Your whole department, not a seat licence. Every team gets a hive of its own — five agents to start, twenty-five at the top, its own Planner, its own records and its own budget — and the organisation pays once.
A person in Finance cannot read Sales's records, not because a policy says so but because they are in a different store. You add a team the way you add a person, hand the lead a panel link, and read the month's bill for all of it in one place.
Teams is built on the same store as everything else here: one binary, no dependencies, and every fact traceable to the bot and model that wrote it. The isolation between teams is the storage engine's, not a promise in a policy document.
Its own records, its own keys, its own Planner and its own model policy. Nothing in one team's hive is readable from another — the store is the boundary.
five agents to start, twenty-five at the topYour identity provider signs your people in — Entra ID, Okta, Google Workspace, Keycloak — and we receive an address and nothing else. No password of theirs reaches us.
sign-in, not another password to manageEach team's spend against its own budget, with your cost centre against every line — as a page, a CSV, or a spreadsheet your finance contact can open and sum.
you pay for what the agents use, at $1 per million tokensThe unit of work here is a team with a recurring job: the same afternoon, every week, done by somebody who would rather be doing something else. The team's Planner builds the agents that do it, the agents keep a record of what they did, and the lead reads the record rather than asking around.
Invoice chasing from your own ledger, one reminder per unpaid invoice with the drafts waiting for a person to send. Expenses and receipts sorted into categories with the receipt behind every line. Cashflow watched, and the month reconciled rather than guessed.
Enquiries answered from your own price list, then followed up twice and no more. Complaints given a timeline before a reply, and a reply that apologises for the right thing. Onboarding a new customer written out as steps somebody follows.
Tenders watched every morning with the notice, the deadline and the address it was read on. The rota checked against everyone's availability and a gap covered before it bites. Suppliers and competitors watched, with what changed and when.
Meeting notes turned into actions with an owner and a date, and decisions kept apart from discussion. Policies drafted from how you actually work, with every judgement call left as a question rather than filled in with a plausible default.
A campaign written — the offer, the calendar, the letters — and left for the team to approve, because a marketing agent here has no way to send at all until somebody gives it one. The Christmas card, the birthday count, the anniversary, all held to the same rules.
The list above is drawn from Agent Possibilities, and it is a starting point rather than a limit. A team asks its Planner for what it needs in its own words; if the job needs a machine of yours or a key of yours, the Planner says which and where it goes.
Add a hive, hand out a panel and take it away again, name the cost centre, set the monthly budget, and choose whether a team runs on our models or its own OpenRouter account. In the panel you already use, under Teams.
A team's monthly budget is checked where every spend is checked, so a runaway stops that team and nothing else. It refuses in words — which team, what it has spent, what its cap is — and it cannot be got round by using a different feature.
A page with a bar per team, a CSV for anything, and a real spreadsheet: one row per team with its cost centre, its budget and what it spent. Read from the same records the budget refuses on, so the document and the refusal cannot disagree.
OIDC against your own provider: your people sign in where they already do, we get an address, and a panel appears in the team's hive. The tokens are never stored, the domains are yours to allow, and every sign-in is written down.
Every action any agent takes is a record carrying the agent, the model, the time and the cost. "What did it do, what did it cost, who asked" is a query, and it is the same records the bill is built from.
Each agent has a declared doorway and the store refuses everything outside it, in writing. That is the answer to "what stops an agent seeing something it should not", and it is a stronger answer than a policy document.
Start with one team that has a boring, recurring job and a lead who will answer for it. Not IT, and not the whole company — the point of the first month is to have something to show the second team.
The organisation's account is opened, the directory is connected, and one team gets its hive. The lead is shown the panel for half an hour and builds their first agent by talking to their Planner — in their own words, about a job they already hate.
The agent runs. The bill goes to the lead and to whoever owns the budget, with the cost centre on it, and it is small enough to be a line rather than a conversation.
What the team did with it, what it cost, what to change, and which team should have one next. A $99 decision can be repeated without a tender, which is the whole reason the price is where it is.
Not yet, and we will not pretend otherwise. What we have is the architecture: the store refuses in writing, every agent has a declared doorway, every action is a record, and the subprocessor list is short. If certification is a gate for you, the same software runs on your own hardware and the question goes away.
Your records stay in your own hive's store. Model calls go to the provider named for that team — ours, or your own OpenRouter account — and every call is a record with the model and the size. A team that must not leave the building can be given models that run inside it.
Those are assistants per person, priced per person. This is agents with a record, a budget and a doorway, priced for the team however many people are in it. They can live alongside each other; we are not asking you to remove anything.
Then start where the work is not personal data: Finance, procurement, tender watching, supplier watch. bring them in before anything reads a mailbox, and we would rather be told no early than be switched off later.
Every record is yours and exportable, and a hive is a file. Backups are taken from the panel and downloaded by you. Leaving is a download, not a migration project.
A plan counts agents, not people, because agents are what cost anything. A hundred people may share one hive's memory.