use no AI.
Can it break the site I test? It is built not to: it never submits, pays, sends or deletes, and stops at payment forms and CAPTCHAs. Test only sites you own or may test.
More in the FAQ.
Issues are welcome: bugs, false positives, misses and ideas. Pull requests are not accepted at the moment. See CONTRIBUTING.md. Security problems go through SECURITY.md, not a public issue.
MIT.