QA agent for web apps. It hooks into Claude Code as a skill: you type
/qa-bot test http://localhost:3000, the agent drives a real browser through
the app, pokes at forms, checks the mobile layout, hits the API directly, and
leaves behind a report folder with screenshots, screen recordings and a log of
every API call it made.
There's no test script to write or maintain. The agent decides what to test
the same way a QA engineer would, and the qab CLI gives it fast hands on a
real browser. Chromium, Chrome and Firefox, on Linux, macOS and Windows.
- Node 20+
- Claude Code
- Google Chrome, if you want to test against real Chrome (
qab engine chrome). Chromium and Firefox come bundled through Playwright, nothing to install.
git clone https://github.com/Bij4n/qa-bot.git
cd qa-bot
npm install
npx playwright install chromium firefox
npm run build
npm link # puts `qab` on your PATHOn a fresh Linux box you may need the browser system deps as well:
npx playwright install --with-deps chromium firefox.
Then register the skill so Claude Code picks it up:
# macOS / Linux
ln -s "$(pwd)/skill" ~/.claude/skills/qa-bot# Windows
New-Item -ItemType Junction -Path "$env:USERPROFILE\.claude\skills\qa-bot" -Target "$(Get-Location)\skill"That's the whole install. Open Claude Code in whatever project you want tested and run:
/qa-bot test http://localhost:3000
The agent will ask how deep to go (quick / standard / exhaustive), work through the app, and hand you the report path when it's done.
Every run creates a folder under reports/ in the project you ran it from,
organized by what was being tested:
reports/2026-07-15-14-30-myapp/
├── report.md health score, findings by severity, repro steps
├── frontend/ desktop screenshots, numbered in test order
├── mobile/ mobile layout screenshots (375x812)
├── backend/
│ └── api-log.jsonl every API call: status, latency, body preview
└── recordings/
├── frontend-chromium.webm one screen recording per phase, per engine
├── mobile-chromium.webm
└── frontend-firefox.webm
report.md scores the app out of 10 and ranks findings critical / high /
medium / cosmetic. Each finding comes with repro steps and a screenshot,
because a bug report you can't reproduce is worthless.
The agent does everything through the qab CLI, and you can too. A
background daemon keeps the browser warm so commands come back in ~100ms:
qab session start myapp # new report folder, recording starts
qab goto http://localhost:3000 # reports console errors + failed requests
qab snapshot # accessibility-tree view of the page
qab click "text=Sign up"
qab fill "#email" user@test.dev
qab shot signup # full-page screenshot -> frontend/
qab phase mobile # switch to 375x812, new recording
qab phase backend
qab api POST http://localhost:3000/api/users '{"email":""}'
qab engine firefox # rerun anything in firefox
qab session end # finalizes the last recording
qab stop # kill the daemonqab help lists everything. Daemon state and logs live in ~/.qa-bot/.
If the app needs an account, the agent will ask you for test credentials and log in through the form like a user would. Use a staging account. Don't point it at production with real customer data.
There's nothing to deploy server-side. Each developer runs the setup above once on their machine and updates with:
git pull && npm run buildIf the daemon acts weird after an update, qab stop and try again.
- daemon failed to start — check
~/.qa-bot/daemon.log - want to watch it work — everything is headless by default;
qab engine chromium headedopens a visible window qab engine chromefails — Chrome isn't installed on that machine. Chromium is the same engine and always available.- recordings look letterboxed — don't resize with
qab viewportduring a session; useqab phase mobile, which starts a fresh recording at the right size