# jqwoutt team

- **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:** [Ivan Golovachev](https://cerebralvalley.ai/u/jqwoutt)
- **GitHub:** https://github.com/jQwout/amuse-pocket
- **Demo video:** https://youtu.be/o74D9q7xLxg
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3699

Plenty of AI chatbots. But where's the voice assistant that runs agents you build yourself?

AI Open Pocket is an everyday Android assistant on Meta's models. You call it like a person, show it a photo mid-call, or hold Home over any app, and it answers out loud. In Auto mode it hands each request to the right agent: a ready-made one (Budget, Tutor, Task list, Reminders, Diary) or one you built yourself.

Building an agent takes no forms and no code. You describe it in plain words, and a builder agent on Muse Spark asks what it needs, a few options at a time, while the agent's passport fills in. You review the whole agent on one card and publish it; from then on the assistant can hand requests to it. Agents run on the phone with real tools (reminders, alarms, calendar, files, memory, cards, web search, photo outlines) and always ask before deleting anything or sending anything off the device.

Short interactive lessons teach the theory along the way: what Meta's models can do and how an agent is built. Mistakes are part of it: every answer explains why, and you can always try again.

Who it's for: people who keep hearing about AI agents but don't know where to start, or feel too shy to ask.

Built during the hackathon (Oct 27 - Nov 4): Meta Model API only, on the person's own key (no key ships in the app); the voice call with Muse Voice Transcribe and photos mid-call; the assistant over any app; Auto routing between agents; the builder chat with the agent passport; apps of their own for Budget, Tutor and Task list; SAM 3.1 as an agent tool; Meta's web_search with sources; Muse Image pictures in replies; the whole course in English. The lesson engine and the first lessons about Meta's models existed before the build window (up to Oct 26); the git history shows what was built when.

Notes for judges
- Try it: clone the repo, run the app configuration in Android Studio (Android 7.0+), and paste your own Meta Model API key in onboarding or Settings. For English: Settings → Language → English.
- Agent screens from the cloud (not in the video): in Settings → Cloud-forged screens enter https://amuse-forge.fly.dev and the token fxVPs7grdtVZaC-tA_LYi8w06F6rdGH7us1HW1--eBk, then on an agent's card pick "Own interface · beta" and describe the screen. A coding agent on Muse Spark writes it, and the app renders it without an update, on your own Meta key.
- Code: agent/loop/ReactLoopRunner.kt, agent/core/agent/BuilderAgent.kt and AgentRouter.kt, ui/call/CallScreen.kt, agent/tools/PhotoOutlineTool.kt.
- Limits: Meta's API has no text-to-speech, so spoken replies use the phone's own engine; Voice Transcribe is used for English. Waiting for the model is sped up in the video.

## More from Meta Global AI Developer Hackathon

- [Valentin Englert - lwnzhn.com](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3696)
- [Diligent 619](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3697)
- [Winning Startup](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3698)
- [Unpile](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3700)
- [Jatt Builds](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3701)
- [Witness](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3702)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3699. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt