LLM Fusion — AI skill for Claude Code

llm-fusion — Fusion Proxy

**English** | [Русский](./README.ru.md) | [Українська](./README.ua.md)

**Three models argue over your prompt, a judge scores their answers, and one final model writes the reply you actually see.**

llm-fusion is a small proxy you run on your own machine, in front of any OpenAI-compatible provider — [Ollama Cloud](https://ollama.com), which the shipped config points at, or OpenRouter, Groq, DeepInfra, Together, DeepSeek, your own endpoint, or several of them at once. Your coding tool connects to it the way it would connect to any single model — same API, same streaming, same tool calls. The difference is hidden behind the model name: ask for `fusion-coder` and you get a panel of four models answering in parallel, a judge comparing what they said, and a synthesizer writing the final answer. On hard questions, that panel gives you several decorrelated viewpoints from different model families instead of one, an impartial judge that adjudicates their disagreements and flags fragile or hallucinated claims, and a synthesizer that reconciles them into a single answer — plus within-provider failover and completeness guards that keep a single stalled or rate-limited model from ending the turn.

Running that panel on every request would be expensive, so there is also `fusion-agents`: a fast router looks at each request and decides whether it deserves the full panel or just one cheap call. Reading a file gets one call. Recovering from a failed test run gets the whole panel.

No database, no build step, no accounts. Node 24, one YAML config file, and an API key for whichever provider you point it at. It speaks both the OpenAI Chat Completions API and the Anthropic Messages API, so OpenCode, Claude Code, Continue, Cline, Aider — or your own agent loop — all work unchanged.

Familiar with **OpenRouter Fusion**? Same idea — *many models think, one answers* — but self-hosted, transparent, and built to survive long agent loops. Full comparison [below