OpenClaw Consensus MCP

9-LLM consensus inside Claude — a hallucination guardrail you can call as a tool.

Production-grade 9-LLM consensus with disagreement scoring and cheapest-route picks to fight hallucinations.

What it does

OpenClaw runs the same prompt across 9 frontier LLMs (Claude, GPT, Gemini, Llama, Mistral, etc.), then returns:

• a consensus answer (with confidence + which models contributed),
• a disagreement score (high = your single LLM is probably about to hallucinate), and
• a cheapest route (pick the smallest model combo that still hits your quality bar).

This MCP server exposes those three capabilities as tools so Claude Desktop / Claude Code can call them mid-conversation.

Why consensus?

A single LLM can confidently make things up. Nine models rarely make up the same thing. When 9 models agree, you can trust the answer; when they fan out, you have a cheap, calibrated hallucination signal — before the user sees the wrong answer.

Install

pip install openclaw-consensus-mcp
# or
uv pip install openclaw-consensus-mcp

You also need a RapidAPI key for the OpenClaw Consensus API: https://rapidapi.com/yanmiayn/api/openclaw-consensus

Set it in your environment:

export RAPIDAPI_KEY="your-rapidapi-key"

Claude Desktop config

Add to ~/.claude/claude_desktop_config.json (macOS/Linux) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "openclaw-consensus": {
      "command": "openclaw-consensus",
      "env": {
        "RAPIDAPI_KEY": "your-rapidapi-key"
      }
    }
  }
}

For Claude Code:

claude mcp add openclaw-consensus -- openclaw-consensus

Tools

consensus(prompt, mode="balanced")

Get a 9-LLM consensus answer.

• prompt (string) — the question.
• mode (string, default balanced) — deep (9 models), balanced (5), or fast (3).

Returns

{
  "answer": "string",
  "confidence": 0.0,
  "models_used": ["claude-opus-4.7", "gpt-5.1", "..."],
  "disagreement": 0.0
}

disagreement_score(prompt)

How much the 9 models disagree on a prompt — a hallucination warning signal.

Returns

{
  "score": 0.0,
  "per_model": { "model-id": "answer summary" }
}

cheapest_route(prompt, target_quality=0.85)

Cheapest model combo that meets a quality threshold (0..1).

Returns

{
  "models": ["..."],
  "estimated_cost_usd": 0.0,
  "estimated_quality": 0.0
}

Local development

git clone https://github.com/yanmiayn/openclaw-consensus-mcp
cd openclaw-consensus-mcp
uv venv && source .venv/bin/activate
uv pip install -e ".[dev]"
pytest

Smoke-test the server with the official MCP Inspector:

npx @modelcontextprotocol/inspector openclaw-consensus

Publish

uv build
uv publish      # to PyPI
mcp-publisher publish   # to the official MCP Registry

License

MIT — see LICENSE.