nutrition-mcp

A 3-tier nutrition lookup MCP server with local SQLite caching. Searches 326K+ foods locally, falls back to the USDA FoodData Central API, and caches results for instant future lookups.

Setup

npm install
npm run build

Database

The local database is seeded on first use via the nutrition_seed tool (triggered automatically by the companion skill). The first seed downloads and imports the OpenNutrition dataset (~326K foods), which takes a few minutes in the background. Tools work immediately — USDA API results are available while seeding runs.

The database is stored at ~/.nutrition-mcp/nutrition.db.

To manually rebuild (e.g. after a dataset update):

npx nutrition-mcp build-db

Rebuilding preserves any previously cached USDA and web results.

USDA API Key (recommended)

A free USDA FoodData Central API key is strongly recommended. Without it, searches are limited to the local OpenNutrition dataset and many branded/packaged foods won't be found.

export USDA_API_KEY=your_key_here

With direnv, add it to your project's .env file so it's automatically available.

MCP Server

With Claude Code

Add to your project's .mcp.json:

{
  "mcpServers": {
    "nutrition-mcp": {
      "command": "npx",
      "args": ["-y", "nutrition-mcp"]
    }
  }
}

Or for local development:

{
  "mcpServers": {
    "nutrition-mcp": {
      "command": "node",
      "args": ["--import", "tsx", "src/mcp.ts"],
      "env": {
        "USDA_API_KEY": "your_key_here"
      }
    }
  }
}

Tools

nutrition_search

Search foods by name. Returns matching foods with macros.

Returns an array of results, each shaped like:

{
  "id": "on_fd_vUZSKMYWbRkt", "name": "G2G Coconut Almond Protein Bar", "brand": "G2G",
  "source_tier": "local", "serving_size": "1 bar",
  "calories": 300, "protein": 18.0, "fat": 12.0, "carbs": 22.0,   // scaled to the serving
  "basis": "per_serving", "basis_weight_g": 70,                   // never inferred — always check this
  "weight_source": "column",                                     // "column" | "parsed_grams" | "parsed_mass" | "parsed_volume"
  "per_100g": { "calories": 429, "protein": 25.7, "fat": 17.1, "carbs": 31.4 },
  "atwater_delta_pct": 0.2, "is_correction": false,
  "verified_fields": null, "superseded_by": null,
  "data_quality": null, "macro_mass_g": null
}

If serving_weight_g isn't stored for a food, the server tries to derive it by parsing serving_size (e.g. "42GRM" -> 42g, "8 oz" -> 226.8g, "1 cup" -> 240g assuming water-like density) before falling back to per_100g. weight_source tells you which: "column" (stated) and "parsed_grams"/"parsed_mass" (deterministic) are as trustworthy as a stated weight; "parsed_volume" assumes 1.0 g/mL density, so treat it more cautiously — and note that foods matching an oil/honey/syrup-class name never get a parsed_volume weight at all (a 1.0 g/mL guess is wrong enough there that they fall back to per_100g instead) — see docs/CALLER-GUIDE.md. Only when no weight can be resolved at all (e.g. serving_size: "1 bottle", or an oil/honey-class food with only a volumetric serving size) does basis fall back to "per_100g", with the headline values being the canonical per-100g numbers — never silently mislabeled as a serving. per_100g is always present regardless of basis. See docs/CALLER-GUIDE.md for how to use basis, weight_source, atwater_delta_pct, is_correction, and superseded_by.

