LLM Council
A pattern (popularised by Andrej Karpathy’s karpathy/llm-council repo) where several LLMs answer independently, review each other anonymously, and a “chairman” model synthesises one final answer.
Stages
- Divergence: the query goes to several models in parallel (e.g. GPT, Claude, Gemini, Grok via OpenRouter); no shared context.
- Critique: each model reviews and ranks the others’ anonymised answers.
- Synthesis: the chairman model compiles the final response.
Status notes
- The original repo is a local web app (Python backend, React frontend, OpenRouter). Its author states it is a “vibe coded” weekend project that he will not support (“provided here as is for other people’s inspiration”). ~25.1k stars at access time; licence not stated on the page.
- Many community derivatives exist, including Gemini CLI extensions and grounded/research variants. The “MCP” tag on this note is loose: the original is not an MCP server. Earlier text of this note was a pasted social-media post (source unknown) and has been replaced.
Sources
- https://github.com/karpathy/llm-council (accessed 2026-09-30)