SYNOD - Sistema de Deliberação Multi-Agent
quantsquirrel/claude-synod-debate
Convoque um conselho: Claude, Gemini e OpenAI debatem para respostas melhores
Ver no GitHubPlataforma que convoca um conselho de três LLMs diferentes (Claude, Gemini, OpenAI) para debater problemas através de deliberação estruturada em três atos: solução independente, crítica adversarial e veredito final. Reduz confiança excessiva e melhora qualidade das decisões via triangulação de perspectivas.
Instalação
npx convez-ai quantsquirrel/claude-synod-debate1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Análise de segurança e revisão de código crítico
- Decisões de negócio com alto risco (financeiro, legal, estratégico)
- Problemas complexos que exigem múltiplas perspectivas
Veredito

Projeto inovador que explora deliberação multi-agent como estratégia para superar limitações de LLMs individuais. Bem executado tecnicamente, mas requer múltiplas chaves de API e setup complexo — ideal para equipes que priorizam qualidade de resposta sobre simplicidade operacional.
README
A tradução para o português deste README ainda está na fila.
When one AI isn't enough, convene the council.
😵💫 Single LLMs are overconfident → ⚔️ Make them debate → ✅ Better decisions
🎭 THE THREE ACTS
Every deliberation follows the same dramatic structure
%%{init: {'theme': 'dark', 'themeVariables': { 'primaryColor': '#1e3a5f', 'secondaryColor': '#4a1d1d', 'tertiaryColor': '#1a3d1a'}}}%%
flowchart TB
subgraph ACT1["🎬 ACT I · SOLVE"]
G1["🔵 Gemini → Solution A"]
O1["🟢 OpenAI → Solution B"]
end
subgraph ACT2["⚔️ ACT II · CRITIQUE"]
G2["🔵 Gemini attacks B"]
O2["🟢 OpenAI attacks A"]
end
subgraph ACT3["⚖️ ACT III · VERDICT"]
C["🟠 Claude → Final Answer"]
end
ACT1 --> ACT2 --> ACT3
style ACT1 fill:#1e3a5f,stroke:#3b82f6,stroke-width:2px,color:#fff
style ACT2 fill:#4a1d1d,stroke:#ef4444,stroke-width:2px,color:#fff
style ACT3 fill:#1a3d1a,stroke:#22c55e,stroke-width:2px,color:#fff
| Act | What Happens | Why It Matters |
|---|---|---|
| I | Independent solutions emerge | No groupthink — maximum diversity |
| II | Cross-examination begins | Weaknesses exposed — biases challenged |
| III | Adversarial refinement | Best ideas survive scrutiny |
⚡ 60-SECOND SETUP
# 1️⃣ Clone the repo
git clone https://github.com/quantsquirrel/claude-synod-debate.git
cd claude-synod-debate
# 2️⃣ Local model bridges
# - log in to Antigravity CLI (`agy`)
# - run CLIProxyAPI on localhost:8317
# 3️⃣ Run setup (installs deps, configures CLI tools, tests models)
/synod-setup
# 4️⃣ Summon the council
/synod review Is this authentication flow secure?
That's it. The council convenes automatically.
🔧 INITIAL SETUP TEST
Verify your models work before deliberating
/synod-setup
| Check | What It Does |
|---|---|
| CLI | Verifies all 7 provider CLIs exist |
| API Keys | Checks all provider API keys |
| Response Time | Tests each model with 120s timeout |
| Classification | Labels models: ✓ Recommended / ✓ Usable / ⚠ Slow / ✗ Failed |
[Synod Setup] 초기 설정을 시작합니다...
Step 0/3: Python 의존성 확인
✓ openai 설치됨
✓ httpx 설치됨
Step 1/3: CLI 도구 설치 (~/.synod/bin)
✓ agy-cli 설치됨
✓ cliproxy-cli 설치됨
✓ gemini-3 설치됨 (legacy fallback)
✓ openai-cli 설치됨 (legacy fallback)
Step 2/3: 로컬 세션/프록시 확인
✓ Antigravity OAuth/session (설정됨)
✓ CLIPROXY_API_KEY (optional; CLIProxyAPI localhost:8317) (설정됨)
Step 3/3: 모델 응답 시간 측정 (타임아웃: 120초)
Provider Model Latency Status
───────────────────────────────────────────────
gemini 3.5-flash 3.2초 ✓ 권장
openai gpt55fast 2.8초 ✓ 권장
[완료] 2/2 모델 사용 가능
Synod를 사용할 준비가 되었습니다!
