Agent prompt · 01
“I have beef. Build a vegetable-led pairing map and explain the bridge ingredients.”
find_pairings
Onion, black pepper, oregano and tomato lead; parsley bridges seven of eight clusters.
Epicure for AI agents
Give your agent a palate.
Connect any MCP client to Epicure's read-only ingredient intelligence through one public endpoint.
Public MCP
Checking live service
Remote endpoint
https://epicure-mcp.kaikaku.ai/mcp- Transport
- Streamable HTTP
- Access
- Public
- Tools
- Checking
Connection object
{
"name": "Epicure",
"url": "https://epicure-mcp.kaikaku.ai/mcp"
}- Read-only tools
- 13
- Ingredients
- 1,790
- Model dimensions
- 300
- Recipe corpus
- 4.14M
A real MCP response
Test a pairing before you connect.
This form opens a live MCP session and calls the same pairing tool available to your agent.
Run the live tool to see how the model positions this relationship inside its corpus.
A measured taste signal appears here.
The complete tool surface
From an ingredient question to a structured answer.
Every operation is deterministic, idempotent, and read-only. No recipe or ingredient data is changed.
Agent prompt · 02
“Move tomato gently toward basil, without losing tomato's identity.”
list_targets → morph
At 30°, the nearest landing points are basil, olive oil, oregano and parsley.
Agent prompt · 03
“Which cuisine directions does tamarind express most strongly?”
cultural_profile
South Asian scores 0.3096 and Southeast Asian 0.2591 in the live model.
Explore
Find useful relationships and place ingredients inside the atlas.
- find_pairings
- Build a cluster and bridge graph around one or more ingredients.
- Example: beef → onion, black pepper, oregano; parsley bridges 7 of 8 clusters
- neighbors
- Return the nearest ingredients in the 300-dimensional model.
- Example: saffron → parsley, tomato, bay leaf, green olive
- where_on_atlas
- Locate an ingredient on the UMAP atlas with nearby peers.
- Example: saffron → curry, cauliflower, zucchini within the 2D atlas
- closest_mode
- Identify the named flavour region an ingredient occupies.
- Example: goat cheese → European melting and table cheeses
Measure
Quantify affinity, cuisine direction, and learned flavour structure.
- pairing_score
- Measure cosine affinity between two ingredients with a corpus percentile.
- Example: pineapple + black pepper → 0.1451 · high affinity
- compare_on_axis
- Compare two ingredients on a verified culinary axis.
- Example: mango vs tomato on sweetness → +0.404 toward mango
- cultural_profile
- Measure an ingredient against eight cuisine directions.
- Example: tamarind → South Asian 0.3096; Southeast Asian 0.2591
- flavour_correlations
- Inspect correlations between the model's flavour axes.
- Example: citrus ↔ woody → 0.845 directional correlation
- ingredient_on_factor
- Project an ingredient onto a residual flavour factor.
- Example: dark chocolate on factor 0 → percentile 16, pole B
Transform
Navigate toward a target while preserving a clear culinary starting point.
- morph
- Rotate an ingredient toward another ingredient, mode, or direction.
- Example: tomato 30° toward basil → basil, olive oil, oregano, parsley
- pareto_navigate
- Balance closeness to a seed against a named flavour pole.
- Example: miso → a frontier of close options across its strongest poles
- list_targets
- Discover valid axes, modes, and directions before transforming.
- Example: list cuisine directions → Japanese, Mediterranean, East Asian…
- list_factors
- Inspect the available residual flavour factors and their poles.
- Example: factor 0 → Indonesian pantry identity, with high coherence
How the connection runs
Public at the edge. Private at the origin.
Cloudflare carries each request to Epicure's self-hosted compute without exposing an origin port.
Your AI client
Sends a standard MCP request from its reasoning loop.
Cloudflare edge
Terminates TLS and carries the request through the private tunnel.
Epicure MCP
Runs one of thirteen deterministic, read-only tools on self-hosted compute.
Flavour atlas
Returns a structured result from the 300-dimensional ingredient model.
The website, API, and MCP remain separate services behind the same local cluster and Cloudflare tunnel.
Explore the kitchen