Fusion

Ask several AI models the same question at once, then have one model read all their answers and write you the best combined reply. It is a plugin for OpenCode.

Inspired by OpenRouter Fusion, which found that a panel of models, compared and combined, beats any single model on hard questions.

Why use it

One model has blind spots. Three models rarely share the same ones. Fusion sends your question to a few models at the same time, lets a "judge" model compare their answers, and your assistant writes the final reply from that comparison. You ask once and get a stronger answer.

Use it for research, big decisions, or anything where being wrong is costly. Skip it for quick edits.

Install

npm install opencode-fusion

Then add it to your OpenCode config (opencode.json):

{
  "plugin": ["opencode-fusion"]
}

Restart OpenCode. You need API keys for the models you want on the panel. Run /connect in OpenCode for each provider (for example Anthropic, OpenAI, DeepSeek).

How to use it

Just ask your assistant a hard question. It decides on its own when a question is worth a panel and calls fusion. You can also ask directly:

Use fusion to compare Postgres and SQLite for my app

Configure (optional)

It works out of the box. To choose your own panel models, judge model, and other settings, create ~/.config/opencode/fusion-config.json. The full list of options, examples, and how each part works lives in llms.md.

Defaults

The default configuration used out of the box:

{
  "judge": {
    "providerID": "anthropic",
    "modelID": "claude-opus-4-8"
  },
  "panel": [
    { "providerID": "anthropic", "modelID": "claude-opus-4-8", "label": "Claude Opus 4.8" },
    { "providerID": "openai", "modelID": "gpt-5.5", "label": "GPT-5.5" },
    { "providerID": "deepseek", "modelID": "deepseek-chat", "label": "DeepSeek" }
  ],
  "panelTools": ["read", "grep", "glob", "list", "webfetch"],
  "judgeTools": ["webfetch"],
  "panelMaxSteps": 16,
  "judgeMaxSteps": 12,
  "panelTimeoutMs": 600000,
  "judgeTimeoutMs": 600000,
  "maxTokensPerPanel": 4096,
  "judgeMaxTokens": 8192,
  "temperature": 0.7
}

License

MIT