The AI assistant

Every AI surface in Openbook — the assistant in your rooms, the build agent, a board's chat — draws from the same pool of models your organization has configured. This page is about where the assistant turns up day to day. The setup it depends on is one page over, in AI configuration.

What has to be true before it answers

Open Organization → AI self hosted and check the switch at the top. It decides whose key every model call is paid with, and there are two answers:

• Self hosted — your organization's own provider keys and its own models. Your provider bills you directly. Somebody has to add a key and activate a model before anything works.
• Managed by us — the platform's key and the platform's models. There is nothing to configure; the assistant works as soon as the organization is switched to it, provided the platform has a model configured.

Only the organization's owner can change either the mode or the models. Everyone else can read the page, because a model picker in a room needs the list.

The full walkthrough — the providers, what each of them can and cannot do, your own OpenAI-compatible server, and what each mode means for the bill — is in AI configuration.

The models the assistant can pick

Whatever mode you are in, the assistant picks from the active models in the list on that page, and prefers the starred default.

• Deactivating a model withdraws it everywhere at once, without touching individual spaces or rooms.
• Starring one makes it the default, and also activates it — a default nobody can pick is a trap.
• A model must support tools to be useful to the build agent: one that cannot call tools cannot read or edit a file.
• New models do not appear on their own. In self hosted mode you browse the provider and add them deliberately, so nobody ends up on a model nobody reviewed. In managed mode the platform decides, and the page shows you what that is.

Each model also carries a rate, a number from 1 to 100 left over from the retired credit system. Nothing charges against it any more — the token allowance counts raw tokens. See AI usage & the token allowance.

Where the assistant appears

The assistant works inside your rooms, with the content that is already there:

• Retro rooms — an AI summary of the session, so nobody has to take minutes.
• Kanban rooms — a board chat that answers questions about the board, and that can turn a project description into a backlog of tickets and a user-flow diagram. See Kanban.
• Gantt rooms — generate a whole plan from a text description.
• Feedback rooms — similar-post detection when somebody writes a request.

Two bigger AI features have their own pages: the build agent, which lives in its own room type and works against a repository, and the concierge, which builds workspaces for you.

If the assistant will not run, the refusal names the cause and the fix. The usual ones are:

• In self hosted mode — no provider key has been added at Organization → AI self hosted, or no model there is marked active.
• In managed mode — the platform has no model configured, which is a platform administrator's to fix at the Bookbag account service.
• In either mode — the model that is running cannot call tools, which the agent needs.

What this costs

In managed mode, everything the assistant does counts against the token allowance — fresh input, output and cached input, all of it, against the organization owner's account at the Bookbag account service. When the allowance is spent the assistant is refused with a sentence naming the figures, the day it turns over and the two ways out.

In self hosted mode nothing is metered at all. Your provider bills you directly, and no figure the assistant reaches will ever refuse it.

The allowance caps what the platform pays for, not what you may use. The full explanation, and the local guardrails that bound a run before the allowance needs to, is in AI usage & the token allowance.

Openbook's credits are gone. The assistant, the build agent and workflows used to be "metered but never charged"; there is no ledger to charge now, and in managed mode they are counted against the allowance like everything else. See What happened to credits.

Next steps

Once the mode is settled and a model is active, the assistant works immediately — there is no per-room setup. From here:

• Add an AI agent room and connect a repository to put the build agent to work.
• Review the per-space agent gates at Organization → AI self hosted if you want to restrict where and how heavily the agent runs.
• Point your own Claude or ChatGPT at a board with the Board MCP server.