Start here · ~10 minutes to a thinking agent
From zero to an agent
that actually knows.
This page is the whole journey: what mozg is and why it exists, then every step — take a brain, connect your agent, prove it works, build your own. Each step shows the screen you'll be looking at.
First, the problem
Your agent answers from memory.
Memory has a date on it.
Models are trained months ago; your stack moved last week. So the agent answers about the old version — with the same confidence either way. The usual fix is pasting docs into context files, which are expensive (you pay for every word, every session), rot silently, and can't tell you what they actually cover.
mozg does three things differently: knowledge lives in a brain the agent searches over MCP (only the notes a task needs enter the context), every brain sits an examso “trained 92%” is a measured fact with the failures listed, and brains learn from use — a question nobody could answer becomes an exam question automatically.
Make an account
why: the free tier is real: read the catalogue, connect agents, one trial brain — no card.
mozg.sh/sign-in — email, Google or GitHub. You land in your workspace: a left rail with your brains, balance and tokens.
Take a brain from the catalogue
why: someone already built and examined the brain for your stack — starting from a proven one beats building blind.
Open the catalogue and pick your stack — Next.js App Router, Expo, Svelte 5, Tailwind v4, the MCP spec. Every card shows the same three honest numbers:
web · mozg
Next.js App Router
trained 84% · 1,213 notes · free
still learning: middleware edge cases · turbopack config
That “still learning” line is the point of the whole product: the brain tells you where not to trust it before you ask. Hit Add to my brains — free ones join your shelf instantly, families arrive whole.
Connect your agent
why: one command; after it, every agent session can search your shelf without you pasting anything.
Grab a token at settings → tokens, then in your terminal:
Codex, Cursor, Cline, Kimi, Qwen, VS Code — exact copy-paste for each lives on /connect. Claude Code users can install the plugin instead — it also nudges the agent to check the brain before answering.
Prove it works
why: never trust wiring you haven't seen carry current.
Ask your agent something spec-level from the brain you added:
The tell: the agent cites the brain instead of reciting. Searching costs zero context tokens — retrieval runs on our side, only the notes it picked enter your session.
Build your own — from one link
why: your project's real knowledge isn't in any public doc; the trial brain shows the whole loop before you pay anything.
New brain → paste one documentation URL. The crawler finds every page (GitHub repo, llms.txt, sitemap, or a link walk), reads them into searchable notes, drafts a goal — and then the important part happens:
exam · sat automatically
✓ 19 passed · ✕ 7 failed · trained 73%
failed: webhook retry order · rate limit headers · sandbox auth…
The failures are not a bug report — they are a shopping list. Feed those topics (more pages, a PDF, or just tell the brain in the teach flow) and re-sit. Free accounts get one trial brain, 30 pages, one exam sitting — the full loop, honestly sized. Your brain is private: only you and your agents can reach it, unless you ask for the catalogue (that goes through review).
Let it get smarter — and learn it yourself
why: a brain is not a file: it improves from use, and you can study the same material your agent queries.
From here everything is automatic: searches that find nothing become exam questions, corrections your agents file arrive for your review, re-reads keep sources current, and /mind shows everything your agents can know on one screen. And the same notes double as a human course — learn.mozg.sh: read, recall, quiz, streaks, a certificate at 80%, and a scoreboard against your own agent's exam score.