Quickstart

Point comber at a preview URL and it explores your app like a real user, then hands you the broken click with a repro. No test scripts to write.

1 · Run it against any URL

Bring your own Anthropic (or OpenAI) key and point comber at a running app you own:

# set your key once export ANTHROPIC_API_KEY=sk-ant-... # crawl a preview deploy and report what broke npx comber check https://pr-482.your-app.vercel.app

comber walks every reachable state, watches the console, network, and accessibility tree, and prints findings with a repro, a screenshot, and a trace. The expensive run happens once — it leaves cheap deterministic replays behind.

2 · Run it in CI on every preview deploy

Drop a step into your pipeline so every preview deploy gets checked before it merges. GitHub Actions example:

name: comber on: [deployment_status] jobs: qa: if: github.event.deployment_status.state == 'success' runs-on: ubuntu-latest steps: - run: npx comber check "${{ github.event.deployment_status.target_url }}" env: ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}

A dedicated GitHub Action and a Vercel integration are on the way — for now the one-line npx step runs comber in any CI.

3 · Or let us run it

Don't want to manage keys or CI? On the Managed plan you type a URL, click Run, and we crawl it on our infra — pick your model, pay per run. Findings land in a hosted dashboard with history and baselines.

What a finding looks like

✗ /convert 500 · after "Upload" fail · server-error POST /api/convert → 500 repro: "Get started" → upload sample.pdf → 500 screenshot + trace saved → r/8fa2c1

Every finding is a real reproduction, not a flaky assertion — the exact steps, the request that failed, and a trace you can replay.

Only test what you own

comber exercises real controls in the app you point it at. Only run it against applications you own or are explicitly authorized to test — see the Terms. For untrusted or shared targets, use comber's strict, write-gated mode.

Questions? Support · support@codecomber.io