Food Facts MCP Server

An MCP (Model Context Protocol) server that gives AI assistants real-time access to nutrition data from two sources: USDA FoodData Central (whole foods, branded products, restaurant items) and FatSecret (extensive branded and fast-food coverage). Built for HooHacks 2026.

The problem

AI assistants like Claude know a lot about nutrition in general, but they cannot reliably answer specific nutrition questions without this server. Three concrete failure modes:

1. Hallucinated numbers. Ask Claude "how much protein is in a Chick-fil-A Deluxe Sandwich?" without live data access and it will produce a plausible-sounding number from training data — which may be wrong, outdated, or for a different serving size. There is no way for the model to know it's wrong.

2. No citations. Any nutrition claim an AI makes from memory is uncitable. For dietary tracking, research, or anything that matters, you need a traceable source. Without this server, Claude cannot point you to a specific USDA FDC record or FatSecret entry — it can only say "according to general knowledge."

3. Stale data. Restaurant menus and product formulations change. An AI's training data has a cutoff; it has no way to reflect a menu item that was reformulated last quarter. This server fetches live data every time (and caches it), so the numbers are current.

This MCP server solves all three by connecting the AI directly to authoritative, live databases — USDA FoodData Central (the US government's official nutrition database) and FatSecret (2.3M+ branded and restaurant foods) — and attaching a citation to every response.

Related MCP server: Food Data Central MCP Server

What it does

Ask Claude (or any MCP-compatible AI) questions like:

• "What are the nutrition facts for a Chick-fil-A chicken sandwich?"

• "Compare broccoli and spinach for iron content."

• "What are the top 10 foods highest in vitamin C?"

• "Give me an APA citation for USDA data on raw almonds."

The server fetches live data, caches results locally in SQLite, and returns structured responses with citations.

Tools (13 total)

USDA FoodData Central (8 tools)

FatSecret (2 tools)

Cache Management (3 tools)

Data Sources

USDA Dataset Types

Setup

1. Get API keys

• USDA FDC — free at api.data.gov/signup — 1 000 req/hr

• FatSecret — free at platform.fatsecret.com — 5 000 req/day

2. Install

pip install -e .

3. Configure

cp .env.example .env

Edit .env:

USDA_FDC_API_KEY=your_fdc_key
FATSECRET_CLIENT_ID=your_fatsecret_id
FATSECRET_CLIENT_SECRET=your_fatsecret_secret

# Optional cache settings
CACHE_DB_PATH=          # default: ~/.cache/food_facts_mcp/food_facts.db
CACHE_TTL_DAYS=         # default: no expiry
CACHE_ENABLED=true      # set false to disable caching

Usage

Claude Desktop (stdio)

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "food-facts": {
      "command": "python",
      "args": ["-m", "food_facts_mcp.server"],
      "cwd": "/path/to/HooHacks2026",
      "env": {
        "USDA_FDC_API_KEY": "your_key",
        "FATSECRET_CLIENT_ID": "your_id",
        "FATSECRET_CLIENT_SECRET": "your_secret"
      }
    }
  }
}

Restart Claude Desktop — the server appears under the MCP tools icon.

HTTP server

food-facts-server --transport streamable-http --port 8000

MCP endpoint: http://127.0.0.1:8000/mcp Health check: http://127.0.0.1:8000/

MCP Inspector (for testing without Claude Desktop)

npx @modelcontextprotocol/inspector food-facts-server

Caching

Responses are cached in SQLite after the first API call. Subsequent calls for the same food/query return instantly with no network request.

First call:   search_foods("broccoli")  →  ~1-2s  (USDA API)
Second call:  search_foods("broccoli")  →  <1ms   (SQLite cache)

The cache is source-agnostic — USDA FDC and FatSecret responses are stored in the same DB under separate source keys. Use get_cache_stats to inspect, list_cached_foods to browse, and clear_cache to reset.

Deployment

Public HTTPS with Caddy + DuckDNS

1. Register a free domain at duckdns.org

2. Start the MCP server:

food-facts-server --transport streamable-http --host 127.0.0.1 --port 8000

3. Create a Caddyfile:

yourdomain.duckdns.org {
    reverse_proxy 127.0.0.1:8000
}

4. Run Caddy (handles TLS automatically via Let's Encrypt):

caddy run --config Caddyfile

5. MCP endpoint is now live at https://yourdomain.duckdns.org/mcp

ChatGPT Connector

In ChatGPT → Settings → Connectors, point to https://yourdomain.duckdns.org/mcp.

Extra CORS origins

food-facts-server --transport streamable-http \
  --allow-origin https://myapp.com

Development

pip install -e .
python -m pytest tests/ -v   # 69 tests

Project structure

src/food_facts_mcp/
  server.py            — FastMCP server, tool/resource/prompt registration
  tools.py             — All tool implementations (FDC + FatSecret)
  fdc_client.py        — USDA FoodData Central API client
  fatsecret_client.py  — FatSecret OAuth2 API client
  cache.py             — SQLite cache layer (FoodCache)
  citations.py         — APA + MLA citation builders
  resources.py         — MCP resource handlers + static data
  prompts.py           — Prompt template definitions
  sampling.py          — Sampling request builders
tests/
  test_tools.py        — FDC tool unit tests (mocked)
  test_fatsecret.py    — FatSecret tool unit tests (mocked)
  test_cache.py        — SQLite cache unit tests (in-memory)
  test_resources.py    — Resource + citation tests
  test_http_server.py  — HTTP transport + CORS tests

API References

• USDA FDC API Guide

• FDC OpenAPI Spec

• FatSecret Platform API Docs

• Model Context Protocol