A seat is a place at the table. There are four built in — Claude, ChatGPT, Gemini and Grok — and any of them can be re-pointed at a different model, including one from OpenRouter or your own endpoint. A seat you switch off does not take part and does not cost anything.
The lead
One seat leads. It answers first and, on the desktop app, it is the only seat that writes files or makes commits. That is deliberate: four models editing the same tree at once is a merge conflict, not a collaboration.
The lead is your account's setting, overridable per discussion from the seat panel, and per project if a project names its own. If the lead cannot answer — an outage, a bad key — the chair is handed to another seat under its own name, so the transcript never misattributes work.
The reviewers
Every other seat reviews. A reviewer reads everything said before it in the round, including the other reviewers, so the second one can disagree with the first rather than repeating it.
On the desktop app a reviewer has its own hands. It runs commands to check the lead's claim — your test suite, a type check, a grep for the thing it suspects — and its verdict carries what it ran. It cannot write files; a reviewer that could fix what it found would stop reporting it.
Directed turns
Start a message with a seat's name to ask that one seat only: "@gemini is this regex right?" One pass, one answer, no review — sometimes you want the quick thing.
Guests
Invite a model for a single answer without giving it a seat: type @guest and pick from the OpenRouter catalogue. It answers once, wearing its maker's mark, and is not part of the room after that. Useful when you want a specific model's opinion on one question without rearranging your setup.