KAIKAKU-AI / epicure-mcp

What is epicure-mcp?

This is a public MCP server that gives AI assistants access to a food ingredient embedding model. Think of it as a bridge between AI coding tools and culinary knowledge -- you can ask questions like "what goes with miso?" or "compare soy sauce and fish sauce on the umami axis" and get answers grounded in a 1,790-ingredient embedding space trained on 4.14 million recipes. The server is stateless and deterministic: every query is a pure function of the request plus bundled data, with no external model calls or user state. It runs in Python using FastMCP and exposes 13 tools covering ingredient similarity, pairing graphs, flavor axis comparisons, and UMAP-based atlas navigation. Data (~13 MB) is committed directly to the repo, so deployment is self-contained.

Why is it gaining traction?

The hook is turning a food embedding model into a tool that AI assistants can actually use. Instead of building your own integration, you point Claude, Cursor, or ChatGPT at the endpoint and let them query ingredient relationships directly. The pairing graph algorithm is particularly interesting -- it builds clusters and "bridges" connecting different flavor families, surfacing non-obvious combinations that a human might miss. The server also handles free-text ingredient matching ("fresh ginger" resolves to canonical names) and dietary filters (vegan/vegetarian) out of the box. Azure deployment with scale-to-zero keeps costs bounded.

Who should use this?

Developers building AI-assisted recipe tools, meal planners, or culinary chatbots will find this most useful. Food tech teams prototyping pairing recommendations can use it as a backend without hosting their own embedding pipeline. Researchers exploring flavor space structure might appreciate the axis comparisons and factor navigation. It's less useful if you need write operations, user-specific recommendations, or a production SLA -- the rate limiting is best-effort and there's no authentication layer.

Verdict

At 12 stars, this is early-stage but the documentation is thorough and the deployment story is production-minded (Azure Container Apps, OIDC-based GitHub Actions, health checks, structured logging). The 0.85% credibility score reflects the project's youth and small community. If you're evaluating MCP servers for culinary applications, this is worth a look -- just budget time to validate the embedding quality against your specific use cases before committing.