Chat assistant

A new organization starts with a checklist of nine steps: a registry, a first prompt, an environment, deployment provider credentials, an LLM provider, an agent, a deployment, a fleet and a governance policy. You can work through it in two ways, and switch between them whenever you like.

Choose Begin (or Continue) in the checklist and pick how to set up:

• I'll do it myself. Go through the steps one by one. Each step opens a short form.
• Chat assistant. A panel opens where you describe what you are building, and the assistant proposes the setup that fits.

Only organization admins can use the checklist's actions and the assistant.

How the assistant works

The assistant asks one question at a time. It looks at what is already set up, so it never suggests something that exists. If you create something by hand between two messages, it notices on the next turn, and the progress bar at the top of the panel updates as steps are completed.

Nothing is created until you say so. When the assistant wants to change your setup, it shows a proposal:

• Accept creates it.
• Edit lets you change the details first, then accepts.
• Reject drops it. The assistant carries on from there.

Every proposal also links to the page where you can do the same thing yourself. Once a proposal is accepted, the card shows what was created and where to find it.

Credentials

The assistant never asks you to type a key into the chat. If you paste one anyway, it is removed from the message before it is saved. When a step needs credentials, such as a deployment provider key or an LLM provider key, the assistant shows a secure form: the same form the checklist uses, which sends the value straight to the platform. You can also store a secret yourself under Secrets and refer to it as ${SECRET_NAME}.

Closing and coming back

Press Escape or use the close button to hide the panel. Your conversation is saved, and opening the panel again picks up where you left off. The manual checklist stays available the whole time.

Choosing the model

The assistant always runs on a model provided by the platform, not on your own LLM provider, and its use counts against the organization's platform token allowance. Platform administrators can pick the model with SYSTEM_LLM_ONBOARDING_MODEL_KEY.

Limits and audit

Each conversation has a budget: up to 80 messages and 750,000 model tokens. A message counts when it is sent, even if the reply fails. An organization has a daily allowance too: up to 120 messages and 800,000 tokens across all its conversations in 24 hours, closed ones included, so starting a new conversation does not give a fresh budget. When a budget is used up, the assistant says so and stops answering; the checklist keeps working, so you can finish the remaining steps by hand.

Every proposal you accept is recorded in the organization's audit log, with who accepted it and what it created.