criteria
How the picks
actually get made.
No algorithmic feed. No scraping the top of a trending list and calling it a day. Each edition is written for one person, then delivered on their day, at their time, in their timezone.
Why this exists
GitHub is enormous and most of the interesting stuff never makes it to the trending page. The good projects get buried the week they launch. The popular ones get recycled on Twitter until everyone's sick of them.
finds.dev started from one simple annoyance: the repos worth reading about almost never match the repos the internet wants to talk about. So the whole point of this service is to do the digging on your behalf — and to be honest about what we find, including the parts that don't work.
The promise. 3–5 repos a week, picked for you specifically, each with one honest out-take — what's genuinely good, plus what to know going in. No filler, no "comprehensive solutions", no AI slop.
How the agent works
An AI agent runs once a week per subscriber. For each user, it:
- Reads your interests and what you've clicked before.
- Looks for repos that match — ones it already knows, plus fresh finds from GitHub.
- For anything new, it actually reads the project and forms an opinion on what's good and what isn't.
- Picks the best 3–5 for you — not the best overall — and writes the email.
Once a repo's been properly looked at, that opinion gets reused for everyone else it's relevant to. That's what keeps the service free.
Your preferences, in your own words
The interests field is a plain-language textarea on purpose. Tags are lossy. A list of checkboxes can't express "backend infra in Rust, local-first stuff, anything that makes Postgres weirder — not interested in crypto or generic AI wrappers."
The agent reads your sentences the same way a thoughtful colleague would: it picks up on what you want, what you don't want, and the tone you're looking for. Change your mind? Rewrite the description. The next edition shifts.
Clicks teach the agent what you actually like
Every repo link in the email is a proxied redirect. When you click through to a repo, we record that — quietly, without any analytics middleman — and the next agent run reads it.
Repos in topics you've clicked on before get a relevance boost. Topics you consistently skip get less weight. Over a few editions, the picks sharpen around what you'll actually open, not just what you said you might like.
If you go five editions without a single click, we pause the subscription and send one final email with a reactivate link. Ignoring us means we stop emailing you. That's the deal.
Delivered on your schedule, in your timezone
Pick the day and time you want finds in your inbox. Want it Friday at 08:00 before coffee? Sunday night with a beer? Monday 06:30 on the commute? Whatever suits you.
Your timezone is stored and respected. A subscriber in Lisbon and a subscriber in Tokyo both ask for "Friday 8am" and both get it at Friday 8am local — not in UTC, not in whatever happens to be convenient for the server.
Language
Today, the email and the hosted edition page are written in English. We collect a language preference at signup so non-English delivery can roll out without re-asking everyone — when it ships, the agent will write the opinion directly in your language rather than running it through a translation layer.
The things finds.dev refuses to do
- No "curated insights" voice. If a sentence sounds like a LinkedIn post, it gets rewritten.
- No repeating picks. We track what you've been sent. A repo only returns if it's genuinely changed since last time.
- No hiding criticism. Every out-take ends with a real caveat — the catch you'd want a friend to mention before you click.
- No hidden "sponsored-but-looks-organic" slots. If a spot is paid, it's labelled.
- No list sharing, ever. Your email does not leave this system.
That's the whole thing. One honest weekly email with a handful of repos someone actually read before recommending.