Epicure MCP

Epicure is a public, anonymous, read-only Model Context Protocol server for computational flavour exploration. It exposes deterministic tools over 1,790 ingredient embeddings learned from a 4.14 million-recipe, multilingual corpus.

• MCP endpoint: https://epicure-mcp.kaikaku.ai/mcp

• Health check: https://epicure-mcp.kaikaku.ai/healthz

• Interactive guide: epicure.kaikaku.ai/agents

• Authentication: none

• Transport: Streamable HTTP

The production origin runs on KAIKAKU's local compute cluster and is published through a Cloudflare Tunnel. It is not hosted on Google Cloud or Azure. Every tool reads the bundled model artefacts; the MCP server makes no external AI or data-provider calls.

Connect from Claude

In Claude, open Settings → Connectors → Add custom connector and enter:

Name: Epicure
URL:  https://epicure-mcp.kaikaku.ai/mcp
Auth: None

Try one of these prompts after enabling the connector:

• “Use Epicure to build a vegan pairing graph around tomato and basil.”

• “Compare miso and soy sauce on the savoury axis.”

• “What ingredients are nearest to saffron in Epicure’s flavour space?”

• “Move rice 30 degrees toward the South Asian cuisine direction.”

• “Show where yuzu sits on Epicure’s ingredient atlas.”

• “Find a flavour trade-off from miso along a coherent factor.”

Tool outputs describe learned statistical relationships. They are not food safety, allergy, medical, or nutritional advice.

Related MCP server: Pepesto MCP Server

Tools

All tools advertise read-only, non-destructive, idempotent, closed-world MCP annotations and validated JSON schemas.

Large catalogues are filtered and paginated. Default responses are deliberately compact; callers can request more detail through each tool's documented inputs.

Data and privacy

Epicure has no accounts, cookies, user profiles, or advertising. Tool arguments and result content are processed in memory and are never written to application logs. Minimal operational telemetry contains a rotating pseudonymous client hash, tool name, response size, latency, success state, and exception class.

Read the complete Privacy Policy, Security Policy, Support Policy, and Terms.

Architecture

MCP client
    │  HTTPS / Streamable HTTP
    ▼
Cloudflare edge (TLS, DDoS controls, WAF/rate limits)
    │  outbound-only Cloudflare Tunnel
    ▼
reef-cluster on local KAIKAKU compute
    └── epicure-mcp container
          └── bundled read-only model artefacts

The origin binds to loopback on the cluster host. No inbound router port is opened; cloudflared initiates the tunnel from the private Docker network. See docs/DEPLOYMENT.md for operations and docs/CLAUDE_SUBMISSION.md for the directory listing.

Local development

Python 3.11 or newer is supported (production uses Python 3.12).

python3.12 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"
python scripts/verify_data.py --data-dir data
ruff check src tests scripts
pytest -q
python -m epicure_mcp.server

The local endpoints are:

Inspect the local server with the official MCP Inspector:

npx -y @modelcontextprotocol/inspector \
  --cli http://127.0.0.1:8080/mcp --transport http --method tools/list

Run the complete remote smoke suite against production:

python scripts/smoke_test_remote.py https://epicure-mcp.kaikaku.ai/mcp

Configuration

Bundled artefacts

The repository includes the model bundle, so production does not depend on an upstream service at runtime. It contains embeddings and ingredient metadata, supervised directions, ICA factors, GMM modes, and 2-D/3-D UMAP projections. Validate shape and presence with scripts/verify_data.py after any refresh.

License

MIT © KAIKAKU.AI Limited.