The room has four chairs, one for each vendor: Claude, ChatGPT, Gemini and Grok. You manage them in Settings → Who's In The Room on every app. A seat can be switched on or off, and a seat that is off stays in the room but does not speak.
Two models per seat
Each seat holds two models. The Leader model is what the seat uses when it leads a discussion. The Reviewer model is what it uses when it reviews. Reviewing is three of every four calls and the smaller job, so by default the reviewer is the same maker's model one step below the lead — a cheaper model doing the more frequent work. You can set either to anything the seat's vendor offers.
Which seat leads
One seat leads a discussion; the others review. The lead is chosen in Settings and can be changed per discussion from the lead chip beside the composer. A seat that is not ready — no key, not installed, benched — is skipped, and the room says so rather than waiting on it.
Testing a seat
Every seat has a Test button. It makes one small call with the seat's current model and key, or on the Mac re-runs detection for a CLI seat, and reports what it found. When a seat is not behaving, Test is the first thing to press.
A seat holding another model
A chair does not have to hold its own vendor's model. Put any catalogue model or a custom endpoint in a seat and the seat is named for that model everywhere — the seat card, the @ menu, the message headers. See "Putting any model in a seat".