Contengine

Contengine is a production-grade MCP (Model Context Protocol) server that functions as an orchestration gateway between users and a local NVIDIA inference server running Qwen3.6-35B-MoE via vLLM. The system handles JWT authentication, API key management, rate limiting, tool routing, and multi-user isolation across two hosts (a VPS and a local LAN DGX Spark connected via Tailscale). The architecture is built around a custom FastAPI gateway that intercepts `tool_use` messages from the vLLM inference engine, resolves tools to isolated MCP tool containers, and loops results back to the model. Each user gets dynamically provisioned Docker Compose stacks for complete isolation. The system was designed for asymmetric resilience — external users retain access to their data layer even during WAN drops, since the inference engine remains accessible locally. The project includes 20+ MCP tools written with the MCP SDK, full Observability stack with Prometheus/Grafana dashboards, a Qdrant vector store with sentence-transformers embeddings, and a graphrag corpus management system for multi-user context injection.