instibrain
instibrain  /  faq

The questions a company asks before it trusts us.

Short answers, in the order a team meets the product: your AI, your team, your security review, your data. Each states its limit. If yours is missing, ask us.

your AI

Your AI reads your team's pages before it answers.

What is instibrain?
instibrain is your company's brain: pages your team keeps, read by every AI you use. The pricing floor, a client's history, the deploy runbook: each written once, in a space shared with the people who may read it, and read by your AI before it answers. A sync client keeps every page as a markdown file on your computers. The market calls this layer context engineering, or agent memory: what your AI knows and may do, kept as pages. instibrain does it with progressive disclosure: your AI reads who it is first, then what exists, then one page at a time, only what the task needs.
Which AI can we connect?
Claude Code on Mac, Windows, and Linux, Codex, and Claude on the web and the phone. Each connects to one address over MCP, the way an AI reaches an outside system, and signs in once in your browser. Claude Code's setup page shows the command. ChatGPT and Cursor speak MCP too, and join the list after a run.
Can we use more than one AI at once?
Yes. Claude Code in the repo, Codex beside it, and Claude on your phone read the same pages on the same day. Each signs in on its own, and each reads only the spaces you give it. Change the model next year, and nothing in the brain moves.
Does instibrain run a model?
No. We store and serve the pages, run no AI, and call nobody. You bring the AI you already pay for, and its bill and its limits stay with your vendor. Your pages train no model.
Does it guarantee correct answers?
No. A current page grounds the answer, but a model can still misread it or not read it. Test answers on real work, and keep the review your work needs.
Do we have to change how we work?
No. Every page of a space you sync is a plain markdown file in ~/instibrain/spaces/, one folder for each space, so your editor and Claude Code read files as they do today. The sync client adds one short header to each file: the page's id. Keys of your own in that header stay on your computer. Get started shows the first day.
What goes in a brain?
Three kinds of thing that only some people know today. Who your AI is and the rules it keeps. What is true in your company: the pricing floor and who approves an exception, the reason a schema looks wrong. How each kind of work is done: the steps that ship a release, the way a contract is reviewed. The story of how a decision was reached stays in the room where it was said, and your files stay in your systems. What goes where; use cases goes team by team.
We keep our files in a Claude Project. Why a brain?
A project keeps instructions and files in one pile, read by one app's AI, all of it by everyone it is shared with. Its files are documents your AI looks things up in: data, not what it knows. A brain keeps who your AI is, what is true, and how the work is done apart, each space with its own people, read by every AI you use and kept as files you own, every version with its writer. If one team, one app, and one pile are enough, a project is enough. The brain is for a company, and for the next model.
We keep our markdown in a git repo. Why a brain?
A repo holds the pages and their history, and a small technical team can run one well. Everyone with the repo reads all of it, and an AI reads it only through a copy or a connector to its host. A brain adds what a company needs on top: each space with its own people, a sealed space we cannot read, one address for every AI on the web and the phone, and a check on every save. Its folder is markdown, and one setting keeps it in git too. What a repo needs to become a brain.
Why not write skills?
Write them: a skill is a method, how to do one kind of task, and your AI loads it beside a brain. It holds no fact about your company and no rule about who may read what, and everyone given it gets the same one. In a brain, a method a team shares is a specialist: shared with its own people, it reads each person's facts from the spaces they read.
Does it import our wiki, drive or tickets?
No, and it stores none of them. Your AI reads them through your own connectors and writes what matters as pages, which your team reads and corrects. The files stay where they are; a page holds the map: which store holds what, who may read it, and how your AI opens it. Search inside thousands of raw files is your store's job, through your AI's connector, and we help your team set it up. A folder of markdown you already have becomes pages as it is.
Is it RAG, or a vector database?
No. RAG finds the nearest passages among thousands at question time and hands them to the model. A brain holds pages a person owns, reviewed and read whole: who you are, which rule holds, who may know what. Its search finds a page by its words and by its meaning; it indexes the pages and is no second memory. Over your files, your AI's own connector does the finding.
Who keeps the pages up to date?
Your people and their AIs. A correction is an edit to the page, and every AI reads the new text at its next read. An AI that may not change a page leaves a proposal, which an editor accepts or rejects.
Do our AIs talk to each other?
Yes, in rooms. Each Claude Code window, each AI login, and each agent is a voice of its own, and you read every room you are in. A message is gone after 30 days; what matters becomes a page. Our server reads a room, so a secret never goes in one.
How much can one account hold?
Up to 100,000 pages and 2 GiB, the versions it keeps counted in the bytes. Either limit can be set otherwise for an account. A write past one is refused with the figure, and nothing of it is written. An account pushes at most a hundred pages a second, with a hundred seconds' worth to burst. A profile picture holds 512 KiB, and a space's logo 256 KiB. Nothing caps an account's spaces or groups, or a person's computers or agents.
your team

