# Bean

- **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:** [Rayan Banni](https://cerebralvalley.ai/u/RayanBn)
- **GitHub:** https://github.com/RayanBn/Bean
- **Demo video:** https://youtu.be/pak68r1TycU
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1772

Bean is AI that works for you before you ask. It lives in your Mac's notch, runs a team of agents, and you never open a chat.

Every AI tool today waits for a prompt. Bean watches your Mac live: a window opens, you switch apps or desktops, a call starts. In about 60 ms a local model decides if it matters. If it does, Muse Spark picks the right agent, or creates one, places it inside the right app, and Bean reports back once.

Why this wasn't possible before:
- Always-on AI was too slow, too costly and too invasive: every event would hit the cloud. Strands, a 2B model on Apple silicon, now filters each event locally in ~60 ms, so only real work reaches the cloud.
- Agents that "use your computer" take over your mouse and screen. Bean's agents work through each app's own scripting and Accessibility, in the background, while you keep working.
- No model could run a whole job alone. Muse Spark now plans and calls tools over many steps, Muse Voice hears the call, Muse Code writes the code: one family, one API, zero clicks from "call ended" to "draft PRs open".

What it does:
- Any work: one sentence creates an agent. Calls to tickets, code and PRs, quotes filled from live hotel pages, overdue invoices chased. No code per use case.
- Any app, any service: agents sit in each app's toolbar, and any remote MCP server is a one-click connector (Notion, Linear, GitHub, Sentry, Vercel, Atlassian...).
- A real team: each agent has its own page with plan, changes with Undo, live terminal, files, memory and "while you were away". It learns from your corrections.
- Private: can run on a local model; off-limits apps are never read.
- Safe: Bean's code, not the model, asks before anything new, outward or irreversible. Merging is never automatic.

Native Swift app, built during the hackathon.

## More from Meta Global AI Developer Hackathon

- [Tomotaka Matsumoto](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1769)
- [Augustine Alfred](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1770)
- [Agent Beans](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1771)
- [DevDiva](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1773)
- [ShortEditor](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1774)
- [Chatfront](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1775)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1772. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt