What is the safe agent workflow for logging a meal?

Check nourish_connection_status, inspect synthetic shapes with nourish_demo, estimate the meal, show confidence and unresolved items, ask the user to confirm, then call nourish_log_intake with explicit_user_intent true. Read summaries with privacy_mode summary unless item-level detail is necessary.

Where do my food logs live?

Locally, under ~/.wellness-nourish (or NOURISH_LOCAL_DIR if set). There is no hosted account, no cloud sync, no shared vault. JSONL and CSV exports are local files.

Which food databases does Nourish use?

USDA FoodData Central for generic foods (provider attribution preserved) and Open Food Facts for packaged barcode lookups (ODbL attribution + share-alike obligations preserved). Both are public datasets.

Can I scan a barcode photo or describe a meal?

Yes. nourish_decode_barcode_image and nourish_lookup_barcode_image accept image paths, base64 or data URIs. nourish_estimate_meal_photo accepts agent-vision detected_items and returns conservative estimates with confidence — agents must confirm portions before logging.

How do I use Nourish through a Delx Wellness profile pack?

Run npx -y delx-wellness-hermes setup for Hermes or npx -y delx-wellness-openclaw setup for OpenClaw. The profile pack installs the nourish MCP preset, a delx-wellness-nutrition skill and preview-confirm-log behavior so chat messages do not silently mutate the food log.

Is Nourish medical or clinical?

No. Nutrition estimates are approximate and the agent surfaces confidence and source attribution. Nourish is not a medical device and does not provide diagnosis, treatment or clinical guidance.

What is Delx Wellness?

Delx Wellness is an open-source, local-first MCP connector library and optional hosted beta. It gives AI agents structured, user-authorized context from wearables, health platforms, exports and nutrition tools.

Model Context Protocol is an open standard for connecting AI assistants to external tools and data. Any MCP-compatible client — Claude Desktop, Cursor, ChatGPT Desktop, Hermes, OpenClaw, custom agents — can use a connector that speaks MCP.

Should I use a one-command profile pack or wire individual connectors?

Use Delx Wellness for Hermes or Delx Wellness for OpenClaw if you want a daily wellness agent with onboarding included and you want all 10 wearable/API/export connectors plus Nourish ready in one command. Wire individual connectors if you only need one or two providers, prefer Claude Desktop / Cursor as your client, or want full control over the install. Both paths use the same underlying MCP packages.

Why choose local-first?

Health data is sensitive, and local connectors minimize the secrets stored by a third party. Every open-source connector runs on your machine and stores tokens with chmod 600. The separate hosted beta is opt-in for people who prefer account-based access, remote MCP and setup help.

Can I install all 10 at once?

Yes — and they coexist cleanly. The fastest path is npx -y delx-wellness-hermes setup or npx -y delx-wellness-openclaw setup, depending on your runtime. For Claude Desktop or Cursor, paste the multi-server JSON snippet from the homepage Quick start. When two providers cover the same signal, each tool returns provider-tagged data so your agent can reconcile them sensibly.

No — every connector is unofficial and not affiliated with the device makers. They use public APIs where available. Garmin uses your own Garmin Connect login locally, Apple Health reads a manual export.zip instead of live HealthKit, Samsung Health reads a local CSV/ZIP personal-data export instead of live Health Connect, and Google Health uses Google Health API v4 through your own Google Cloud OAuth client. Google describes that platform as actively evolving, so check its current migration and scope guidance before production changes.

Is this medical advice?

No. These connectors expose user-authorized data for personal AI workflows. They are not medical devices and do not provide diagnosis, treatment or emergency monitoring. Always consult a qualified professional for medical concerns.

Can I contribute a new provider?

Yes. The connector quality standard documents the bar for registry_ready and agent_ready connectors. Open an issue or PR on the delx-wellness repo with the provider you want to add and the data you'd expose.

Back to Delx Wellness

Nutrition module

Nutrition MCP for AI Agents

Name: Delx Wellness for Hermes
Brand: Delx Wellness

Nourish is the Delx Wellness nutrition layer. It is a local-first MCP server for calories, macros, TACO/BR food search, barcode image decoding, meal photo previews, carbon context and nutrition trends, built to sit beside wearable connectors instead of pretending to be another device provider.

Direct answer

Nourish MCP 0.8.2 gives AI agents local nutrition context through a preview-confirm-log contract. Start with verified synthetic data, preserve confidence and provenance, and never write a meal without explicit user intent. It is nutrition context, not medical advice.

v0.8.2USDATACOOpen Food FactsPrivacy modesMeal photos

Install from npm Repository

claude_desktop_config.json

{
  "mcpServers": {
    "nourish": {
      "command": "npx",
      "args": ["-y", "wellness-nourish@0.8.2"],
      "env": {
        "FDC_API_KEY": "${FDC_API_KEY}",
        "NOURISH_OFF_ENABLED": "1"
      }
    }
  }
}

First meal in two minutes

Prove the contract before touching personal data.

