It started with video
We had a real interest in video — and a growing pile of it we wanted to archive and actually search. Somewhere in solving that, it clicked: the video problem was the file problem. The same needle-in-a-haystack lived in our documents, our email, our code — everywhere.
So we did what everyone’s doing now: we pointed AI agents at it. And it got expensive, fast. The agents re-read everything, burned tokens, and we’d still be waiting when all we wanted was to get the thing done. It cost us money and time in equal measure.
So we built the thing we actually wanted: something that reads your material once, understands it, and then hands you — and your agents — the exact page, cell or timecode instead of the whole pile.
The principle: least in, most out
The way it usually runs, you give the most and get the least. Everything you own goes into the model, on every question, and what comes back is an answer you then have to go and verify.
We think the first question should be the other one: what’s the least we can send to get the most back? There was a reason to hand over everything once, when nothing could tell what mattered in advance. There isn’t anymore. Understanding your material at ingest is what makes the small answer possible — and the small answer is cheaper, faster, and leaves less of your business sitting in someone else’s context window.
The same flip explains the rest of the design. Your files stay where they are. Your existing permissions still apply. Nothing gets reorganised, because a system that needs you to tidy up first is a system that assumes a kind of time nobody has when they’re actually in the work.
Where it actually is
Early beta, on Mac, with Windows and Linux behind it. The engine works and the savings are measured — every number we publish is on the evidence page with its methodology and its caveats attached, and the full agent memos behind them are published in the work.
What’s left is scale: running well on your setup rather than the machines we tested on, and staying quick as a collection grows. That’s what the beta is for, and it’s why we’d rather call it a beta than a launch.
Keep your chaos. We’ll find the thing.