# AncientOS

- **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:** [Nathan Greene](https://cerebralvalley.ai/u/ngreene)
- **GitHub:** https://github.com/nathandavidgreene/ancientos-hackathon/
- **Demo video:** https://youtu.be/AqqmyBX3_5E
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/566

**AncientOS: a personal AI that works on your real data without you handing over the keys**

I bookmark hundreds of posts on X and never go back to them. I wanted an agent that could actually work through them: read the posts and the images in them, label what's worth keeping, and remove what isn't. And I wanted it to do that without me giving it my accounts and hoping it behaves.

AncientOS is my answer. It's a Rust rebuild of my private, roughly 180,000-line Python assistant, designed so the safety rules are part of the code rather than promises in a prompt:

- **Every request has a data class.** Sensitive text stays on a local model, and only sanitized text goes to Meta's hosted models, always through a single router. A test enforces that sensitive data never reaches the hosted model.
- **Anything that costs money is off by default.** Paid calls run only when I turn on an explicit switch, under a spend cap. Paid execution needs an approval token, and the code won't compile without it.
- **Every action is logged in a hash chain,** so tampering breaks verification. Signing is available.
- **Sign-in is hardened** with cookie, CSRF and rate-limit checks, all covered by tests.

In a real run, Meta's models read 61 media items across 50 of my bookmarks and produced 407 findings, each tied to its source image, for 26 cents of billing. Over the whole hackathon that came to about 11,000 requests, all within the free credit. On the labels page I review those findings and click Delete, and the post comes off my X bookmarks for real, with every attempt logged.

It's an early alpha with more than 1,000 tests. The repo also lists its known gaps honestly; for example, code can still call a model without going through the router. AncientOS is model-agnostic and self-hostable, and the goal is a private intelligence that belongs only to its user.

## More from Meta Global AI Developer Hackathon

- [PulsePilot](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/563)
- [FixLine](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/564)
- [Pimp My Social](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/565)
- [Whiteout](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/567)
- [Team ProtocolSmith](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/568)
- [Plan2Cost](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/569)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/566. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt