Gistrel collects the AI, model and agent posts you save on X, Hacker News and YouTube. It tells you what is actually new to you, whether a long read is worth it, and explains it at the depth you want. Everything you keep stays findable.
Just saved
from your last 20 minutes on X
What we learned running 400 coding agents in parallel
Article · 38 min read · 212 HN comments
Gist: 2 of 9 ideas are new to you, mostly in the section on cache-aware task scheduling. Comments add a first-hand cost breakdown.
Building an evaluation harness from scratch (full talk)
YouTube · 1 h 04 min
Thread: a sign-in animation built with only CSS
Post with a 14 s video
Reads where you already read
How it works
You skim fast and save what looks promising. Gistrel does the slow part: reading the whole thing, checking it against what you already know, and keeping the useful pieces.
Bookmark on X, send a link, or click the extension button. You decide which signals count, so bookmarks you use for other things stay yours.
Each item gets concrete ideas, evidence and first-hand experience pulled out, then sorted into new to you, familiar, outdated, or worth a closer look.
Useful material goes into your collections, with notes, images and clips. Ask for it in plain words months later, even if you forget the title.
Today’s stories
Tweets, articles and Hacker News threads about the same event are merged. You get what people who tried it actually experienced, which opinions have evidence, and the practical nuances a headline leaves out.
From 9 posts, 2 articles and 1 Hacker News thread
Cache the system prompt and tool definitions so repeated turns are cheaper. The less obvious move is to schedule tasks so that ones sharing a prefix run back to back, which kept our hit rate above 80%.
⋯ 4 paragraphs of setup history, skipped. Expand
We also found that summaries written by the agent itself drift after a few hours.
Familiar
You marked this known on Sep 30.
New to you
Ordering by shared prefix, not just caching.
New twist
Contradicts a post you saved in August. Compare
Worth reading?
An intriguing title often hides obvious ideas or stale advice. Gistrel reads the article and the discussion first, then tells you which ideas are new, which are outdated, and where the real substance is.
When you do read, one click gives every article the same calm typography, with notes in the margin where the ideas occur.
Long videos
Gistrel reads the transcript, pulls out what is new for you, and points to the exact moments. When the transcript is missing or partial, it says so instead of guessing.
Building an evaluation harness from scratch
Conference talk · 1 h 04 min · transcript complete
▶ 11:20 – 15:50
Grading agent runs by replaying tool calls against a frozen sandbox. New to you.
▶ 39:05 – 42:10
Why their pass rates fell when they switched judge models. First-hand numbers.
It learns you
Mark an idea as known, and Gistrel recognizes it the next time someone phrases it differently. If a later post adds a real twist, it explains what changed instead of hiding the whole item.
“Prefix caching cuts the cost of repeated agent turns.”
Interest depth
Different topics deserve different depth.
A big text-to-speech price drop still reaches you. The architecture paper behind it does not.
Collections & search
Save a post with a short UI animation and almost no text. Gistrel looks at the images, links and video, describes what is in them, and files it where you want. Search uses those descriptions and your notes, not just the original words.
Someone shared sign-in flow animations a while ago
Morphing sign-in button
Matched: video shows a sign-in form whose button expands into a loader.
Passkey-first onboarding
Matched: your note “nice auth step transitions”.
Animated sign-up with avatar
Matched: in your UI inspiration collection, bucket “auth flows”.
Your reading stays yours
Read-only
Gistrel asks X for read access. It never posts, likes or follows on your behalf.
You choose the intake
Turn bookmarks or likes on or off, or only send items explicitly.
Never sold
No ads and no data sales. Your interests and notes are not shared with other users.
Delete anytime
Disconnect a source or delete your account, and the data goes with it.
Details are in the Privacy Policy.
People who follow AI, models and agents closely but have little time to read. It started as a personal tool for one developer who scans X and Hacker News while coding agents run, and it is opening to a few more readers.
No. Keep scanning where you like. Gistrel picks up what you save and is where you go to understand it.
From what you have saved and read, and from quick feedback such as “already knew” or “not interesting”. It treats three questions separately: is this recent, is it new to the field, and is it new to you.
Gistrel uses AI models, which can be wrong. Every assessment links to the passage, comment or timestamp it is based on, and it says when evidence is missing or incomplete.
The web app comes first and works on desktop and phone. A Chrome extension for one-click capture and in-page annotations is planned.
Early access is free. If that changes, you will be told before anything is charged.
Gistrel is in early access. Tell us what you read and we will let you in as space opens.
Request early access