Wellness Nourish

Local-first nutrition MCP — food search, barcode lookup, intake logging, hydration. Works without OAuth.
Local-first MCP server — tokens never leave your machine.

📈 Published on npm and used by AI agents and MCP clients — see the live download badge above for current numbers.
_{If Nourish helps your agent, a ⭐ on this repo makes it easier for other AI builders to find.}

⚡ One-command install — pick your runtime:

• Delx Wellness for Hermes: npx -y delx-wellness-hermes setup
• Delx Wellness for OpenClaw: npx -y delx-wellness-openclaw setup

Both preconfigure this connector and the full Delx Wellness stack into a dedicated profile. Or wire it standalone into Claude Desktop / Cursor / ChatGPT Desktop — see the install section below.

Want runnable agent examples? Use the Delx Wellness hub for prompt packs, MCP client configs and local-first workflow templates.

What's new in 0.7.1: Offline fixture demo works from the published npm package (fixtures shipped + path resolved from package root). 0.7.0: Brazilian TACO 4 meal estimator for pt-BR foods (cafezinho, feijão, concha, churrasco-style meals) plus Smithery install. Offline demo: NOURISH_FIXTURE_MODE=1 npx -y wellness-nourish search banana. Notes: CHANGELOG.md · eval: docs/evals/pt-br-meal-estimator.json (52 cases).

Public proof: Nourish is tracked in the Delx Open Source Growth Snapshot alongside downloads, stars and next-action priorities. If this saves you setup time, star this repo so other agent builders can find the local-first nutrition path faster.

HTTP (v2 stateless)

Default is stdio. Optional Streamable HTTP — no session id, JSON responses, loopback only:

npx -y wellness-nourish --http
# GET  http://127.0.0.1:3000/health
# POST http://127.0.0.1:3000/mcp   (sessionless)

Env: NOURISH_MCP_HOST, NOURISH_MCP_PORT, NOURISH_MCP_TRANSPORT=http.

Local-first nutrition MCP for AI agents — food search, barcode lookup, photo-assisted meal estimation, intake logging, hydration, goals and coach-style workflows. No OAuth, no hosted account.

Front door

• Install one connector — npx -y wellness-nourish setup --client claude
• Run it in Claude · Cursor · ChatGPT · Hermes · OpenClaw — see the client examples.
• Local-first — your tokens and food logs never leave your machine (privacy).
• Which connector should I use? — see the front-door guide.

Quickstart (60 seconds)

npx -y wellness-nourish@0.8.0 doctor
npx -y wellness-nourish@0.8.0 search banana
npx -y wellness-nourish@0.8.0 barcode 0000000000000
npx -y wellness-nourish@0.8.0 log --preview "2 ovos, banana e café preto"

doctor checks readiness, search/barcode hit the food providers, and log --preview estimates a meal locally without writing anything.

How you can help

Zero-secret demo (offline, no API key)

NOURISH_FIXTURE_MODE=1 serves the bundled fixtures/ instead of calling USDA or Open Food Facts, so you can see the exact shape of every response with zero network access or keys:

$ NOURISH_FIXTURE_MODE=1 wellness-nourish search banana
Bananas, raw	usda	89 kcal/100g
BANANA	usda	312 kcal/100g

Try it with your agent

Three copy-paste prompts, all backed by existing tools:

• "Estimate the calories and protein in 2 eggs, a banana and black coffee." → nourish_estimate_meal
• "Look up the barcode 737628064502 and tell me what it is." → nourish_lookup_barcode
• "What should I eat next today, given my goals?" → nourish_daily_coach / nourish_suggest_next_meal

Mutating tools (log intake, water, goals, clear-day) never run without explicit user save intent — they return USER_ACTION_REQUIRED until the agent passes explicit_user_intent: true.

Tools

Nourish exposes food search, barcode lookup (text + image), photo-assisted meal estimation, intake logging, hydration, goals, exports, daily/weekly summaries, personal meal memory, and coach-style workflows over stdio (default) or Streamable HTTP (POST /mcp).

• Full CLI (20+ commands), install, client configs & ChatGPT dashboard → docs/cli.md
• Hermes / Telegram personal setup (10-step flow) → docs/telegram.md
• Data providers & attribution (USDA, Open Food Facts, ZXing) → docs/providers.md
• pt-BR meal-estimator eval set (52 examples) → docs/evals/pt-br-meal-estimator.json
• Reproducible Telegram/Hermes demo transcript → docs/telegram-demo-transcript.json

Food photo decision tree

Agents should route Telegram/Hermes/OpenClaw food photos by the strongest signal they can extract:

1. Barcode is visible and image bytes are available: call nourish_lookup_barcode_image.
2. Barcode is blurry or no product is found: ask for sharper barcode digits, or call nourish_analyze_food_image with barcode_observation plus any OCR/meal clues.
3. Nutrition facts are readable: OCR the label and call nourish_analyze_food_image with product_name and nutrition_label_text.
4. It is a plate or unpackaged food: describe visible foods/portions and call nourish_analyze_food_image with detected_items or image_description.
5. Never log from an image response until the user confirms the product or meal, serving size and save intent.

Image tools accept exactly one of these input forms:

{ "image_path": "/tmp/telegram-food-photo.jpg" }

{ "image_base64": "<base64 image bytes>", "image_mime_type": "image/jpeg" }

{ "image_data_uri": "data:image/jpeg;base64,<base64 image bytes>" }

If barcode decoding fails, the response includes fallback and next_actions so the agent can ask the user for the typed digits, OCR the nutrition label, or route the photo as a meal without silently inventing a food.