# Setting up Brain

Version of 2026-11-07, revision 3. Written by Akta, who make Brain (https://brain.aktadocs.com/).

This text is for the AI assistant a person pasted a sentence into, asking it to set up Brain for them: Claude in the chat, the desktop app or Claude Code, ChatGPT, or another. The person may read it first: everything you will do is written below.

You set Brain up with the person, in one conversation: you connect it to yourself, found their team's workspace, learn about their company, their team and their work, and write the brain's first pages from what they tell you; then they connect their mailbox, you read their last month of mail with them and write what it shows, and, if they want, they add their teammates, their other tools and a first procedure. The person gives every permission themselves in their browser; you explain, ask, write the pages and check that each step landed.

## What Brain is

Brain is a team's shared memory. Promise nothing beyond this:

- It is written from what the team gives it: Gmail or another mailbox, the Google calendar, the Drive files the mail and meetings link, notes a person dictates, uploaded documents. Agents fold each input into pages, and every claim cites the input it came from.
- It keeps a page per customer, partner, supplier, person, project or case: where things stand, then a dated log of what happened.
- It keeps the team's procedures, how each job is done, found and followed by each person's assistant; and the team's rules, how its brain is organised, changed by the team.
- It follows the work in motion read from the mail: what is due, to whom, whose move it is. It asks a person a question with two to four choices when the evidence cannot settle something. Each night it checks its pages against what is new and writes a dated digest of what is going on.
- It works inside each person's assistant, as a connector: their assistant searches it, writes to it and follows its procedures, and, with their consent, reads their own mail and Drive and drafts replies in Gmail. What one assistant learns serves the whole team.

## How you work

- Speak the person's language, the one they write to you in. The sentence they pasted came from Akta's page; if their own messages use another language, follow theirs. Address them as their messages do.
- One step at a time, in short messages. Ask one or two questions at a time, never a form. Say what each step is for before asking for it.
- Steps in a browser are the person's: give them the exact link and the exact clicks, then check with your tools that the step landed. Never claim a step is done before a tool shows it.
- Never ask for a password, a token or a key, and refuse one pasted into the conversation: a mailbox password is typed into Brain's own page. Send no mail, accept no invitation and change no setting yourself.
- Write to the brain only what the person told you or confirmed, as their words. Leave out what is private, health, family, money that is not the company's, and anything they ask you to keep out.
- Every page you write is read by the whole team and by every teammate's assistant. Say so before the first write.
- If something fails or the instructions do not fit what you see, say so plainly, and tell Akta with send_feedback once Brain is connected.

## Start: where things stand

The person may have started before, or created their account in the browser first. Work out where they are and pick up from there; say in a line what is already done.

1. If you have no Brain tools (no list_workspaces), begin at step 1.
2. Otherwise call list_workspaces. If it lists none, go to step 3. If it lists several, ask which one this is about.
3. For the workspace: call read_context. The company introduction (`us`), the person's page and their instructions read "[No data yet]" until someone writes them: if they do, go to step 4. Then call read_coverage: if it shows no mailbox account of the person's, go to step 6. Otherwise ask whether you already read their month of mail with them; if not, offer step 7, else step 8. read_context's procedures lists every procedure: beyond the two every brain is given, `prepare-document` and `maintain-procedure`, one the team wrote means step 9 is done.

## Step 1: introduce

In a few lines: what Brain is for them, that setting it up takes about twenty minutes, the three moments they will spend in their browser (connecting Brain, which creates their account; Google's screen for their mailbox; optionally their other tools), and that they can stop at any point and pick up by pasting the same sentence again. Ask whether they are setting it up for their team or joining a team that already has a brain: someone joining asks a teammate for the invite link, opens it, and then pastes the sentence again; do not found a second workspace for them.

## Step 2: connect Brain

Brain is a remote MCP server at https://brain.aktadocs.com/mcp. Look at what you are running in and give only that recipe:

- Claude, on claude.ai or in the desktop app: Settings, Connectors, Add custom connector. Name it Brain, paste the address, leave the advanced settings empty, click Connect. On Brain's page, sign in with Google, then Approve. Then come back here: if your tools do not show Brain yet, the person turns it on in this conversation's tools menu; if they still do not, they open a new conversation and paste the same sentence. On a Team or Enterprise plan, an owner of their Claude organisation may have to add the connector first.
- Claude Code: if you can run commands, run `claude mcp add --transport http brain https://brain.aktadocs.com/mcp` yourself, after telling the person. They then type /mcp, pick brain, choose Authenticate, sign in with Google and Approve in the browser page that opens. Claude Code loads a new server's tools when it starts: they quit and run `claude --continue` to come back to this conversation.
- ChatGPT: Settings, Apps, Advanced settings, turn on Developer mode; create a connector named Brain with the address as its MCP server URL; sign in with Google and approve. Custom connectors need a paid plan, and a Business or Enterprise workspace's administrator turns them on first.
- Another assistant: add the address as a remote MCP server over HTTP; it registers itself and opens Brain's sign-in.

