3-AI Debate
topprismdata/3-ai-debate
Conselho de IA multi-provider: Claude + Gemini + MiniMax com protocolos de síntese.
Ver no GitHubServidor MCP que executa um conselho deliberativo entre três CLIs de IA (Claude, Antigravity/Gemini, MiniMax) em paralelo e sintetiza respostas via protocolos estruturados como debate, votação e síntese. Utilitário para comparação qualificada de múltiplos modelos sem API keys, usando autenticação CLI nativa.
Instalação
npx convez-ai topprismdata/3-ai-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
- Comparar respostas de múltiplos modelos para decisões críticas
- Debate estruturado entre perspectivas de diferentes LLMs
- Síntese automática de consenso e divergências entre modelos
Veredito

Ferramenta sólida para equipes que já usam múltiplas CLIs de IA e precisam de deliberação estruturada. Reduz custo de API calls ao aproveitar autenticação nativa, mas requer setup das três CLIs instaladas localmente.
README
A tradução para o português deste README ainda está na fila.
3-AI Debate
Multi-CLI deliberation council: send a prompt to three AI CLIs in parallel (Claude, Antigravity, MiniMax) and combine their responses through structured protocols. Ships as an MCP server and Claude Code skill.
Fork of rachittshah/llmcouncil — replaced API-based providers (OpenAI/Anthropic/Gemini SDK) with CLI-based providers (
claude,agy,mmx).
Demo
import { runCouncil } from "3-ai-debate";
const result = await runCouncil({
question: "微服务架构值得采用吗?用 100 字内回答",
config: {
models: [
{ provider: "claude-cli", model: "claude", label: "ModelA" },
{ provider: "agy", model: "gemini-3.5-flash", label: "ModelB" },
{ provider: "mmx", model: "MiniMax-M2.7", label: "ModelC" },
],
protocol: "synthesize",
},
});
Real output:
=== 协议: synthesize
=== 参与模型:
[ModelA] claude-cli/claude - 21493ms
[ModelB] agy/gemini-3.5-flash - 11091ms
[ModelC] mmx/MiniMax-M2.7 - 12426ms
=== 综合答案 ===
**Synthesized Answer:**
**共识:** 三位专家核心定义一致——404 表示服务器无法找到请求的资源。
**分歧:** 列举的成因略有差异(URL 错误 / 链接失效 / 页面删除),
但本质相同,均为客户端请求的资源在服务器上不可用。
**最终答案(40 字):**
> HTTP 404 表示服务器未找到请求的资源,常见于 URL 错误、
> 页面已删除或链接失效。
**无显著异议。**
With protocol: "debate", the chairman explicitly reasons about the trade-offs between opinions:
=== 协议: debate
=== 最终综合 ===
**合成回答:**
值得采用,但绝非银弹。**核心价值**是支持多团队并行开发与独立部署,
而非单纯解决扩容;当业务迭代快、模块边界清晰、团队具备 DevOps
能力时收益最大。**否则**应从模块化单体起步,警惕数据所有权分裂、
分布式事务等隐性成本,避免为拆而拆。
**裁决说明:**
- **三方共识**:非银弹、有条件、需匹配团队与业务成熟度。
- **采纳 B/C**:将"多团队并行 + 独立部署"定位为微服务的本质价值
- **保留 A 的关键补充**:模块化单体作为更稳妥的起点
- **显化分歧**:A 重"成本/起点",B/C 重"组织/速度"
Why CLI instead of API?
- No API key juggling — uses your existing CLI auth (Antigravity IDE login, MiniMax config, Claude Code session)
- No usage cost concerns — all three are subscription-based
- What you see is what you get — the same model + prompt combination your terminal would run
The 3 CLIs
| CLI | Model | Setup |
|---|---|---|
claude | Claude Opus 4.7 / Sonnet 4.6 | Claude Code |
agy | Gemini 3.5 Flash | Antigravity IDE — login once, CLI inherits |
mmx | MiniMax-M2.7 | mmx-cli from npm, see below |
Install
# Claude Code (already installed if you're using Claude Code)
npm install -g @anthropic-ai/claude-code
# Antigravity CLI (preinstalled with Antigravity IDE)
# or: https://github.com/google-antigravity/antigravity-cli
# MiniMax CLI
npm install -g mmx-cli
mmx auth login --api-key <your-key>
Protocols
| Protocol | Behavior |
|---|---|
vote | All 3 answer independently, then anonymously rank each other. Winner by first-place votes. |
debate | Multi-round argumentation. Chairman synthesizes consensus. |
synthesize | Fan out, then chairman produces an authoritative synthesis. |
critique | Peer review of each response. |
redteam | Adversarial probe for flaws, hallucinations. |
mav | Model-as-Verifier cross-checks a candidate answer. |
Quick Start
As MCP server (Claude Code)
# Build
npm install
npm run build
# Configure
export CLAUDE_CLI_PATH=$(which claude)
export AGY_CLI_PATH="$HOME/.local/bin/agy" # or wherever agy is
export MMX_CLI_PATH="$HOME/npm-global/bin/mmx"
# Add to Claude Code
claude mcp add three-ai-debate node /path/to/3-ai-debate/dist/server.js
Then in Claude Code:
> Use council_deliberate to answer "should we use microservices?" protocol=synthesize
As a library
import { runCouncil } from "3-ai-debate";
const result = await runCouncil({
question: "What's the best cache eviction policy?",
config: {
models: [
{ provider: "claude-cli", model: "claude", label: "ModelA" },
{ provider: "agy", model: "gemini-3.5-flash", label: "ModelB" },
{ provider: "mmx", model: "MiniMax-M2.7", label: "ModelC" },
],
protocol: "synthesize",
},
});
console.log(result.synthesis);
Architecture
Same as the upstream llmcouncil, but providers are CLI-based:
flowchart TD
A[User / Claude Code] -->|MCP stdio| B[3-AI Council Server]
B --> C{Protocol Router}
C -->|vote| D[Parallel Query + Peer Review]
C -->|debate| E[Multi Round Debate]
C -->|synthesize| F[Fan Out + Chairman Synthesis]
C -->|critique| G[Peer Critique]
C -->|redteam| H[Adversarial Red Team]
C -->|mav| I[MAV Verification]
D --> J[claude CLI]
D --> K[agy CLI]
D --> L[mmx CLI]
E --> J
E --> K
E --> L
F --> J
F --> K
L
F -->|synthesis| M[Chairman: claude]
subgraph Providers
J
K
L
end
Configuration
CLI Paths
If your CLIs aren't on PATH, set the env vars:
export CLAUDE_CLI_PATH=/custom/path/to/claude
export AGY_CLI_PATH=/custom/path/to/agy
export MMX_CLI_PATH=/custom/path/to/mmx
Or pass to the provider constructor (advanced).
What about the broker?
The upstream llmcouncil has a peer-discovery broker for cross-network council runs. This fork drops it — all 3 CLIs run on the same machine.
Differences from upstream
| Aspect | llmcouncil (upstream) | 3-ai-debate (this fork) |
|---|---|---|
| Providers | OpenAI / Anthropic / Gemini SDK | CLI spawn (claude / agy / mmx) |
| Token tracking | Yes | No (CLIs don't expose) |
| Pricing | Yes | No (subscription-based) |
| Broker (peer discovery) | Yes | No (single-machine) |
| Models | GPT-5.4, Gemini 2.5 Pro, Claude Sonnet 4.6 | Claude, Gemini 3.5 Flash, MiniMax-M2.7 |
License
MIT (inherited from upstream)