OpenFusion

Introduction

How it works ​

OpenFusion is a fusion engine, not an agent. It doesn't browse the web, click around, or call tools. Your agent does the legwork — reads the code, runs the searches, gathers the facts — then hands OpenFusion a prepared question. OpenFusion does one thing with it: turns many single-shot answers into one considered answer.

The three stages ​

  1. Fan out. Your prompt goes to 2–5 candidate models in parallel, each answering once. Workers get no tools and no second tries — they're opinions, not oracles. Each candidate runs under its own timeout, so a stalled model can't hold up the panel.

  2. Analyze. The judge model — a model you pick — reads all of the surviving answers and produces a structured comparison: where the candidates agree (consensus), where they disagree (contradictions), what only some of them covered, and what all of them missed (blind spots).

  3. Synthesize. The judge writes one consolidated answer using only the candidates' answers and its own analysis — no new information, no second round of calls. This two-step split (analyze, then write) is where most of the quality lift comes from.

The rules the engine keeps ​

  • At least two candidates must survive. If timeouts or errors leave fewer than two answers, fusion stops with an error rather than presenting a shaky majority as consensus.
  • Both judge steps use the same model. You configure one judge; it does the analysis and the writing.
  • Everything is logged. Each fusion writes one activity record plus one row per sub-call (each candidate + the two judge steps) into a local SQLite database — which model said what, cost how much, took how long.
  • One call in, one answer out. From your agent's point of view, fusion is just a tool that takes a prompt (plus optional context) and returns text.

What it is not ​

  • Not an agent. No tools, no browsing, no memory between calls. Bring the context; it brings the panel.
  • Not a router. It doesn't pick one model's answer — it writes a new answer from all of them.
  • Not a cloud service. The engine, the dashboard, the database, and your keys all live on your machine. API traffic goes only to the providers you configure.

Why the judge matters

OpenRouter's research found that roughly three-quarters of the quality lift comes from the synthesis step itself, not just from model diversity. That's why OpenFusion treats "compare the candidates" and "write the final answer" as two deliberate stages.

Next up