Install the current package, inspect readiness, then ask for synthetic response shapes with nourish_demo. For a real meal, estimate first and show the user calories, macros, confidence, warnings and unresolved items. Log only after confirmation, then read a summary with the least detailed privacy mode that serves the request.

nourish_connection_status checks providers and the local store.
nourish_demo returns verified synthetic shapes — no personal data required.
nourish_log_intake requires explicit_user_intent: true.
privacy_mode: summary minimizes item-level detail in agent context.

first-meal.mcp

npx -y wellness-nourish@0.8.2 setup --client claude

nourish_connection_status({})
nourish_demo({})
nourish_estimate_meal({ text: "2 eggs and a banana", locale: "en" })

// Only after the user confirms the estimate:
nourish_log_intake({
  text: "2 eggs and a banana",
  explicit_user_intent: true
})

nourish_daily_summary({
  privacy_mode: "summary",
  response_format: "markdown"
})

Food data

Search USDA FoodData Central and keep provider attribution in the output.

Barcode lookup

Decode barcode images locally, then use Open Food Facts for packaged foods with license metadata preserved.

Goals

Preview text or photo-assisted meals, then compare calories, macros and hydration against local goals.

Local-first

Food logs stay local and mutating tools require explicit user intent.

Agent workflow

Preview. Confirm. Log. Summarize.

Nourish can write to your local log, so the agent contract is stricter than wearable connectors. Agents check status, estimate conservatively, surface confidence — and only mutate after explicit user intent.

Inspect nourish_agent_manifest or nourish_capabilities before building a client against the tool set.

Use nourish_demo for verified synthetic response shapes before reading or writing personal data.

Preview a meal estimate before writing to the local log.

Install Hermes personal mode for Telegram with setup --client hermes.

Search TACO/BR local food data for Brazilian staples before falling back to USDA.

Search USDA FoodData Central before estimating generic foods.

Lookup packaged foods through Open Food Facts with ODbL attribution.

Compare days, inspect trends, undo the last entry, or bulk-log a Telegram food recap.

Summarize meal carbon footprint when food names have kg CO2e matches.

Decode barcode photos from image paths, base64 images or data URIs before Open Food Facts lookup.

Use agent vision observations for meal photo previews, then confirm portions before logging.

Use daily coaching and remembered meal shortcuts only with clear user context and explicit save intent.

Read summaries with privacy_mode=summary unless the user explicitly needs item-level detail.

Summarize calories, macros, hydration and goals in agent-friendly markdown.

Export local intake data as JSONL or CSV.

Profile packs

Personal nutrition tracking from chat.

Hermes and OpenClaw setup write a dedicated MCP server block, pin the package path and install a Nourish skill so chat flows default to preview, confirmation and local logging, including barcode photos and meal photo estimates.

Use one local store per personal Telegram profile.
Agents prefer direct MCP tools instead of shell workarounds.
Meal messages are previewed before write unless save intent is explicit.
Barcode photos and meal photos use MCP preview tools before any local log write.

profile-packs.sh

npx -y wellness-nourish setup --client hermes --profile personal --local-dir /root/.hermes/nourish/personal
npx -y wellness-nourish doctor --client hermes --json
hermes mcp test nourish

npx -y delx-wellness-openclaw setup
npx -y delx-wellness-openclaw doctor --profile delx-wellness --run-openclaw
openclaw --profile delx-wellness mcp show nourish --json

MCP tools

nourish_agent_manifest
nourish_capabilities
nourish_quickstart
nourish_demo
nourish_profile_get
nourish_profile_update
nourish_onboarding
nourish_connection_status
nourish_search_food
nourish_lookup_barcode
nourish_decode_barcode_image
nourish_lookup_barcode_image
nourish_estimate_meal
nourish_estimate_meal_photo
nourish_analyze_food_image
nourish_log_intake
nourish_daily_coach
nourish_suggest_next_meal
nourish_remember_meal
nourish_list_memory
nourish_bulk_log_intake
nourish_undo_last
nourish_compare_days
nourish_carbon_summary
nourish_daily_summary
nourish_weekly_summary
nourish_log_water
nourish_hydration_summary
nourish_set_goals
nourish_goal_progress
nourish_export_data

Resources

nourish://agent-manifest
nourish://capabilities
nourish://privacy-audit
nourish://usage-guide

Privacy

Food logs, hydration and goals stay under ~/.wellness-nourish unless NOURISH_LOCAL_DIR is set.
Profile and goals persist locally to ~/.delx-wellness/profile.json (mode 0600); shared with all 14 Delx Wellness connectors and never uploaded.
No hosted account is required.
Agents must preserve source attribution, confidence and safety warnings.
Mutating MCP tools require explicit user intent.

Known limits

Nutrition estimates are approximate.
Meal photo estimates depend on the agent vision layer and always require user confirmation before logging.
No hosted sync or clinical guidance.
Open Food Facts data carries attribution and share-alike obligations.
TACO values ship as an attributed curated subset while redistribution-license confirmation remains a 1.0 gate.