new·Update the plugin: three new commandsThe mozg plugin is at 0.4.0. New: /mozg:update re-reads a brain against its sources (only pages that actually changed cost anything), /mozg:sync writes your shelf into .mozg/brains.md so every session starts knowing your brains offline, and /mozg:add shelves a catalogue brain without opening a browser. Training got a five-phase protocol that skips what the brain has already read.all news →
mozg.betaSign in

Start from zero

Never heard of
any of this?
Good. Start here.

No jargon, one idea at a time. By the end of this page you will know what an AI agent is, why it gets confident things wrong, what MCP means, and what this service actually does about it. If a word here needs explaining, it gets explained — and there is a list of them at the bottom.

One

An AI agent is a colleague with two odd properties.

Imagine hiring someone who has read almost everything ever published — every manual, every forum, every tutorial — and can write code, edit your files and run commands. That is an AI agent: Claude Code, Cursor, Codex. Genuinely useful, and strange in two specific ways.

It stopped reading on a date. Its knowledge was fixed when it was trained. Anything published after that — a new version of the tool you use, an option that got renamed, the API that changed last Tuesday — it has never seen, and it does not know that it has not seen it. So it answers from what it remembers, with total confidence, and sometimes that answer is a year out of date.

It forgets everything between conversations. Explain your project's rules today and tomorrow it starts blank. Everything it knows about your work has to be handed to it again, every single session — and every word of that costs money to read.

Two

Which is why the answers look right and are not.

Two kinds of knowledge are missing, and they fail differently. Public knowledge that moved — the documentation of whatever you are building with, which changed after the model stopped reading. And knowledge that was never public at all— your project's conventions, your company's internal system, the way your studio does things. No model has ever seen that, so everything it says about it is invention.

The usual fix is to paste it all in — a long file of instructions the agent reads every session. That works until it does not: the file grows past what fits, you pay for all of it whether today's task needed it or not, and worst of all it cannot tell you when it went stale. A document is silent about its own age.

Three

MCP is a socket. That is genuinely all it is.

Agents need to reach things outside themselves — your files, a database, a service. Every agent used to do that its own way, so every service had to be built once per agent. MCP (Model Context Protocol) is the agreed shape of that connection: build the service once, and any agent with the socket can use it. Like USB, and about as interesting — what matters is what you plug in.

Claude CodeCursorCodexMCPbrain84%3 notes
One connection, any agent, and only the few notes the question needed

Four

We keep the knowledge, and we measure it.

A brainis a small library about one subject — SvelteKit, your company's billing service, the way your studio builds games. Inside it are notes: one fact each, written to stand alone. Your agent connects once and then asks the brain whenever a question depends on that subject, getting back the two or three notes that answer instead of a document it has to read whole.

The part nobody else does is the exam. Every brain is given questions drawn from what it claims to cover, made to answer them, and marked — so it carries a percentage, and a public list of the questions it fails. That is the difference between “here is a document, good luck” and “this knows 84% of its subject, and here is exactly where it does not”. An agent that is told where knowledge ends can say “I don't know” instead of inventing.

Where do brains come from? Some are made already — the catalogue holds the documentation of a hundred popular tools, free to use, kept current. Some you make: paste a documentation link and it gets read for you, or point your own agent at your own files and let it write the notes. And some people sell theirs.

Five

What it looks like in a real morning.

You do not have to know the right word for what you want. That is the whole point of asking something that actually knows the subject.

“Make the spin feel more exciting when two of the special symbols land. I have no idea how these games do that.”

Without a brain

A shake and a louder noise once the reels have stopped — reasonable, generic, and pointed at the wrong moment. Nothing was wrong with the request; the agent simply answered exactly it.

With the brain

It asks the brain first and comes back with the name of the thing: anticipation. The last reel slows down and plays a build-up animation while it is still spinning, which is where the tension in these games actually lives — plus the two house rules about when it must not fire. You had never heard the word, and got what somebody with ten years in the industry would have specified.

Six

Three steps, about ten minutes.

  1. Take a brain. Open the catalogue and find the tool you are working with. Free, no card, and you can read what it knows before deciding.
  2. Connect your agent. One command, copied from the connect page. It works with Claude Code, Cursor, Codex and anything else that speaks MCP — this is where that socket earns its keep.
  3. Ask normally. You do not change how you talk to your agent. It searches the brain when a question touches that subject, and tells you when the brain does not know.

Reading and connecting is free, forever, and so is the whole catalogue. Money only appears when you want our AI to read documentation for you — and even then you can point your own agent at it instead and pay us nothing. The pricing page says exactly which is which.

Words you will run into

in plain language
AI agentA program you talk to that can also do things — read your files, write code, run commands. Claude Code, Cursor and Codex are agents. The chat window in a browser is not: it can only talk.
ModelThe part that does the thinking, trained once on a huge pile of text. Training has an end date, and after that date the model learns nothing new on its own.
ContextEverything the agent can see right now: your question, the files it opened, the notes you pasted. It is limited, it is refilled from scratch every session, and everything in it costs money to read.
MCPModel Context Protocol — an agreed way for an agent to plug into an outside service. Like USB: one socket, and anything that fits it works with any agent that has the socket.
BrainA small, searchable body of knowledge about one subject. The agent asks it a question and gets back the two or three facts that answer, instead of reading a whole document.
NoteOne fact inside a brain, with a title, written so it stands alone. Notes are what the agent actually reads.
Exam scoreA percentage on every brain. Questions are generated from what the brain is supposed to cover, the brain is made to answer them, and a judge marks it. 84% means it answered 84% of them.
GapA question the brain fails. Listed publicly, on purpose — knowing where knowledge ends is worth more than a promise that it does not.
TokenTwo meanings, unluckily. Text is billed in tokens (roughly ¾ of a word). And an access token is a password-like string that lets your agent into your account.

Still not sure this is for you?

Fair. Read how people actually use it — an artist selling his method, a company teaching the AI its own software, a game studio, a maintainer, an agency. One of them is probably closer to your situation than anything on this page.

Look at the catalogueSet it up, step by stepHow people use it