🤖 SUPPORTED PROVIDERS
v3.0: Now supporting 7 AI providers
| Provider | CLI | Best For | Status |
|---|---|---|---|
| 🔵 Gemini | agy-cli | Antigravity Gemini 3.5 Flash | Required |
| 🔵 Gemini legacy | gemini-3 | Previous direct API defaults | Fallback only |
| 🟢 OpenAI | cliproxy-cli | CLIProxyAPI / ChatGPT Pro OAuth | Required |
| 🟢 OpenAI legacy | openai-cli | Previous direct API defaults | Fallback only |
| 🟣 DeepSeek | deepseek-cli | Math, reasoning (R1) | Optional |
| ⚡ Groq | groq-cli | Ultra-fast inference (LPU) | Optional |
| 🌐 OpenRouter | openrouter-cli | Multi-model fallback | Recommended |
| 🔶 Grok | grok-cli | 2M context window | Opt-in |
| 🟠 Mistral | mistral-cli | Code, European deployment | Opt-in |
# Optional: Add more providers to your council
export DEEPSEEK_API_KEY="your-deepseek-key" # DeepSeek R1
export GROQ_API_KEY="your-groq-key" # Groq LPU
export OPENROUTER_API_KEY="your-openrouter-key" # OpenRouter (Recommended)
# Opt-in Providers (requires explicit activation)
# Grok (2M context window)
export SYNOD_ENABLE_GROK=1
export XAI_API_KEY="your-xai-key"
# Mistral (code specialization)
export SYNOD_ENABLE_MISTRAL=1
export MISTRAL_API_KEY="your-mistral-key"
🎯 FIVE MODES OF DELIBERATION
Choose your council configuration
| Mode | Summon When... | Configuration | |
|---|---|---|---|
| 🔍 | review | Analyzing code, security, PRs | Gemini 3.5 Flash ⚔️ CLIProxy gpt55fast |
| 🏗️ | design | Architecting systems | Gemini 3.5 Flash ⚔️ CLIProxy gpt55fast |
| 🐛 | debug | Hunting elusive bugs | Gemini 3.5 Flash ⚔️ CLIProxy gpt55fast |
| 💡 | idea | Brainstorming solutions | Gemini 3.5 Flash ⚔️ CLIProxy gpt55fast |
| 🌐 | general | Everything else | Gemini 3.5 Flash ⚔️ CLIProxy gpt55fast |
# Code review
/synod review "Is this recursive function O(n) or O(n²)?"
# System design
/synod design "Design a rate limiter for 10M requests/day"
# Debugging
/synod debug "Why does this only fail on Tuesdays?"
# Brainstorming
/synod idea "How do we reduce checkout abandonment?"
📜 ACADEMIC FOUNDATION
Not just another wrapper — peer-reviewed deliberation protocols
| Protocol | Source | What Synod Implements |
|---|---|---|
| ReConcile | ACL 2024 | 3-round convergence (>95% quality gains) |
| AgentsCourt | arXiv 2024 | Judge/Defense/Prosecutor structure |
| ConfMAD | arXiv 2025 | Confidence-aware soft defer |
| Free-MAD | Research | Anti-conformity instructions |
| SID | Research | Self-signals driven confidence |
Synod calculates trust using the CortexDebate formula:
Credibility × Reliability × Intimacy
Trust Score = ────────────────────────────────────────
Self-Orientation
| Factor | Measures | Range |
|---|---|---|
| C | Evidence quality | 0–1 |
| R | Logical consistency | 0–1 |
| I | Problem relevance | 0–1 |
| S | Bias level (lower = better) | 0.1–1 |
Interpretation:
T ≥ 1.5→ Primary source (high trust)T ≥ 1.0→ Reliable inputT ≥ 0.5→ Consider with cautionT < 0.5→ Excluded from synthesis
📦 INSTALLATION
# Clone the repo
git clone https://github.com/quantsquirrel/claude-synod-debate.git
cd claude-synod-debate
# Prerequisites: agy logged in, CLIProxyAPI running on localhost:8317
# Optional override if your proxy uses a non-default key:
export CLIPROXY_API_KEY="your-local-proxy-key"
# Run setup inside Claude Code (auto-installs Python deps, creates CLI wrappers, tests models)
/synod-setup
Skills auto-load from plugin.json when you open Claude Code inside this directory. /synod-setup handles the rest: Python dependencies (openai, httpx), CLI tool wrappers in ~/.synod/bin/, local auth validation, and model connectivity testing.
git clone https://github.com/quantsquirrel/claude-synod-debate.git
cd claude-synod-debate
pip install openai httpx
# Create CLI wrappers and test models
python3 tools/synod-setup.py
# Optional if CLIProxyAPI uses a custom local key
export CLIPROXY_API_KEY="your-local-proxy-key"
# Optional
export SYNOD_SESSION_DIR="~/.synod/sessions"
export SYNOD_RETENTION_DAYS=30
🔒 COMPATIBILITY
| Environment | Status | Notes |
|---|---|---|
| bash | ✅ | Fully supported |
| zsh | ✅ | Fully supported (v3.0.1+) |
| MCP Plugins | ✅ | Guard directives prevent routing interception |
| OMC (oh-my-claudecode) | ✅ | CODEX-ROUTING opt-out built-in |
Synod executes external models (Gemini, OpenAI) exclusively via CLI tools (agy-cli, cliproxy-cli, legacy fallback gemini-3/openai-cli). If your environment includes MCP routing plugins that redirect model calls through ask_codex or ask_gemini, Synod's built-in defense-in-depth guards prevent interception:
allowed-toolsfrontmatter — Schema-level restriction excludes MCP tools- Markdown directives — Explicit prohibition in skill entry point and Phase 0/1
- Automated tests — CI validates guard presence against configuration drift
No additional configuration needed — protection is automatic.
🗺️ ROADMAP
- MCP Server — Native Claude Code integration
- VS Code Extension — GUI for debate visualization
- Knowledge Base — Learning from debate history
- Web Dashboard — Real-time debate monitoring
- More LLMs —
Llama, Mistral, Claude variantsv3.0: 7 providers supported!
🤝 JOIN THE COUNCIL
Issues · Discussions · Contributing
@software{synod2026,
title = {Synod: Multi-Agent Deliberation for Claude Code},
author = {quantsquirrel},
year = {2026},
url = {https://github.com/quantsquirrel/claude-synod-debate}
}
MIT License · Copyright © 2026 quantsquirrel
Built on the shoulders of ReConcile · AgentsCourt · ConfMAD · Free-MAD · SID
"In the multitude of counselors there is safety." — Proverbs 11:14