OpenFusion

Introduction

What is OpenFusion? ​

OpenFusion asks several AI models the same question, then has one more model read all of their answers and write a single, better one. It runs on your machine, inside the coding agent you already use — Claude Code, Cursor, Cline, Zed, Codex, Gemini CLI, and many more.

It appears to your agent as an ordinary MCP server with one useful tool: fusion. When your agent calls it, OpenFusion sends the prompt to 2–5 models of your choice in parallel, has a judge model compare the answers, and returns one consolidated response.

Where the idea comes from ​

OpenRouter studied this pattern — they call it Fusion — and found two things:

  • A panel of cheap models, fused together, outperformed solo frontier models on deep-research tasks, landing within about 1% of the best at a fraction of the cost.
  • Most of that lift doesn't come from having many models — it comes from the synthesis step: splitting "analyze the candidates" from "write the final answer".

OpenFusion brings that exact architecture to your own machine with your own API keys. No subscription, no middleman — the calls go from your computer straight to the providers you configure.

What you get ​

  • The fusion tool — your agent calls it whenever a question deserves more than one opinion. It returns one answer, with progress updates along the way.
  • The open_dashboard tool — opens a local web dashboard at http://localhost:9077.
  • A local dashboard — configure models and keys, run fusions by hand in the Playground, and inspect every past fusion: which model said what, cost how much, took how long.
  • An agent skill — a small guide installed next to your agent that teaches it when fusion is worth a call, so it doesn't over-use it.

When fusion earns its keep ​

  • Architecture and design decisions, where a second and third opinion change what you build.
  • Deep research, where answers need synthesizing rather than retrieving.
  • High-stakes calls, where independent models agreeing is real confidence.
  • Budget quality — a panel of cheap models, fused, can beat one expensive call.

And when to skip it: routine edits, simple lookups, quick one-off questions. Fusion is 2–3× slower and costlier than a single call, so it pays for itself only on questions that deserve the panel.

What you need ​

  • Node.js 22 or newer
  • Any MCP-capable agent or client — see Connect your agent
  • An API key for at least two providers — or none at all, if you use the built-in keyless providers (rapid-mlx for local Apple-Silicon inference, ollama-cloud)

Next up