Winnow 作者： Science

Winnow


Your YouTube feed, winnowed down to what you actually want to watch. AI scores your subscriptions and recommendations against your own interest profile — no autoplay, no infinite scroll, no engagement bait. Bring your own AI key; everything runs in your browser.


YouTube's algorithm optimizes for minutes watched. Winnow optimizes for something different: videos you actually want to watch and feel enriched by afterward.

How it works

1. Winnow reads your real YouTube data — your subscriptions feed and homepage recommendations — using your own logged-in browser session. No OAuth setup, no YouTube API keys, no quota.
2. Each video is scored by an AI model (Anthropic or OpenAI — you bring your own API key) against your free-text interest profile: what you want more of, what you want less of. When a transcript is available it's weighed heavily — it reveals whether the content delivers on the title's promise.
3. You get a calm, bounded feed in tiers: Top picks, Worth a look, and a collapsed Winnowed out fold. Nothing is deleted — every filtered video stays one click away, with the reason it was filtered, so the curation is always auditable.

Deliberately absent, forever: autoplay-next, infinite scroll, engagement-ranked anything. The feed has a bottom, and says so.

Private by construction: there is no Winnow server and no telemetry. The only network traffic is to youtube.com (as you, for your data) and directly to the AI provider you configured, under your own key. Full policy: https://github.com/science/winnow/blob/main/PRIVACY.md

You'll need: a youtube.com login in the same browser, and an Anthropic or OpenAI API key (a cheap model works well — a cold start on a ~200-video feed costs on the order of $0.10, daily refreshes cents).

Source code: https://github.com/science/winnow