Quorum with Claude Code

Add Quorum to Claude Code with one command, then tell Claude Code when to convene a panel.

Claude Code calls Quorum as a tool. It sends a question to a panel of models from different labs and gets back one answer plus the places the panel split.

Set up

Run this in your terminal. Replace YOUR_QUORUM_API_KEY with a key from your organization's usage page. If you have no organization yet, create one.

claude mcp add --transport http quorum https://www.quorum.dog/mcp --header "Authorization: Bearer YOUR_QUORUM_API_KEY"

Quorum runs as a hosted server at https://www.quorum.dog/mcp. Authentication is the API key, sent as a Bearer token. API keys always bill your organization's API account.

To share the server with your team, commit a .mcp.json at the repository root. Each person's key is read from their environment, so it stays out of the diff.

{
  "mcpServers": {
    "quorum": {
      "type": "http",
      "url": "https://www.quorum.dog/mcp",
      "headers": { "Authorization": "Bearer ${QUORUM_API_KEY}" }
    }
  }
}

Check it works

Ask Claude Code to call list_modes. list_modes is free and returns the Modes your key can call. A response confirms the URL and the key.

Tell Claude Code when to convene

Claude Code reads CLAUDE.md at the start of a session. Add this block to it:

## Quorum

Quorum is connected as an MCP server. It convenes a panel of models from different labs.

When to convene: Convene a panel (deliberate) when the work is judgement rather than execution: a consequential assumption, options close enough that the numbers no longer separate them, a confident diagnosis that cannot be checked, a decision someone will have to defend, or "what have I missed". Do NOT convene one for lookups, syntax, formatting, conversions, arithmetic, or any answer that is a verbatim artifact to reproduce — a synthesis pass rewrites those and makes them worse.

Loop: estimate, deliberate, get_receipt, act

1. Call `estimate` first. It is free. Read `deliberation_value`.
2. If it says `likely_helps`, call `deliberate` with the model `QRUM:STAN`.
3. Call `get_receipt` with the returned `request_id` to see which models sat and where they split.
4. Report the disagreement before acting. Do not merge a split into one answer.

The tools

Quorum exposes 13 tools. The loop above uses these four.

Tool	Use
`estimate`	Quotes the whole call for the question (the engines, the platform fee and the Mode's fee) and says whether a panel is likely to help, free of charge.
`list_modes`	Lists the Modes you can call, with their prices and the providers behind them.
`deliberate`	Sends your question to a panel of models and returns where they agree and where they split.
`get_receipt`	Shows which model sat in each seat of a past deliberation, with judge scores, cost and latency.

The other tools cover testing and building Modes. The full list is in the API reference.

Questions worth a panel

A decision that is hard to reverse. Schema changes, auth models, a public API shape.
An answer Claude Code already gave. Pass the answer to the panel and ask what is wrong with it.
A risk you want named. get_receipt shows which seats flagged it.

Run estimate first. Keep the panel for questions with no single right answer.

Escalation router shows how to convene a panel only on hard questions. Connect Quorum to your tools covers every other client.

The paper

This guide is also a paper: Quorum with Claude Code (PDF).

More papers are on the whitepaper shelf.