A test for every bug
The path is shortened to the fewest steps that still fail, then written as a standalone spec.
Autonomous exploratory QA
Finds the bugs your test suite never visits.
Nomad explores a web app the way a curious user would. Every move is chosen by a three-tier decision cascade, every anomaly is replayed three times to confirm it, and every confirmed bug comes back as a minimal Playwright test.
npm run nomad -- scan http://localhost:4100 --steps 45- TypeScript
- Playwright
- NestJS
- PyTorch → ONNX
- Qwen2.5 on Ollama
- Next.js
One loop, run once per step.
Each step observes the page, decides what to try, acts, checks for anything broken and reports what it can prove.
- 01Observe
Fingerprints the page by structure, visible text and URL pattern, so a revisit is recognised as the same state.
state graph - 02Decide
Rules take the obvious calls. A trained ranker scores the rest. A local model settles what the ranker is unsure of.
rules → ranker → LLM - 03Act
Clicks, fills and submits through Playwright, and finds targets by meaning when their selectors drift.
self-healing locator - 04Detect
Watches console, network, rendering and copy, then folds related signals into a single finding.
9 detectors - 05Report
Shortens the path, replays it three times and writes a Playwright test for every confirmed bug.
replay ×3 → .spec.ts
Evidence you can act on.
Selectors that survive a redesign
When an element disappears, candidates are scored by meaning, role and position on the page.
Regression sweeps
Replay every bug of a run against another build and get a verdict for each: fixed, still broken or flaky.
Point it at a URL.
Runs locally with Node 20. The ranker and the local model are optional.
- Install and build
npm install && npm run build - Scan an app
npm run nomad -- scan <url> --steps 60 - Gate CI on it
--fail-on major --min-recall 1 - Or run everything
docker compose up --build