OpenClaw Consensus MCP server

OpenClaw Consensus MCP

Multi-model consensus inside MCP clients: compare answers, surface disagreement, and escalate only when needed.

OpenClaw Consensus MCP wraps the OpenClaw Consensus API as three Model Context Protocol tools. It is designed for workflows where a maintainer wants a second opinion before accepting a risky answer, review summary, or routing decision.

<!-- mcp-name: io.github.MICONNM/openclaw-consensus-mcp -->

What it does

OpenClaw runs the same prompt across multiple models, then returns:

• a consensus answer with confidence and model response metadata,
• a disagreement heuristic derived from the deep consensus response, and
• a cheapest route recommendation that tries smaller model sets before escalating.

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

Why consensus?

A single model can produce a confident but incorrect answer. Comparing multiple responses does not prove correctness, but disagreement is a useful signal that a maintainer should review the output more carefully.

Install

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

You also need a RapidAPI key for the OpenClaw Consensus API:

Set it in your environment:

export RAPIDAPI_KEY="your-rapidapi-key"

Claude Desktop config

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

{
  "consensus": "string",
  "confidence": 0.0,
  "models_responded": 5,
  "votes": []
}

The consensus tool returns the upstream API response as-is. Fields may expand as the endpoint evolves.

disagreement_score(prompt)

How much the deep consensus response disagrees on a prompt.

Returns

{
  "disagreement": 0.0,
  "confidence": 1.0,
  "models_responded": 9,
  "votes": []
}

cheapestroute(prompt, targetquality=0.85)

Try fast, balanced, and deep modes in order until the confidence threshold is met.

Returns

{
  "selected_mode": "balanced",
  "models_used": 5,
  "confidence": 0.9,
  "answer": "string"
}

Local development

git clone https://github.com/MICONNM/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

See CONTRIBUTING.md for the development workflow and docs/maintainer-workflow.md for triage, review, security, and release responsibilities.

Limitations

Shortened. The full README is on GitHub.