Show HN: Gretel – Topic specific feeds built from YouTube recommendations

A developer released Gretel, a beta desktop app that builds topic-specific YouTube feeds by creating separate profiles for different interests without requiring additional Google accounts. Gretel uses YT.js to access YouTube, runs locally so requests originate from the user's machine, searches YouTube for each profile's topics, collects candidate videos, and generates embeddings to filter unrelated content while continuously updating topic embeddings as videos are watched. The app supports sign-in with Google for included embedding credits or a user-supplied OpenRouter key, with other providers and local embedding models planned.

One of the annoying things about YouTube is that it is really good at catching my attention, but I have little control over what kind of feed it can show me, so it becomes distracting instead. I don't want to hide recommendations entirely though, I just don't want all my interests mixed into one feed. So I wondered if I could have separate YouTube profiles for different interests, but make creating one as simple as picking a few topics instead of having to create or connect another Google account. So I built Gretel, a desktop app. It uses YT.js to access YouTube, and it runs locally so the requests come from my own machine rather than a server. I create profiles, each with their own topics. Gretel then searches YouTube for the topics, collects candidate videos, and creates an embedding that is used to filter videos later on. From the candidate videos, Gretel fetches related videos and continuously expands outward while filtering unrelated videos. This creates the base feed. After that, as I watch videos, the topic embeddings are continuously updated, and at the same time Gretel persists the YouTube sessions so the recommendations adapt as well. I have been working on Gretel on and off for a while. It's beta: usable but still rough. I am curious if, after using a profile for a while, the feed actually starts converging toward what you wanted. Bug reports and contributions are welcome too. You can sign in with Google to use the included embedding credits, or bring your own OpenRouter key. Other providers and local embedding models are coming. Comments URL: https://news.ycombinator.com/item?id=49982629 https://news.ycombinator.com/item?id=49982629 Points: 2 Comments: 0