A seat is a chair, not a vendor. Four are set up for you — Claude, ChatGPT, Gemini and Grok — and each can hold something else.
Re-pointing a seat
Open the seat panel and pick a different model. It can be:
- Another model from the same maker. A cheaper one for reviewing, a stronger one for leading.
- Anything in the OpenRouter catalogue. Hundreds of models, including open-weight ones. The seat then wears that model's name and its maker's mark, so the transcript says who actually answered rather than which chair it sat in.
- Your own endpoint. Any OpenAI-compatible URL, including something you host. The the help centre article has the details.
Each seat keeps two choices: the model it uses when it leads and the one it uses when it reviews. Reviewing is the cheaper job and does not need the same model.
CLI seats
If you already pay for Claude Code, Codex, the Gemini CLI or Grok's CLI, a seat can run through that instead of through an API key. The app detects the ones you have installed and signed in. Those turns cost you nothing extra here — your subscription already paid for them — and the receipt says so in words rather than showing a misleading zero.
Guests
@guest invites one model for one answer without giving it a seat. It answers, wearing its maker's mark, and is not in the room afterwards.
Which model actually leads well
Rate answers with the thumbs. Over a few weeks the operator view shows which seat and which model earn their place on your kind of work, which is the question the leader and reviewer pickers are really asking you.