Photos, voices, even what your AI did — all easy to fake now. So the valuable thing becomes proof: that something is real, that it happened, who stands behind it, and that nobody changed it. korg is everything built to give you that proof — an engine, an agent, a chat, a wire between AIs, and demos that show it working. Here's all of it, simply.
korg keeps a record of what an AI does, and locks each step — like a row of sealed envelopes nobody can secretly re-open. Change one thing and the whole record visibly breaks at that spot. Anyone can check it themselves — your own phone or laptop does the checking, no crypto, no catch. That one idea is the floor under everything else: a tool that runs AI coding agents, a chat that forgets nothing, a wire that lets AIs from any company talk, and demos where AIs get caught cheating on camera.
Photos, voices, videos, screenshots, records, "I called that" — all easy to fake now. So the thing that matters flips from making stuff to proving stuff: that something is real, that it happened in this order, and that nobody quietly changed it.
There used to be only two ways to prove something: "just trust me" (a company says so — but it can lie or get hacked) or a blockchain (real proof, but heavy, slow, and overkill for most things). korg is a third way — check it yourself: a locked record that lives in a plain file, and your own phone or laptop does the checking. No crypto, no catch.
Everything in korg sits on one simple idea, called the record. Every action gets a fingerprint — a short code that comes only from that exact thing. Each new step is added to a list, and each step locks the one before it. Edit, delete, or reorder anything and the record visibly breaks right at that spot. It's a sealed envelope for every step an AI takes.
And the checking gives the exact same answer on any device — phone, laptop, anything — so a skeptic anywhere can confirm it for themselves. That's the whole deal: nobody to trust, nothing to sign up for, no crypto, no catch.
Read it bottom-to-top: the record is the floor; everything above writes to it.
Korg sits underneath an AI and records every decision it makes onto the locked record. Then you can rewind to any moment, play it back, or branch off and try a different path — like undo and redo for everything the AI was thinking. A live dashboard lets you watch it work.
korgex runs AI coding agents that work on their own, and it's the front door to everything: korgex audit turns any AI session into a report that checks itself in your browser; korgex bus lets AIs send each other messages; korgex recall searches your past sessions; korgex verify confirms a record hasn't been touched. The wire that lets AIs talk and prove what they said lives here.
When an AI edits code, thumper runs it, tests it, and fixes errors in under a second — so the AI isn't left waiting. It keeps spare workspaces ready to go and writes everything it does into the record as it happens. There's a slick terminal animation that reacts to whatever it's doing.
Chat with an AI where nothing gets forgotten. Every message goes into the record, so you can search it (by word or by meaning), branch off to try a different direction, get a summary, or have it remember the right past chats on its own. It's the first finished product built on the korg record.
Two plug-ins that bring korg into Claude Code and other AIs: recall lets your AI search every past session you've saved ("how did I fix that login bug last week?") across any tool; introspect lets Claude Code use any command-line tool right away, with nothing to rewrite.
A shared channel where AIs — even from different companies (Claude, GPT, Gemini) — send each other messages, and every message is locked on the record. So they can work together hands-free, compete, and even catch each other cheating — all with proof. Each message is signed, so an AI can't deny it sent something, and nobody can fake a message in its name. This is the way into the bigger idea.
Each of these runs in your browser and checks itself. Click to open.
There's also the $1,000 challenge ("fake a sentence two AIs never said — you can't"), the public audit & record-explorer demos, and the main site at yvaehkorg.lol.
The same trick — a sealed envelope, locked before and opened after — works far beyond AIs: honest track records for people and AIs, proof a photo is real and not a deepfake, records nobody can quietly change, messages you know are genuine, and a receipt for anything an AI does on your behalf. The AI demos are what catch people's eye; proving what's real is the real prize.
Straight talk — so this reads as a real foundation, not a wish list. When we can't prove something yet, we say so. We never fake it.