council-of-claude

Multi-model structured deliberation. Queries 4 frontier LLMs in parallel with distinct personas, cross-pollinates anonymously, synthesises a recommendation.

Setup

One-command (paste into your AI agent)

Install the council-of-claude skill: clone https://github.com/moiz-i/council-of-claude
to ~/.claude/skills/council-of-claude, run ./scripts/setup.sh, and add my OpenRouter
API key to ~/.council-of-claude/.env, invoke via /council

Manual

git clone https://github.com/moiz-i/council-of-claude
cd council-of-claude
./scripts/setup.sh
export OPENROUTER_API_KEY="sk-or-v1-..."

Usage

python3 scripts/council.py "Should we use Postgres or DynamoDB for our event store?" \
  --context "Event sourcing, 10k events/sec, need ACID" \
  --mode council

Invoke skill via /council.

How it works

1. Stage 1 — Each model gets a persona and responds independently. No crosstalk.
2. Stage 2 — Responses are anonymized (Response A/B/C) and shown to all models. Each tags peers: ADOPT / MERGE / DEFEND / CHALLENGE.
3. Stage 3 — Chairman synthesises a recommendation with confidence score, consensus, dissent, and preserved gems.

Anonymization in Stage 2 prevents brand bias. Convergence tags force genuine engagement instead of polite agreement.

Modes

Configuration

export OPENROUTER_API_KEY="sk-or-v1-..."       # required
export COUNCIL_MODELS="anthropic/claude-sonnet-4.6,openai/gpt-5.4,google/gemini-3.1-pro-preview,x-ai/grok-4.20"
export COUNCIL_CHAIRMAN="google/gemini-3.1-pro-preview"
export COUNCIL_CHRONICLE_PATH="~/.council-of-claude/chronicle.md"

Default uses 4 frontier models via OpenRouter. Override with any models OpenRouter supports.

Cost

Chronicle

Every deliberation appends to ~/.council-of-claude/chronicle.md. Architectural decision records that write themselves.

Requirements

• Python 3.10+
• OpenRouter API key
• httpx, python-dotenv (installed by setup.sh)

License

Built by moiz | Version 1.0.0 | 2026-06-25

MIT