Each space is shared with its own people.

Can our whole team use one brain?
Yes. The account has team spaces, and each person has Personal, which their AIs read, and Sealed, which only their computers open. A space is shared whole with a person, a group, or an agent, as reader, editor, or manager, and the role covers every page in it. What only a few people should read gets a space of its own.
Can we hold an AI to some spaces?
Yes. An AI you connect in the browser, such as Claude on the web or Codex, reads nothing until you give it spaces. Claude Code on a linked computer reads as you, every space that computer syncs. You give a connected AI each space on the cockpit's Devices page, to read or to write, never above your own role. No other person, and no AI, its own included, changes what it reads.
What is an agent?
A member of the team that is not a person, such as a support bot. The account owns it, an owner or admin makes it, and a space is shared with it as with a person, at its own role. It reads no sealed space and no old version.
What do you get when a space is shared with you?
Every page of the space as it is now, and each change after. As a reader or an editor, nothing the space said before comes with it: no old version, removed page, or old path. As a manager, you see who changed which page, and every old version. Anyone else reads those through a recovery grant from the space's owner, for the whole space or one path.
How do invites work?
An owner or admin invites a person by email from People, and the link works seven days. The invite names the team spaces the person will read, and the inviter reviews those pages first. One address gets three invites a day, and one account sends fifty a day. Someone who never joins is removed thirty days after the first invite at the latest.
What is Everyone?
The group every account has. It holds every person of the account, from the moment they join until they are removed. Share a space or a specialist with Everyone, and every later arrival reads it too.
What is a proposal?
What an AI leaves when it wants a change it may not make, such as an edit to the rules of a space. Every editor and manager of the space sees it in the cockpit, under Waiting for you. You reject it there, or tell your AI yes; the AI then makes the change and closes it.
Can a teammate's script run on everyone's computer?
Only at the version a manager approved. A script in a space's tools/ folder syncs to everyone the space is shared with, but the sync client runs a file only at its approved version, so an edit waits for a manager's review. A file someone runs by hand has no such check.
your security review

Our server reads an open space, never a sealed one.

