Nutrition MCP

Introduction

A calorie and macronutrient tracker built around an MCP server. A single Bun process exposes a streamable-HTTP MCP API backed by the OpenNutrition dataset (SQLite + FTS search) and serves a Vue 3 web app — search foods, build recipes, compute macros, and log what you eat.

Tools exposed to your LLM client:

• searchIngredient — FTS-backed food search with category/tag filters

• getIngredientMacros — per-100g and per-serving nutrition with measured/unmeasured flags

• computeRecipeMacros — deterministic macro summation for a recipe

• filterFoods — nutrient ranges and dietary presets (keto, high-protein, …)

• convertUnits — unit conversion, food-aware where possible

• getServerRecipes — returns every recipe from the server's recipes/ directory (one JSON file per recipe), validated against the export envelope; invalid files are skipped and reported. Directory is read fresh on each call, so newly added files appear without a restart.

All arithmetic happens server-side, so the model never does math or invents nutrition data. Single-user by design: logs, recipes, and favorites live in your browser's IndexedDB and never leave your machine.

Related MCP server: nutrition-mcp

Usage

Requirements: Bun and Docker (for server deployment only).

1. Clone repository

2. Download latest OpenNutrition dataset from https://www.opennutrition.app/download and extract to local directory (e.g. external/opennutrition-dataset-YYYY/)

3. (First run only) Install dependencies and create OpenNutrition.sqlite database

   bun install && bun run build:sqlite external/opennutrition-dataset-2025.1/opennutrition_foods.tsv

   The TSV is gitignored — obtain the OpenNutrition dataset release first and place it at that path. The build writes opennutrition.sqlite next to the TSV.

   Recommended organization:

   mkdir -p data/recipes
   
   mv external/opennutrition-dataset-YYYY.MM/opennutrition.sqlite data/

4. Dev Server (Vite) Copy or symlink database to server/data/ and recipes folder to server/data/recipes/

5. Docker Deploy Adjust the bind mount paths in docker-compose.yml:

   ...
   volumes:
   - "./data/opennutrition.sqlite:app/server/data/opennutrition.sqlite:ro"
   - "./data/recipes:/app/server/data/recipes:ro"
   ...

6. Docker Start the container: docker compose up -d --build On Linux, if your user is not uid 1000 and you bind-mount the metrics store, match the container's bun user to your uid so mounted files stay owned by you: HOST_UID=$(id -u) HOST_GID=$(id -g) docker compose up -d --build

Web app: http://<host>:3000/nutrition-tracker/

MCP endpoint: http://<host>:3000/mcp — point your chat client (Claude Desktop, OpenWebUI, etc.) at this URL

Metrics & logging

Opt-in server usage/performance metrics (ADR-0031), off by default. Enable with NUTRITION_ENABLE_LOGGING=1 (all variables are documented, commented, in examples/example.env):

What gets recorded — per request, in a separate SQLite store: UTC timestamp, endpoint (http:/mcp, http:static, http:static:404, mcp:tool:<name>), duration, error flag, a per-day anonymous client hash, and a coarse client bucket (browser, claude-code, codex, opencode, other, unknown). Never recorded: IPs, request payloads, tool arguments, full User-Agents, or anything linkable to a person. Unique clients are counted via a salted hash whose salt is discarded daily — it cannot be reversed, linked across days, or recomputed later.

Output, single-line JSON on stdout:

{"type":"usage-hourly","window":"2026-10-03T12:00:00–12:59:30-04:00","timezone":"America/New_York","requests":{"http:/mcp":37,"mcp:tool:searchIngredient":25},"total":159}
{"type":"usage-daily","window":"2026-10-02T00:00:00–23:59:30-04:00","timezone":"America/New_York","endpoints":{"http:/mcp":{"requests":812,"avgMs":23.4,"errors":3},"mcp:tool:searchIngredient":{"requests":700,"avgMs":41.2,"errors":1}},"uniqueClients":4,"uaBuckets":{"browser":132,"claude-code":668},"total":812}

Hourly snapshots emit at H:59:30 local, the daily snapshot at 23:59:30 — so every line belongs to the local day it describes. Idle hours print nothing; snapshots missed by a restart are backfilled from the store on startup.

License

See LICENSE.md

Contributing

See CONTRIBUTING.md