A deep dive into the design of a Java MCP server for urgency scoring, built for a JavaOne 2026 demo. The server exposes a single tool that scores support ticket complaints on a 0–10 urgency scale. It supports two embedding strategies behind the same stable MCP interface: in-process MiniLM embeddings via DJL (384-dimensional, no external dependencies) and hosted OpenAI embeddings (1536-dimensional, production-quality semantics). Each embedding path has its own DeepNetts feed-forward scorer trained specifically against that representation, making the two scorers incompatible with each other. The Helidon framework handles MCP server bootstrapping via annotations (@Mcp.Path, @Mcp.Server, @Mcp.Tool), while the inference stack is loaded lazily. The in-process path is recommended for development, integration testing, and CI; the hosted path is used for semantic validation when credentials and network access are available.
