# Harbor

- **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:** [Tymon Zarski](https://cerebralvalley.ai/u/tymzar)
- **GitHub:** https://github.com/tymzar/harbor
- **Demo video:** https://youtu.be/0_bqs513JHs
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4837

Content is personalized everywhere: feeds, recommendations, ads, and now LLM answers are tuned to each of us. The software itself is not. Every site still ships one layout for everyone, and the only way to change it is to write CSS selectors by hand. Harbor closes that gap: a Chrome side-panel extension that reshapes any web page around how you use it, with Meta's Muse Spark 1.3 doing the planning. Set up your Harbor once with three picks: the page shape you'd rather read, what you reach for first, and what to tidy (ads, pop-ups, "more like this", sidebar extras, comments). Then click Harbor's icon on any site. Muse Spark reads the page's element list plus your picks as short rules and returns a strict JSON layout plan: hide, fold behind a "Show…" button, compact, grid, emphasize. Every change says why ("because your harbor: sidebar extras folded", or "Muse Spark's own idea: …"). Preview outlines the changes on the real page; untick anything; Keep.

What you keep comes back by itself: a saved layout is matched by structure, not URL, so a change to a site's top bar follows you from the feed to a messaging thread, with no model call. On a site you follow, a page of a new kind gets a suggestion drafted automatically, at most once per page shape per day. You can also ask in your own words ("I only read the feed: hide both sidebars"), and answer Harbor's follow-up questions, which adjust the plan locally. An opt-in, on-device model counts which items you open and, after eight clicks on a row of similar items, offers your own order (e.g., Messaging, Jobs, Home), applied visually only.

Harbor never touches cookies, consent or sign-in prompts, a page's main video, or forms you are filling in; form contents are never read or sent. Restore puts the page back exactly as the site shipped it.

It is tested, not just demoed: an evaluation set of 54 scenarios on 20 popular live sites (YouTube, GitHub, Wikipedia, Amazon, Spotify, …) with 124 machine-checkable expectations, including 52 "keep" checks that catch over-hiding. In the last round, Harbor met 93% of expectations and fully completed 84% of scenarios with no over-hiding, against 42% and 0% for leaving the page unchanged. Region-level fingerprints re-find saved layouts after a page changes (96.2% correct, 0.9% wrong on held-out changes).

## More from Meta Global AI Developer Hackathon

- [Tini](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4834)
- [Spidey Sense](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4835)
- [Jazzy Fox](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4836)
- [Jehsh Philemon Mulwahali](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4838)
- [OfferLetter](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4839)
- [Pyx](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4840)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4837. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt