llm-deliberate-mcp 0.1.0

MCP server for multi-LLM deliberation with anonymized peer review

LLM Council MCP Server

Multi-LLM deliberation with anonymized peer review -- as an MCP tool.

Quick Start

1. Install

uvx llm-deliberate-mcp
# or
pip install llm-deliberate-mcp

2. Get an API Key

Get an OpenRouter API key (provides access to all major LLM providers).

3. Add to Your MCP Client

{
  "mcpServers": {
    "llm-deliberate": {
      "command": "uvx",
      "args": ["llm-deliberate-mcp"],
      "env": {
        "OPENROUTER_API_KEY": "sk-or-v1-..."
      }
    }
  }
}

How It Works

User Question
    |
    v
Stage 1: 4 LLMs answer independently (parallel)
    |
    v
Stage 2: Models peer-review anonymized responses
    |       (Response A, B, C, D -- no model names)
    v
Stage 3: Chairman synthesizes the best final answer
    |
    v
Final Answer + Metrics + Rankings

When Deliberation Helps (and When It Doesn't)

Works best for:

• Architecture decisions with genuine tradeoffs
• Complex problems requiring multiple perspectives
• Code review and design critique
• Strategic planning and risk assessment
• Security review of authentication flows

Works poorly for:

• Simple factual questions (use a single model)
• Calculations and data lookups (use a single model)
• Tasks under time pressure (30-90 sec overhead)
• Repetitive/routine tasks

Honesty note: Research (DeliberationBench, Jan 2026) shows that for simple QA tasks, a single strong model outperforms multi-model deliberation. LLM Council is designed for complex decisions where multiple perspectives add genuine value.

Configuration

• OPENROUTER_API_KEY (required)
• COUNCIL_MODELS (optional, comma-separated)
• CHAIRMAN_MODEL (optional)
• COUNCIL_TIMEOUT (optional, seconds)
• COUNCIL_MAX_TOKENS (optional)

Client Configuration Examples

Claude Desktop

{
  "mcpServers": {
    "llm-deliberate": {
      "command": "uvx",
      "args": ["llm-deliberate-mcp"],
      "env": {"OPENROUTER_API_KEY": "sk-or-v1-your-key-here"}
    }
  }
}

Claude Code

{
  "mcpServers": {
    "llm-deliberate": {
      "command": "uvx",
      "args": ["llm-deliberate-mcp"],
      "env": {"OPENROUTER_API_KEY": "sk-or-v1-your-key-here"}
    }
  }
}

Cursor

{
  "mcpServers": {
    "llm-deliberate": {
      "command": "uvx",
      "args": ["llm-deliberate-mcp"],
      "env": {"OPENROUTER_API_KEY": "sk-or-v1-your-key-here"}
    }
  }
}

Generic MCP client (Python)

import asyncio
from mcp import ClientSession, StdioServerParameters
from mcp.client.stdio import stdio_client

async def main():
    server_params = StdioServerParameters(
        command="uvx",
        args=["llm-deliberate-mcp"],
        env={"OPENROUTER_API_KEY": "sk-or-v1-your-key-here"},
    )

    async with stdio_client(server_params) as (read, write):
        async with ClientSession(read, write) as session:
            await session.initialize()
            result = await session.call_tool(
                "council_deliberate",
                arguments={"prompt": "Should we use microservices or a monolith?"},
            )
            print(result)

asyncio.run(main())

API Reference

• council_deliberate(prompt, models=None, chairman=None, depth="standard", contract_stack=None)
• council_status()
• council_configure(stage1_models=None, stage2_models=None, chairman_model=None, timeout_seconds=None)

Cost Transparency

Each deliberation uses roughly 10-15 API calls and typically costs about $0.03-$0.15 depending on model mix and prompt length. Responses include cost_estimate_usd.

Self-Learning

Deliberation metadata is stored locally in ~/.llm-council/deliberations.db for pattern learning. No learning data is sent externally.