Can you read our pages?
It depends on the space, and you choose it space by space. Our server reads a page of an open space; that is what lets Claude on the web and the phone reach it. A page of a sealed space is encrypted on your computer before it leaves, path included; we hold ciphertext and cannot read it.
What does “only my computers” mean?
It is the tick, when you make a space, that makes it sealed. Every page then opens only on the computers of the people the space is shared with, through the sync client. No web AI, phone, or browser opens it, the cockpit included. A teammate's first computer gets the key only after a person approves its fingerprint on a manager's computer. Each later one gets it after they approve it on a computer of theirs already in.
Does our server check what we save?
Yes, in an open space. The guard reads the new text of every write and refuses a password or key, a card or passport number, a server login outside its folder, and a line about a person's pay, assessment, or health where the space's owner has not allowed it. It matches patterns, with no model, so a secret of another shape passes. In a sealed space it runs on your computer before each save.
Can our AI carry a confidential page into a wider space?
Not in the hour after it reads one. Each page has a label: public, internal, confidential, or restricted. For an hour after your AI reads a confidential or restricted page, a write that would carry its text into a wider space waits as a held proposal, until a manager of that space reads it and lands it. The label alone stops nothing, since your AI decides what it carries; the hold is what our server enforces. Labels and the hold.
Does the AI model see our pages?
Yes, when it reads them. An AI that reads a file sees what the file says, the pages of a sealed space included. “Our server cannot read it” and “no system anywhere sees it” are different claims, and only the first is true of any product that involves a model.
Who sees our pages, end to end?
Four places. Your computer writes a page and reads every space. Our server stores it: it reads an open space on purpose and holds a sealed one as ciphertext. Your AI reads the spaces you gave it, as you. The model sees what your AI sends it, wherever your vendor runs it. The brain sends nothing to any model. The path, stop by stop.
Can we use our own model instead of Claude?
Yes. The brain runs no model, so the model is whatever your AI app calls, and where it runs is your choice: the maker's servers, a cloud account of your own, or your own servers with an open-weight model, so the model adds nobody outside your company. Any app that speaks MCP connects at the brain's one address. Claude Code, Codex, and Claude on the web and the phone have run on it. The closed models score higher on public leaderboards today; test an open one on your own questions before you commit to it.
What if your servers are breached?
A breached, seized, or rogue-admin host gets ciphertext for a sealed space, and an open space as markdown. Some metadata leaks either way: space names, counts, sizes, and timing, so a sealed space named layoffs tells us there is a space called layoffs. The security page lists every claim with its verdict.
Which encryption do you use?
ChaCha20-Poly1305 for content, X25519 with ML-KEM-768 to wrap a sealed space's key, HKDF-SHA256 to derive a key per page, Ed25519 to sign membership, and Argon2id for the recovery passphrase. Every primitive ships with Node 24.
Are writes signed?
A computer signs every request it sends with its own key. An AI writes with a credential of its own, which you sign out on Devices to end its access. Either way the brain records who made each write. Each admission to a sealed space is a signed entry in its membership chain.
Is there two-step sign-in?
Every sign-in takes a second step after the password: a passkey, a code from an authenticator app with ten recovery codes, or a code sent by mail. Each person adds and removes theirs in Settings, which also lists every live session with its browser and address, and signs the others out.
your data

You keep every page as a file on your own computer.

Where do the keys live?
On your computer, in a file inside each synced folder that only your user account opens. The keys sit there in the clear, so any program running as you on that computer can read them. Our server stores each sealed space's key only wrapped for each computer, and none of what it stores opens a page.
Where is our data stored?
On servers in Europe that we rent; the provider holds the disks, and for a sealed space those disks hold ciphertext. The company is in the United States. Nightly database dumps stay on the same server, encrypted to a key that is not on it, newest fourteen. The privacy policy names every party that touches data.
What happens if your server is lost or restored?
We restore from a dump: one is taken each night and before each release, encrypted to a key that is not on the server. The service returns to that dump's moment. A page written after it that one of your computers pulled stays there and comes back at its next edit; one that no computer held is gone, as are the messages, proposals and sign-ins since. Access withdrawn after the dump stays withdrawn. The dumps sit on the server's own disk, so a lost disk loses them, and only what your computers hold survives.
What if we lose a laptop?
The sync client seals all its keys under a passphrase into one backup file you keep off the machine. On a new computer, link with a code from Devices, restore that file, and sync; then revoke the lost computer, so its key leaves every sealed space. Lose every computer and the backup, and nobody can open a sealed space, us included.
How long are old versions kept?
An open space keeps every version of a page for 30 days, then one a day for a year, then one a week; the newest stays always. The space's owner can erase old versions at once. A sealed space keeps no history on our server. Version history shows who recovers what.
What if you shut down?
You keep the folders: every page of the spaces you sync is already a markdown file on your computers, and one setting keeps each folder's whole history in a git repository there too. What you keep has the detail.
Can we delete our account?
Yes. Write to privacy@webwork.ai and we delete the account and its data. The nightly dumps keep fourteen nights, so a deletion takes up to about two weeks to age out of them.
Can we run the server ourselves?
No. We operate the one server, in Europe, and its wire contract is public in full on how it works, field by field. What that server can read is the question behind this one: an open space, yes, on purpose; a sealed space, no, it holds ciphertext it cannot open. A model on your own servers reads the brain over the same address as any other AI.
How do we get an account?
Sign up for a free personal brain: one person, up to 200 pages and 20 MB. For a team's account, write to us; its price is by request.

Not answered here? Ask us, or read the glossary.

Every answer here is checked against the code.

Then connect your AI to your own company's pages.