Put them in a room and let them argue.
Any model can take a seat. Each gets a role — sceptic, researcher, adversary, auditor — and reads what the others said before it speaks.
Four steps, no configuration screen.
Set the request
One clear brief. Choose whether you want discussion, a recommendation, or a forced vote with dissent recorded.
Seat the panel
Any model, any role. A panel of Claude, Gemini and DeepSeek is a legitimate configuration.
Watch them go in order
Each seat sees every answer before it. Round two is genuinely different from round one because the panel has read itself.
Steer it
Pause, interject, redirect, or fork a round to explore the path not taken.
Seats
2–8
Typical cost
About 41 credits per run
Best for
Decisions where being wrong is expensive and consensus is suspicious
What it does and does not do.
The agreement signal shows where the panel overlapped. It is a navigation aid and nothing more — models share training data and share blind spots, so consensus is weak evidence. What is useful is the shape of the disagreement: which claims split the room, and on what grounds.
When you ask for a recommendation, dissent is recorded by name rather than averaged away. A minority view that turns out to be right is the whole reason to run a panel instead of asking one model twice.
The other tools.
See it running.
The demo workspace shows this tool on sample data. No account, no card.
Open the demo