openfusion

OpenFusion

Open-source multi-model fusion orchestration engine. OpenAI-compatible API that dispatches requests to multiple LLMs in parallel and synthesizes the best answer.

⚠️ Research / experimental code. OpenFusion is a personal research project exploring multi-model fusion patterns. It works in production at the author's environment but comes with no guarantees. Use at your own risk.

📄 Licensed under MIT — see LICENSE.

🌐 github.com/alitrack/openfusion

Like OpenRouter Fusion, but self-hosted on your own infrastructure.

Features

Quick Start

# 1. Get a Brave Search API key (free, 2,000 queries/month)
#    https://api.search.brave.com/

# 2. Configure
cp config.example.yaml config.yaml
# Edit config.yaml: add providers, API keys, presets

# 3. Build
go build -o openfusion ./cmd/openfusion/

# 4. Run
./openfusion -config config.yaml

# 5. Test
curl http://localhost:8080/v1/models
curl http://localhost:8080/v1/dashboard

Configuration

server:
  addr: "127.0.0.1:8080"
  auth_token: "${OPENFUSION_API_KEY}"

providers:
  openrouter:
    base_url: "https://openrouter.ai/api/v1"
    api_key: "${OPENROUTER_API_KEY}"
    plugin: "generic"

presets:
  dir: "presets"
  items:
    budget:
      description: "Cost-effective fusion with 3 diverse models"
      panel:
        - provider: openrouter
          model: deepseek/deepseek-v4-pro
        - provider: openrouter
          model: google/gemini-3-flash
        - provider: openrouter
          model: kimi/kimi-k2.6
      judge:
        provider: openrouter
        model: openai/gpt-5.5-preview

fusion:
  default_timeout: 120
  panel_timeout_per_model: 60

API Reference

POST /v1/chat/completions

Standard OpenAI chat completions format. Set model to a preset name (e.g. openfusion/budget).

With request-level overrides:

{
  "model": "openfusion/budget",
  "messages": [{"role": "user", "content": "Analyze transformer architecture"}],
  "panel": [
    {"provider": "openrouter", "model": "deepseek/deepseek-v4-pro"},
    {"provider": "openrouter", "model": "google/gemini-3-flash"}
  ],
  "judge": {"provider": "openrouter", "model": "openai/gpt-5.5-preview"}
}

With web search injection:

{
  "model": "openfusion/research",
  "messages": [{"role": "user", "content": "Latest AI papers 2026"}],
  "stream": true
}

Preset Management API

Other Endpoints

Architecture

Client → POST /v1/chat/completions {"model":"openfusion/budget"}
          │
          ▼
     API Layer
          │  1. Parse request + auth + validation
          │  2. Apply request-level overrides (panel/judge)
          ▼
     Cache Check (SHA-256 key)
          │  Miss → continue  |  Hit → return cached response
          ▼
     Web Search (if enabled)
          │  Brave Search API → inject context as system message
          ▼
     Panel Dispatch (parallel goroutines)
          │  goroutine 1: DeepSeek V4 Pro
          │  goroutine 2: Gemini 3 Flash
          │  goroutine 3: Kimi K2.6
          │  (graceful: individual timeout/failure doesn't block)
          ▼
     Judge Synthesis
          │  Judge reads all answers
          │  Output: consensus, contradictions, unique insights, blind spots
          │          + synthesized final answer
          ▼
     Return OpenAI-compatible chat.completions JSON

Skills

Skills define routing strategies via .skill.yaml files. They use trigger matching for automatic routing.

# skills/code-review.skill.yaml
name: code-review
description: "Multi-model code review with consensus analysis"
triggers:
  - categories: "code|review"
    min_tokens: 200
mode: fusion
strategy:
  panel:
    - provider: openrouter
      model: anthropic/claude-sonnet-4
      system: "You are a code review expert."
    - provider: openrouter
      model: openai/gpt-5.5-preview
      system: "Focus on security and performance."
  judge:
    provider: openrouter
    model: openai/gpt-5.5-preview
    enabled: true

Set model: "auto" or model: "openfusion/auto" to use skill matching.

MCP Knowledge Integration

OpenFusion can retrieve domain knowledge from external MCP (Model Context Protocol) servers before dispatching to panel models. This enables knowledge-augmented fusion — panel models receive relevant context from knowledge bases, databases, or ontologies before generating their answers.

Configuration

Add mcp_knowledge sources to any .skill.yaml file:

# skills/domain-advisor.skill.yaml
strategy:
  mcp_knowledge:
    sources:
      # Stdio mode: spawn a subprocess
      - server_cmd: "python scripts/mcp-knowledge-server.py"
        tool_name: "search_knowledge"
        max_tokens: 4000

      # HTTP mode: connect to a remote MCP server
      - server_url: "http://localhost:8080/mcp"
        tool_name: "Ask"
        max_tokens: 8000
  panel: ...
  judge: ...

Supported Transports

Example Knowledge Sources

Architecture

User Question
    │
    ▼
OpenFusion (Skill Matching)
    │
    ├── MCP Client ──→ ChatSQL (NL2SQL query database)
    ├── MCP Client ──→ OntoMind (ontology concept lookup)
    └── MCP Client ──→ Knowledge Base (document RAG)
    │
    ▼ (all context injected into system prompt)
Panel Models (parallel)
    │
    ▼
Judge Synthesis
    │
    ▼
Final Answer

See internal/mcp/ for the MCP client implementation.

Tests

go test ./... -v -count=1

Benchmark

Validated against 3 tasks × 7 modes with DS V4 Pro blind judge:

Key finding: Budget fusion (DS V4 Pro + Flash at ~$0.05) beats Opus 4.6 on code tasks (88 vs 85) at 1/12 the cost.

Full report: claw/wiki/articles/openfusion-benchmark-2026.md

License

MIT