io.github.AIDataNordic/food-recipe-mcp

Official2 tools

by AIDataNordic · Python

Semantic search across 50,000+ food recipes with hybrid retrieval and reranking.

Semantic search across 50,000+ food recipes using hybrid retrieval and reranking.

Topics

fastmcpfoodmcprecipessemantic-search

Captured live from the server via tools/list.

ping

Simple connectivity test. Returns a greeting to confirm the server is running.

Parameters (1)

namestring
Arbitrary label included in the response, e.g. 'healthcheck'

search_recipes

Search a database of recipes using hybrid semantic search (dense + sparse) with reranking. The database contains ~50,000 recipes from Food.com covering a wide range of cuisines, meal types, and cooking styles. Recipes include nutritional information, difficulty ratings, and user ratings. Use natural language in the query to describe what you are looking for — cuisine, style, main ingredient, occasion, or mood all work well. Norwegian and English are both supported natively. Examples: 'quick Italian pasta for weeknight dinner' 'Swedish meatballs with gravy' 'healthy high-protein chicken bowl' 'easy chocolate cake for beginners' 'something with salmon and lemon' 'Indian curry chicken' 'traditional Norwegian kjøttkaker' 'hurtig pasta med kylling' 'enkel sjokoladekake' Args: query: What you are looking for — describe the dish, cuisine, main ingredient, cooking style or mood freely. Any language is supported. diet: Optional — filter by dietary requirement: 'vegetarian', 'vegan', 'gluten-free', 'dairy-free', 'low-carb', 'keto', 'paleo' max_minutes: Optional — maximum total time in minutes, e.g. 30 difficulty: Optional — 'easy', 'medium' or 'hard' servings: Optional — not used for filtering (servings vary), but include in query for scaling context, e.g. 'pasta dish for 6 people' limit: Number of results to return after reranking (default 5, max 20) Returns: List of recipes ranked by relevance. Each result includes rerank_score, rrf_score (hybrid fusion), title, total_time, difficulty, diet labels, ingredients, instructions, nutrition, rating, and source URL context.

Parameters (6)

querystringrequired
Natural language description of what you want, e.g. 'quick Italian pasta' or 'enkel sjokoladekake'
dietstring
Optional dietary filter: vegetarian, vegan, gluten-free, dairy-free, low-carb, keto, or paleo
max_minutesinteger
Optional maximum total cooking time in minutes, e.g. 30. Use 0 for no limit
difficultystring
Optional difficulty filter: easy, medium, or hard
servingsinteger
Not used for filtering — include serving size context in query instead, e.g. 'pasta for 6 people'
limitinteger
Number of results to return after reranking (1–20, default 5)

README.md3 KBView on GitHub

Food Recipe MCP

Semantic search over 50,000+ food recipes — built for AI agents and LLMs. Two-stage hybrid retrieval (dense + sparse BM25, fused via RRF) with cross-encoder reranking. Supports natural language queries in Norwegian and English.

Live endpoint: https://recipes.aidatanorge.no/mcp
Transport: streamable-http
Demo: https://recipes.aidatanorge.no/

Connect

Add to your MCP client config:

json

{
  "mcpServers": {
    "food-recipe": {
      "type": "streamable-http",
      "url": "https://recipes.aidatanorge.no/mcp"
    }
  }
}

Or with Claude Code:

bash

claude mcp add --transport http food-recipe https://recipes.aidatanorge.no/mcp

Quick Test

Try the live demo in your browser:
https://recipes.aidatanorge.no/

No installation or configuration needed.

MCP Tools

`search_recipes`

Semantic search over 50,000+ recipes from Food.com with hybrid retrieval and reranking.

python

search_recipes(
    query="quick Italian pasta for weeknight dinner",
    diet="vegetarian",      # vegetarian | vegan | gluten-free | dairy-free | low-carb | keto | paleo
    max_minutes=30,         # maximum total cooking time in minutes
    difficulty="easy",      # easy | medium | hard
    limit=5                 # default 5, max 20
)
# Returns: rerank_score, rrf_score, title, description, total_time, difficulty,
#          diet, main_ingredient, servings, ingredients, instructions, nutrition,
#          rating, rating_count, source, recipe_id

Query examples:

"Swedish meatballs with gravy"
"healthy high-protein chicken bowl"
"easy chocolate cake for beginners"
"traditional Norwegian kjøttkaker"
"hurtig pasta med kylling"

Search pipeline: Dense embedding (intfloat/e5-large-v2, 1024d) + sparse BM25, fused via Reciprocal Rank Fusion (RRF), reranked by mmarco-mMiniLMv2-L12-H384-v1.

`ping`

python

ping(name="world")
# Returns: "Hello world! Recipe MCP server is running."

Data

Source: Food.com (~50,000 recipes)
Coverage: Wide range of cuisines, meal types, and cooking styles
Nutritional data: calories, fat, protein, carbohydrates, sodium, fiber, sugar per serving
Ratings: user rating + rating count per recipe
Languages: English and Norwegian supported natively in queries

Architecture

code

Food.com recipes → Python ingest → Qdrant (recipe_data_v2 collection)
                                         ↓
                              Hybrid search (dense e5-large-v2 + sparse BM25)
                                         ↓
                              RRF fusion + cross-encoder reranking
                                         ↓
                              FastMCP 3.2 → MCP clients / AI agents

Technical Stack

Embeddings: intfloat/e5-large-v2 (1024d dense) + Qdrant/bm25 (sparse)
Reranker: cross-encoder/mmarco-mMiniLMv2-L12-H384-v1
Vector DB: Qdrant (self-hosted)
Server: FastMCP 3.2 over HTTP
Infrastructure: Ubuntu Server 24 LTS, Cloudflare Tunnel

License

MIT

Install

claude_desktop_config.json

{
  "mcpServers": {
    "food-recipe-mcp": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://recipes.aidatanorge.no/mcp"
      ]
    }
  }
}

Desktop config is stdio-only; this bridges via mcp-remote. Native remote: Settings > Connectors.