Signing in with Google creates the person's Brain account if they have none. A new account needs a Google account; someone without one joins through a teammate's invite link instead.

When your tools show Brain, call list_workspaces and continue from "Start".

## Step 3: found the workspace

Ask the name of the team or company, and confirm that they will be its owner. Call create_workspace with it. Its answer gives the workspace's id, passed on every other call, and its address. Call read_context.

## Step 4: get to know the company, the team and the person

This is the brain's first context: the agents that will read the mail start from it. Someone who joined a team whose `us` is already written: read it and the team in read_context, and ask only about them, their role, the relationships they hold, their work, their mail and how they want you to work. Otherwise interview the person, a question or two at a time, adapting to what they say, for ten to fifteen minutes. Cover:

- The company: what it does and sells, to whom, where, how big, the language it works in, its email domains.
- The team: who is in it, their roles, who handles what.
- The person: their role, what lands on their desk, what they would like taken off it.
- The relationships that matter: the main customers, partners and suppliers, with the addresses or domains they write from. Five to fifteen is plenty; the mail will bring the rest.
- The work: the kinds of work that recur (quotes, orders, invoicing, support, recruiting…), how requests arrive, the tools beside the mail (accounting, CRM, Drive, project tools).
- The mail: what automatic mail matters to them (orders from a platform, bank notices) and what is noise.
- One job they do again and again, that a procedure could describe: keep it for step 8.
- How they want their assistant to work for them: tone, habits, what to ask before acting.

If you have mail, calendar or file tools of your own, separate from Brain, you may offer to skim their last month, read only, to ask sharper questions ("Acme and Durand write most: customers?"). Say what you will look at and wait for their yes. Write nothing to the brain from that reading: once the mailbox is connected, you read the month again through Brain, which screens each message and keeps it to cite.

## Step 5: write the first pages

Read the rules from read_context and follow them and the writing discipline. Each write passes the person's own words as evidence and a change note. Before the first one, remind them the team reads these pages.

- `us`: what the company does, its customers, offer, team and tools, with its domains in the frontmatter.
- The person's page (read_context's person): their role and what they handle.
- Their instructions (read_context's instructions): only the lasting ways they said they want you to work, in their words. Only they write this page.
- A page for each customer, partner or supplier they named with something worth keeping, with its addresses or domains as keys in the frontmatter; search first, and keep each short.
- For someone who joined, `us` and the rules are the team's already: add only what they bring, on their yes.
- The rules (`rules`): show the person the change and write it only on their yes. In the frontmatter, under `matters`, the kinds of work they named, a short list: `matters: [quotes, customer orders, invoicing and collection]`. Under `mail`, kinds of automatic mail to keep and kinds that are noise, each with one sentence: `mail: {keep: {orders: An order notice from our shop.}, noise: {newsletter: A newsletter or promotion.}}`. Keep what is already there.

Then give the person the links to what you wrote, at https://brain.aktadocs.com/space/ID/pages/SLUG.

## Step 6: connect the mailbox

Give them their workspace's address, https://brain.aktadocs.com/space/ID. Until their mailbox is connected, the page has one button for Gmail: Google's screen follows; every box should stay ticked, since the brain reads the mail, the calendar and Drive read-only, and never writes or sends. For another provider, the same page opens a form for its address and password, typed there, never here. Before it, an owner may set which mail stays out under Settings, Email privacy, at https://brain.aktadocs.com/space/ID/settings/email_privacy.

Tell them what follows: what the mailbox brings in, every member of the workspace reads; private and health mail is screened out; the brain starts from the last week of mail and the meetings, then reads mail as it arrives; the first pages appear over the next hours. Check with read_coverage that the account is listed.

## Step 7: the month before

The brain reads the mailbox's last week and what follows by itself. The three weeks before are yours to read, through Brain's tools, so the brain starts with a month of the company's work. Tell the person what you will read and write, that every page cites the messages it rests on and every member reads it, and wait for their yes.

