# Probei

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Aibek Kenzhebayev](https://cerebralvalley.ai/u/aibekken)
- **GitHub:** https://github.com/AibekKen/probei
- **Demo video:** https://youtu.be/JIIAqpIt0Yw
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1006

Probei is a visual QA agent that tests a web app the way a real user does. Automated UI tests break on every redesign: rename a class or move a button and a hundred tests go red while the product still works. Probei has no selectors, so there is nothing to break.

How it works: you write a scenario in plain English (what to do and what should happen). Probei opens the app in a browser and takes a screenshot; Muse Spark looks at the screenshot and the history and returns one action (click, type, press a key, scroll or wait); Playwright performs it; the loop repeats until the scenario passes, a bug is found, or the agent gets stuck. The result is a bug report with steps to reproduce, screenshots, page errors and a video of the run (model thinking time is cut out, so a run plays in seconds).

Features: auto-review that crawls the app and drafts test scenarios; a web UI with projects, saved scenarios, a run-all queue and run history; one-click export of a bug report to a GitHub issue; deploy watch that can rerun all tests after a new deploy; voice input for scenarios; and a CLI for scripts and CI.

Results: we built a demo shop with planted bugs and a ground-truth file the agent never sees. Probei found 2 of 3 bugs with zero false alarms, at about 40 seconds per scenario. Then we fully restyled the shop (new layout, theme, DOM and class names). A classic selector-based Playwright check went from PASS to FAIL, while Probei ran the same scenarios unchanged and still found bugs. Small sample, so scores vary between runs.

Stack: TypeScript, Playwright, Muse Spark via the Meta Model API or Baseten (open weights).

## More from Meta Global AI Developer Hackathon

- [Dastan](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1003)
- [See the Shake](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1004)
- [Breadcrumbs](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1005)
- [GEONOS](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1007)
- [ClearLine](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1008)
- [Setecho](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1009)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1006. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt