One Question. A Council of Models. A Single, Trusted Answer.
Quorum is a professional-grade consensus reasoning engine designed to eliminate AI hallucinations and vendor lock-in. Instead of trusting a single LLM, Quorum orchestrates a council of models (GPT-4, Claude, Gemini, Llama, etc.), clusters their responses semantically, measures agreement, and surfaces the most reliable answer.
Single LLMs are often "confidently wrong." Quorum makes disagreement visible. If models conflict, Quorum flags a Disputed Zone, allowing you to see dissenting opinions rather than blindly trusting a single, potentially incorrect output.
Don't wait for slow models. Quorum starts with local/cheap models and speculatively launches higher tiers if consensus isn't reached within seconds. You get the best of both worlds: extreme speed for easy questions and deep reasoning for hard ones.
Consensus isn't one-size-fits-all.
code&logic: Stricter similarity checks (0.4) to ensure every semicolon counts.creative&prose: Higher semantic tolerance (0.15) to embrace diverse perspectives.
Lexical overlap isn't enough. Quorum uses high-dimensional embeddings to cluster responses by meaning, not just words. With Remote Fallbacks, it automatically uses OpenAI Embeddings if local hardware is unavailable.
Quorum learns which models are experts. Premium models (GPT-4/Claude-3.5) are penalized more heavily for confident failures but rewarded more for leading a correct consensus.
Getting started is now a seamless, interactive experience.
# Clone the repository
git clone https://github.com/magnetoid/Quorum.git
cd Quorum
# Run the guided installer
chmod +x install.sh
./install.sh- Ollama Discovery: Automatically finds and pulls local models.
- Provider Picker: Paste your API keys (OpenAI, Anthropic, etc.) and enable them instantly.
- Council Configurator: Auto-sorts your models into
Local,Cheap, andPremiumtiers. - Health Check: Verifies connectivity and model readiness.
Quorum is fully compatible with MCP, allowing you to use it as a tool directly within Claude Desktop, Cursor, or Claude Code.
Drop-in replacement for any OpenAI client. Simply point your BASE_URL to Quorum.
# Control your budget per-request with custom headers
curl http://localhost:8000/v1/chat/completions \
-H "X-Quorum-Budget: 0.15" \
-d '{"model": "quorum", "messages": [{"role": "user", "content": "How do I scale a Redis cluster?"}]}'| Feature | Quorum | Single LLM |
|---|---|---|
| Hallucination Detection | ✅ Active (via Consensus) | ❌ None |
| Vendor Neutrality | ✅ Multi-Provider | ❌ Locked-in |
| Reliability | ✅ Tiered Fallbacks | ❌ Single Point of Failure |
| Expertise | ✅ Reputation-Weighted | ❌ Generic |
| Speed/Cost | ✅ Speculative Short-circuit | ❌ Linear |
We are evolving from a consensus engine into a complete AI Decision Backbone.
- Speculative Tiered Execution
- Adaptive Domain Thresholds
- Hybrid Semantic Clustering
- Guided Interactive Installation
- Deliberation Round (Multi-Agent Debate): Models critique each other's reasoning to reach a deeper truth.
- Redis-Backed Semantic Cache: Instant answers for semantically identical queries.
- Prometheus & OpenTelemetry: Enterprise-grade monitoring for latency, cost, and consensus drift.
- Automated Judge Feedback: Self-supervision using a "Golden Model" to resolve disputes automatically.
- Continuous Improvement Algorithm (CIA): Self-evolving reputation calibration and council optimization.
- RAG Memory Layer: Infuse councils with private data before they vote.
- Decentralized Councils: Support for multi-node councils across different geographic regions.
We are an open-source project and welcome all contributors! Whether it's fixing a bug, adding a new provider adapter, or improving documentation, your help is appreciated.
See CONTRIBUTING.md for details.
Quorum is released under the Apache-2.0 License.
"In a council of many, there is wisdom."
