Use cases
These are eleven of our favorite ways to use Opyt, written out as the exchange you would actually have with your client. We will keep adding to this page as we and the people using it find more.
Everything on this page works today with the tools the docs describe. Nothing here is a roadmap. Cases 03 and 06 run in the desktop app where Opyt is installed; the rest also work from claude.ai and ChatGPT in a browser, through the connector.
you › I'm about to spend a week building our own RAG evaluation harness. Has anyone already built it, and did anyone I trust call it a dead end?
claude › Built, in part. Ragas covers the scoring half and assumes you already have a labeled set. The paper argues reference-free metrics drift once contexts get long, and the thread you saved says why the author stopped trusting them. The open gap is the labels, and that is where your week should go.
Repos, papers and posts rank in one index, so "someone built it" and "someone credible thinks it fails" arrive in the same list. The check costs one query, before the first commit.
you › Keep me up to date on AI drug discovery.
claude › Three things to watch, and two of your people who mostly write about this. Boltz-2 is busy: most of its repos are small wrappers, so I'll only bring you the ones that matter. Want me to follow it, and check it for you every morning?
A topic watches names, not categories. Each name is searched every day on arXiv, GitHub, OpenAlex, X and Substack, and a find counts only when the name is in its title or text. Measured on 403 labelled results, that cut off-topic finds from 44% to 13%.
you › I was heads-down for ten days. What's new?
claude › Two things worth your time on agent payments: a paper on who is liable when an agent pays the wrong party, and a repo settling x402 payments in stablecoins. On drug discovery, the rentosertib Phase 2a write-up is out, and Freda Duan posted that labs are now buying training data.
Your client checks before it shows. New finds wait outside your library until it reads them; it brings you what is worth your time, adds those to your library, and drops the rest without a word. Opyt runs no model for this.
you › Has anyone I follow said anything about that liability paper?
claude › Two of your people have. Christian Catalini argued in August that card networks will carry the liability because they already price fraud. A thread you saved disagrees: it says stablecoin rails leave no one to charge back. Nobody else has said anything on it yet.
Only what they said, never what they would think. Every line cites the post it comes from, and when none of your people has said anything, that is the answer.
you › What do the people I trust actually think about letting agents write to production?
claude › Six of your people have written on it, and they agree on one thing: never a standing credential. Four want a human approving each write. Two argue short-lived scoped tokens make that unnecessary, and one of those has run it in production for a year, which is the post to read before you pick a side.
One query returns where your people landed, and who dissents. The split comes from six sources you chose, so the minority view here is somebody you already trust rather than a loud stranger.
you › Whose work do I keep saving without noticing?
claude › Six people, ranked by what you already did. The top one you liked six times and bookmarked twice and never followed. The second wrote three of the papers you saved. The endorsement already happened, one save at a time. One confirmation brings in the whole archive.
Every name shows the count that put it there, so you can check the ranking before you trust it. The evidence is clicks you already made, which is how somebody you forgot about still lands on the list.
you › We are moving our API from per-seat to per-token pricing next quarter. What have my sources said that bears on it?
claude › Eleven things, and they split on one variable: whether a customer's usage tracks their own revenue. Four operators who shipped per-token report support load up and churn down. Two argue the opposite where usage is lumpy, and one of those is a post you saved and forgot. The decision stays yours. The homework ran against people you chose.
Your client reasons over full source text, on the subscription you already pay for. When an answer changes your mind, you can open the exact post that did it.
you › Draft the argument that open-weight models win the enterprise, cited only from what I saved.
claude › Nine citations, every one linking back to its source from someone you confirmed. Two places where your store holds nothing are marked as gaps rather than filled in from general knowledge, so you can see what you still have to read.
No claim without a source, and no source you never chose. The gaps stay visible, which is the part a model working from general knowledge will not show you.
you › Ana sent me an Opyt invite. I have to pick an eval setup for retrieval this week. Has she already been through this?
claude › She has. Ana read the framework and its repo back in March, and she saved a thread arguing that reference-free metrics drift once contexts get long, which is the one that bears on your week. All four citations are hers and say so. You hold nothing on this yourself.
A friend's reading becomes searchable the moment you accept. Ana built that store from people she chose, and every hit says it came from hers, so you can weigh it as her judgment and open the source yourself.
you › What have I been reading most since March?
⚙ aggregate(date_from="2026-03", sample=200)
cursor › Tariffs and inflation is your biggest topic since March, with grid and data centers close behind:
The counts come from your store, not from a guess. Your client names the subjects from a sample of what you saved, then checks each one with a search before it reports it.
you › Save this: arxiv.org/abs/2608.01144
you › yes
claude › Saved, searchable beside everything else. Opening it reads the whole paper.
Saving something is one sentence. No folders, no tags, no deciding where it belongs.