LLM CouncilCAUTION

Multi LLM deliberation council: query frontier models in parallel, then combine their responses through structured protocols (voting, debate, synthesis, critique, red teaming, verification). Ships as an MCP server and Claude Code skill.

Architecture

flowchart TD
A[User / Claude Code] -->|MCP stdio| B[LLM Council Server]
B --> C{Protocol Router}
C -->|vote| D[Parallel Query + Peer Review]
C -->|debate| E[Multi Round Debate + KS Stopping]
C -->|synthesize| F[Fan Out + Chairman Synthesis]
C -->|critique| G[Peer Critique]
C -->|redteam| H[Adversarial Red Team]
C -->|mav| I[Multi Agent Verification]

D --> J[OpenAI GPT 5.4]
D --> K[Gemini 2.5 Pro]
D --> L[Claude Sonnet 4.6]

E --> J
E --> K
E --> L

F --> J
F --> K
F --> L
F -->|synthesis| M[Chairman Model]

subgraph Providers
J
K
L
end

subgraph Broker [Peer Discovery :7899]
N[Register] --> O[Peers]
P[Messages] --> O
end

B -.->|optional| Broker

Features

• Vote. Each model answers independently, then all models anonymously rank each other. Winner selected by first place votes with confidence score.
• Debate. Multi round argumentation with optional KS statistic adaptive stopping. Chairman synthesizes the final round into a consensus answer.
• Synthesize. Fan out to all models, then a chairman model produces an authoritative synthesis combining the best elements.
• Critique. Models answer, then peer review each other's responses with structured feedback.
• Red Team. Adversarial variant of critique that actively probes for flaws, hallucinations, and failure modes.
• MAV (Model as Verifier). Cross checks a candidate answer using multiple models. Each model scores the candidate, highest scored response becomes verified output.

MCP Tools