Council
Ask several models the same thing. Compare, let them debate, or have them build together.
- Seats2–4 models
- ModesDebate, Build together
- Promptingnone added
- Opentitle bar or Ctrl+K
S15 · council
Seat the council
- Pick 2–4 models — each seat is a provider, a model and an account, from the usual model picker.
- Choose a mode — Debate or Build together.
- Ask once — type your question and press Start.
Want them to see your code? Attach workspace files, paste text, add images, or let them read the workspace.
Debate
- They answer one after another, in seat order.
- Each round, every model sees what the others said since its last turn, and answers again.
- Answers sit side by side when there’s room.
- Pick 1–5 rounds. Run them all, or pause after each round to steer.
- Talk back in the thread any time: Talk (each answers once) or another debate.
- No summary unless you ask. Choose Summarize… and pick a judge.
Side chats
- Open any model in its own pane and talk to it directly.
- The others don’t see it. Click Share to council under a reply to bring it into the thread.
Build together
Plan, spec, doc or patch — one shared piece of text:
- Propose — everyone drafts their own.
- React — everyone reads and responds to the others.
- Merge — the lead you picked merges the best into one result.
Code with agents — real changes in your workspace:
- Split — the lead proposes who does what; the others comment.
- Approve — you check the split, edit it if you like, then press Build it.
- Build — each model builds its part as a background agent, in its own worktree.
- Review — another model reviews each part, and it gets fixed if needed.
- Check — the lead makes sure everything works together.
Pure model output
- No system prompt, persona or debate rules. Each model gets your words as if you typed them into its own app.
- Between turns, a model only gets the others’ replies, behind a short lead-in.
- Subscription seats stay on the subscription. They never fall back to an API key.
Sessions and hand-off
- Every council is saved and listed on the page.
- Export saves it as Markdown.
- Hand off sends the result to Toad, or starts a background agent with it.
Note
Models on China-hosted endpoints can join a plain conversation, but never get workspace context and can’t build with agents. The same model running locally in Ollama is fine.