Every result is also checked for mass conservation: if protein + carbs + fat (per 100g) exceeds 100g, the row is physically impossible and data_quality is "impossible_macros" — in that case calories/protein/fat/carbs/fiber/sugar/sodium are all returned null (never a corrupt or scaled-up number) and macro_mass_g reports the impossible sum; per_100g still carries the raw stored values for reference. See docs/CALLER-GUIDE.md.

nutrition_lookup

Look up a specific food by ID. Returns the complete food record, shaped like the nutrition_search result above (see there for the basis/per_100g/trust-signal fields) plus full source metadata (type, ean_13, source_id, source_query, alternate_names_text, labels, ingredients, data_source, cached_at, updated_at). If superseded_by is non-null, a correction exists for this exact id — look that id up instead.

nutrition_barcode

Look up a food by barcode. Accepts 12-digit UPC-A or 13-digit EAN-13. Searches locally first, then USDA. Same response shape as nutrition_lookup. If a correction exists for this barcode, returns the correction, not the original.

nutrition_cache_add

Add or update a food item in the local cache. Useful for saving nutrition data found on the web.

nutrition_cache_list

List cached food entries (USDA and web-sourced). Does not include the local OpenNutrition dataset.

Returns entries in the same shape as nutrition_search, ordered by most recently updated — except a correction (is_correction: true) always sorts ahead of the row it corrects.

nutrition_cache_delete

Delete a cached food entry by ID. Refuses to delete local dataset entries (on_ prefix) — use build-db to manage those.

nutrition_override

Correct nutrition data for an existing food — e.g. from a physical label. Creates a corrected web-tier copy that inherits all fields from the original, with your corrections applied. Repeated corrections of the same food update the existing entry (not a new one each time). A correction takes precedence over the original on barcode lookup, and is signaled via is_correction/ superseded_by on search/lookup/cache-list.

Defaults to basis: "per_serving" — paste the numbers straight off a physical label, no need to divide by hand. serving_weight_g is required in the same call whenever you supply any macro field under basis: "per_serving" — the stored weight is never used automatically, since it may be exactly what's wrong.

nutrition_seed

Seed the local database with 326K+ foods in the background. Returns immediately — call again to check progress. Idempotent: no-op if already seeded, returns progress if in progress. No parameters.

nutrition_cache_stats

Returns cache statistics: total foods, count by source tier, last cached timestamp, and seed status (phase, progress percentage). No parameters.

CLI

# Search for foods
nutrition-mcp search "chicken breast"
nutrition-mcp search "protein bar" --limit 20

# Rebuild the database
nutrition-mcp build-db

# Start MCP server (default, used by Claude Code)
nutrition-mcp

Companion skill

The /nutrition skill (in skills/nutrition/SKILL.md) lets Claude Code users type /nutrition chicken breast to get a formatted nutrition table. Install by copying the skill to your project or personal skills directory:

# Project-scoped
cp -r skills/nutrition .claude/skills/

# Personal (available in all projects)
cp -r skills/nutrition ~/.claude/skills/

How it works

For a detailed technical overview, see docs/architecture.md. For guidance on interpreting responses as a calling agent (trusting basis, weight_source, atwater_delta_pct, is_correction; when to verify vs. ask the user; how to correct from a physical label), see docs/CALLER-GUIDE.md.

3-tier search

1. Tier 1 (local) — FTS5 full-text search against the local SQLite database (326K+ foods from OpenNutrition)
2. Tier 2 (USDA) — If local results are insufficient, queries the USDA FoodData Central API. Results are cached locally.
3. Tier 3 (web) — Coming in a future phase. For now, use nutrition_cache_add to manually save web-sourced data.

Data

Nutrition values are stored per 100g internally. What a tool call returns is scaled to the food's serving whenever a weight is known or can be derived from serving_size — check the basis field on every result rather than assuming (see the nutrition_search/nutrition_lookup sections above, and docs/CALLER-GUIDE.md for a full walkthrough for callers). Core macros: calories (kcal), protein (g), fat (g), carbs (g), fiber (g), sugar (g), sodium (mg).

Foods are identified by prefixed IDs: on_ (OpenNutrition), usda_ (USDA), web_ (manually cached or corrected via nutrition_override).

Development

npm run dev          # Start MCP server with tsx
npm test             # Run tests (node:test + tsx)
npm run build        # Compile TypeScript
npm run seed         # Seed/rebuild database

Acknowledgments

Inspired by mcp-opennutrition by deadletterq. This project uses the OpenNutrition dataset which combines data from USDA, CNF, FRIDA, and AUSNUT sources, licensed under ODbL 1.0.

License

MIT