Delx Living Body MCP Server

delx-living-body

Meta-MCP that turns 15 wellness MCPs into one unified body data layer for AI agents.

Today, answering "should I train hard today?" forces an agent to orchestrate WHOOP recovery, Garmin Body Battery, Oura sleep, Nourish nutrition, and cycle phase across five separate MCP servers. That's brittle for users and confusing for the agent.

delx-living-body is one MCP server that:

1. Auto-detects which of 15 Delx Wellness connectors you already have installed locally — no manual config
2. Composes parallel calls to the right subset
3. Synthesizes a natural-language answer plus a structured reasoning trace and per-source confidence — using rule-based reasoning, no LLM calls

Install it once. Get a unified body data layer. Works with whatever wellness MCPs you already have.

If it helps your agent workflow, star the repo. Stars make the single-entry Delx Wellness path easier for other AI builders to find.

---

• Try it now, no accounts — npx -y delx-living-body demo ("Should I train hard today?")
• Run it in Claude · Cursor · ChatGPT · Hermes · OpenClaw — see agent setup examples
• Local-first — delx-living-body never reads your tokens; children read their own creds (privacy)
• Which connector should I use? — start at the Delx Wellness front door

The three flagship connectors this composes over: google-health-mcp (google-health-mcp-unofficial), garmin-mcp (garmin-mcp-unofficial), and wellness-nourish (wellness-nourish).

---

Install

npx -y delx-living-body

That's the whole install. No OAuth flow, no API keys — delx-living-body has no auth of its own. Each child connector handles its own credentials.

See the full agent demo → "Should I train hard today?" (no accounts needed)

npx -y delx-living-body demo

One command, no clone. The demo boots the real MCP server, fakes three installed connectors (WHOOP + Oura + Garmin, backed by a bundled stub child that carries synthetic body data), and drives it over stdio exactly the way an agent does. No real accounts, API keys, or network. Add --scenario=red to see the low-readiness ("back off today") path; demo --help lists options. Captured output lives at examples/demo-what-should-i-do-today.txt:

2) living_body_ask  question="What should I do today?"
────────────────────────────────────────────────────────────────
Recommendation:
   Today at a glance: recovery 74, sleep 83, body battery 68.

Confidence: high   Sources: whoop, oura, garmin

3) living_body_ask  question="Should I train hard today?"
────────────────────────────────────────────────────────────────
Recommendation:
   Green light for a hard session. Recovery and sleep both support high intensity.

Confidence: high   Sources: whoop, oura, garmin

Reasoning trace (rule-based, no LLM):
   Intent classified as: training_readiness
   - (rec_high) Recovery 74 supports a high-intensity day.
   - (sleep_good) Sleep score 83 is supporting recovery.

One question in → one synthesized answer composed across all three connectors, with a stable reasoning trace and zero LLM calls. This is the Body-vertical entrypoint: install once, ask in plain language, get a unified answer.

Tools (6)

How detection works

For each known connector, delx-living-body checks:

1. ~/.<vendor>-mcp/tokens.json exists
2. ~/.<vendor>-mcp/config.json exists (password-based connectors like Eight Sleep)
3. An export file at the path in the vendor's env var (Apple Health, Samsung Health)
4. ~/.delx-wellness/profile.json lists the device

If any check passes → detected. Otherwise → missing (with install hint). Stateless connectors (Cycle Coach) are always considered available.

Detection results cache for 60s (DELX_LIVING_BODY_DETECT_TTL).

Known connectors (15)

Composition flow

When living_body_ask or living_body_compose_context runs:

1. Detect installed connectors.
2. For each, spawn it as a child MCP via npx -y <package> over StdioClientTransport.
3. Call the child's *_wellness_context (or *_daily_summary) tool in parallel.
4. Normalize results into a delx-wellness-context/v1 shape with merged scores.
5. Run the synthesizer (rule-based, offline) to produce a recommendation + reasoning trace.

Critically: delx-living-body never calls an LLM. Synthesis is deterministic so downstream agents can reason on top of a stable trace.

Synthesizer rules

14 heuristic rules, each with a stable rule_id that appears in the reasoning trace:

• rec_low / rec_mid / rec_high — recovery score bands
• bb_low / bb_high — Garmin Body Battery bands
• sleep_poor / sleep_good — sleep score bands
• strain_high — WHOOP strain ≥ 18
• cycle_luteal / cycle_follicular — cycle phase signals
• load_high / load_low — aggregate training load
• no_data — nothing installed, advisory only
• conflict — sources disagree → low confidence

Privacy & security

• delx-living-body never reads child connector tokens or config files — children read their own credentials independently.
• Upstream secret env vars (*_CLIENT_SECRET, *_ACCESS_TOKEN, *_REFRESH_TOKEN, *_API_KEY, *_PASSWORD) are stripped before spawning children.
• Children are spawned with privacy_mode=structured by default. raw is only honored when the caller sets explicit_user_intent: true on living_body_ask.
• Child responses are not logged verbatim — only counts and summary fields.
• Per-child call timeout: 30s. A hanging child is marked timeout and skipped.
• Cache lives at ~/.delx-living-body/cache.sqlite (chmod 600), 5 min TTL. Disable with DELX_LIVING_BODY_NO_CACHE=true.
• No phone-home from delx-living-body itself.

See SECURITY.md for the full threat model.

Env vars

CLI

living-body-mcp-server                # MCP stdio server (default)
living-body-mcp-server --http         # Local HTTP transport
living-body-mcp-server doctor         # Detect installed connectors
living-body-mcp-server doctor --json  # JSON output
living-body-mcp-server setup          # Print profile path + install hints
living-body-mcp-server demo           # Zero-secret end-to-end demo (--scenario=red, --help)
living-body-mcp-server version

Use with Claude Desktop

{
  "mcpServers": {
    "living-body": {
      "command": "npx",
      "args": ["-y", "delx-living-body"]
    }
  }
}

Use with Cursor

{
  "mcpServers": {
    "living-body": { "command": "npx", "args": ["-y", "delx-living-body"] }
  }
}

Not medical advice

Outputs are operational context for training/recovery/sleep/nutrition agents. Not for medical diagnosis or clinical use.

License

MIT — see LICENSE. Built by David Mosiah.