Multi Reasoner — Development skill for Claude Code

Multi-Reasoner

An MCP server that sends one prompt to several frontier models at once (ChatGPT, Gemini, Grok, Mistral) and returns every answer. The point is reducing single-model bias: brainstorming, strategic analysis, code review, and decisions rest on multiple independent perspectives instead of one model's take.

Tools

• ask_model - one question to one model, or model="all" for a full multi-model panel
• code_review - multi-model review of a diff or file set
• code_plan - architecture and implementation planning
• code_debug - debugging with multiple models on the same error
• codex - pair-programming sessions with per-model context across turns
• get_status - which models are available and their current weights

How it works

One adapter per provider, an orchestrator that fans the prompt out and collects responses, and a blind-judge protocol that anonymizes responses before scoring so no model is favored by name. Model weights learn from outcome feedback over time.

Setup

Requires Python 3.11+. Run from a checkout (it is not a pip-installable package).

python3 -m venv .venv
.venv/bin/pip install -r requirements.txt

Provide access for the models you want. Any subset works; a model without access is skipped:

export GEMINI_API_KEY=...      # Gemini
export XAI_API_KEY=...         # Grok
export MISTRAL_API_KEY=...     # Mistral

The ChatGPT lane uses the [OpenAI Codex CLI](https://github.com/openai/codex) in the default (stdio) mode: install the `codex` CLI and log in once. To use the OpenAI API directly instead, set `MULTI_LLM_TRANSPORT=sse` and export `OPENAI_API_KEY`.

Register the server in your MCP config (`.mcp.json`, or `claude mcp add`):

"multi-llm": {
  "command": "/path/to/multi-reasoner/.venv/bin/python",
  "args": ["/path/to/multi-reasoner/server.py"]
}

Cost note

API-key models bill per token under your own keys; they are not covered by chat subscriptions (ChatGPT Plus and similar)