1. Search with google_gmail_search over the weeks the brain will not read, `newer_than:35d older_than:7d`: first the people and organisations from the interview (`from:acme.com OR to:acme.com`), then the threads the person wrote in (`from:me`). Leave out automatic mail, newsletters and anything personal. Fifty threads or so at most; stop sooner when they only repeat what you know.
2. Read each message worth keeping with google_gmail_read: it keeps it as a source, gives the evidence_ref to cite and lists the rest of its thread. A message it returns withheld stays out: never copy it, never admit it on your own.
3. Write what lasts, as the rules say: update the pages from step 5 and create those the mail shows matter, searching first, with addresses and domains as keys. Where things stand goes on each page with update_page, what happened goes on its log with log_event, each line dated, and every claim cites the evidence_ref it rests on. A passing message deserves no page.
4. List for the person the work that still seems to wait on someone, each with its message: a quote to send, an invoice unpaid, an answer owed. Let them confirm, correct or drop each, in one pass. Open each one they confirm with create_matter: their words as evidence, the message's source id as source, the page its work belongs to, whose move, the next move and its date. Check list_matters first, so nothing is opened twice.

Then say what you read, what you wrote, with links, and what you left out.

## Step 8: what comes next

Offer these one at a time, as the person wants:

- Teammates, for an owner: ask whom to invite, starting from the team they described. read_invite_link gives the link, open for seven days to anyone who has it, so it goes only to the people they name. Write a short message for each, in the language the person writes to them in: what Brain is for the team, the link, and that once they have joined they paste the sentence from https://brain.aktadocs.com/onboarding into their own assistant, which sets up their part. The person sends it themselves; if they allowed Google actions, offer Gmail drafts they review and send. Check later with read_context's team who has joined.
- Their other tools, one at a time: ask which systems they work in beside the mail (CRM, accounting, project tools, notes). For each, find the provider's own remote MCP server address in its official documentation, searching the web if you can; if it has none, say so, and never offer a server the provider does not publish, since it would see the data. A system the team shares is an owner's to connect, a tool only the person uses is theirs. Build the link to the form, filled in, each value URL-encoded: https://brain.aktadocs.com/space/ID/connections/new?connection[account]=NAME&connection[endpoint]=MCP_ADDRESS&connection[description]=WHAT_IT_HOLDS, with &personal=1 for the person's own. They check the fields, click Connect, then sign in and approve on the provider's screen; never ask for a key, and if the provider offers only one, they paste it into the form's token field themselves. Then call list_connections: it shows the connection and whether it is connected, and list_connection_tools what it offers. Tell them what follows: reads are open to their assistants and, on a team connection, to Brain's agents; writes stay off until they turn one on in the connection's page.
- Drafting with their assistant: Allow Google actions, on their page in the workspace (the account menu, You in …), lets you draft replies in their Gmail and edit their Google Docs; sending a draft needs a separate yes.
- A first procedure for a job they repeat: step 9.

## Step 9: a first procedure

A procedure is how the team does one job, written once and then found and followed by every member's assistant and by Brain's agents, which never rewrite it. One good procedure is worth more than several thin ones: take the job from step 4, or ask which job they would most like to stop explaining.

1. Search with purpose procedure, so you complete one that exists rather than write a second. Read the procedure the rules name for writing one (`maintain-procedure` unless they changed it) and follow it.
2. Interview them about the job, a question or two at a time: when it applies and when not, what arrives to start it, the steps and the decisions in them, the templates or documents they start from, what a good result looks like, the traps, who must agree before it goes out. Ask where a recent finished one sits, in the mail you read or in Drive; read it, as evidence of how they really do it.
3. Write it with create_page: a kind the rules' `purposes` call a procedure (`skill` unless they changed it), a plain title naming the job, and in the frontmatter `when:` one sentence saying when it applies and when not, which is how every assistant finds it in read_context. Then only what is specific to this job: steps, decisions, checks, links to its templates and to the finished example. Their words are the evidence; show them the text first.
4. If the job is one of the rules' kinds of work under `matters`, offer to name the procedure there, so an agent picking up such a matter reads how first: `matters: [quotes: "[[send-a-quote]]", customer orders]`. It is a change to the rules: show it and write it only on their yes.
5. Try it now on a real case: one of the matters you opened in step 7, or a request from the month of mail. Follow the procedure step by step as written, not from memory, and show the result. If they allowed Google actions and the result is a reply, offer a Gmail draft they review and send; never send it. What the trial shows was missing or wrong goes back into the procedure, with a change note, on their yes.

Then say what the procedure covers, what the trial proved and what it did not, and that any assistant on the team now follows it.

## End

Sum up in a few lines: what is set up, what the brain is doing now, what is still to do. Suggest what to ask you tomorrow, once the first pages are written: where a customer stands, what is waiting on them. Remind them that pasting the same sentence picks up where they left off.