# Neuron

- **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:** [Tanay Kapoor](https://cerebralvalley.ai/u/imtanay)
- **GitHub:** https://github.com/NeuronBots/neuron-hackathon
- **Demo video:** https://www.youtube.com/watch?v=snoBPcwI6kQ
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3326

Neuron is a workspace where you run a project with a team of AI bots instead of a single chat. Describe what you want, e.g. a one-page shop for a terrarium studio. A Master bot sets up the team, plans with you, builds the first version and keeps going. Anything that leaves the app, such as publishing a site, waits behind an Allow/Deny card.

BEFORE vs THIS WEEK
Neuron already had projects, bot teams and chat before the hackathon baseline (commit e0703a5, Oct 27). Everything in this submission and the demo was built after it: [79] commits by Tanay Kapoor and Ayushi Gupta, [216] files changed.

BUILT THIS WEEK
- Plan, approve, build: the Master bot writes a plan stating which bot owns what, and nothing builds until the owner approves. Meta Muse then builds the site in a sandbox with a live preview.
- Hosted sites that take orders, with order routines set up by the bots.
- Payments and Wallet: the owner turns on Razorpay checkout, sets prices, sees orders and approves refunds.
- Allow/Deny approval cards for any action that leaves the app.
- Versioned bot memory (draft, production, rollback).
- Meta Muse model layer: provider in our proxy, per-bot tier routing, retries and back-off, usage limits.
- Rebuilt PR review bot, shared documents, voice call billing.

HOW WE USE META MUSE
All model calls go through one proxy running on Meta Muse: muse-spark-1.3 for the fast, workhorse and judgement tiers (they differ by thinking effort), muse-wonder-1.0 for reasoning only, and muse-image-1.0 for site images. Builds run in an isolated sandbox that calls the Meta API directly. We moved structured setup calls to the fast tier so Muse does not spend its allowance thinking.

WHAT WE MEASURED
From our own dev logs (one machine, 452 calls on Claude vs 3,013 on Muse, chat turns only): the median model call fell from about 8,100 to about 4,700 fresh tokens, and cache hits rose from 51% to 61% of prompt tokens. Caveats: whole prompts are the same size, Muse writes longer answers, and we measured token counts only, not quality, speed or price.

TRY IT
Live: https://beta.neuronbot.in/

Demo Account
judge@neuronbot.in; 6gme-txnh-9nsx-n2wx

## More from Meta Global AI Developer Hackathon

- [BlueOcean SP](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3323)
- [Warp Relay](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3324)
- [Clever Monkey](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3325)
- [Apprentice - Taran Pal Singh](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3327)
- [Crosscheck AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3328)
- [Speed&Inertia](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3329)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3326. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt