mycelium-hq/ai-brain-starter/ingest-health

ingest-health, Apple Health to DuckDB connector

Loads Apple Health data into ~/.claude/health-mcp/health.duckdb so the rest of the substrate (daily-journal, coaching, advisory-panel, patterns, insights) can read biometrics through health-mcp tools.

This is one of the substrate's ingestion connectors (slack, gmail, linear, github, notion, whatsapp, health). Each owns a single source.

When to use

• User says /ingest-health (with or without --mode xml|csv|tcp and a path)
• User asks to import, sync, capture, ingest, or pull Apple Health data
• User wants to set up the health connector for the first time
• After a fresh Apple Health export from iPhone (mode A) or after refreshing the Simple Health Export folder (mode B)

Do NOT use for:

• Querying existing health data (use health-mcp's health_query, health_metric_series, health_recovery_score, etc.)
• Writing to the vault (this skill is read-only with respect to the vault)
• Non-Apple-Health sources (Garmin / Whoop / Oura get different connectors; for multi-wearable see the open-wearables project)

Privacy notes

Apple Health data is among the most sensitive content the user owns. Read this before running:

1. Storage stays local. The DuckDB file lives at ~/.claude/health-mcp/health.duckdb. Treat it as a medical record.
2. Export files are sensitive too. Apple's export.zip contains every metric ever recorded on the device. Don't commit it to git, don't sync it through cloud storage you don't trust, and delete the staging copy after import if you don't need to re-import.
3. Live mode is local Wi-Fi only. The Health Auto Export TCP server is unauthenticated. Run it only on a trusted network.
4. The MCP never writes the vault. This skill imports into DuckDB. Vault-aware tools READ frontmatter; they never write.

If any of these is unclear, do not run the skill. Ask first.

How it works

Three modes. Pick by argument or ask the user.

Mode A: XML export.zip (default; free, manual, universal)

1. User has exported from iOS Health app: Profile → Export All Health Data → export.zip
2. They've moved the zip to disk (AirDrop / Files / share)
3. Skill calls health_import_xml(zip_path)
4. The MCP streams-parses the XML, inserts into DuckDB, returns row counts
5. Idempotent: re-running on the same SHA returns skipped: true unless force=True

Mode B: Simple Health Export CSV folder (free, manual)

1. User has the folder of HKQuantityTypeIdentifier*.csv files from the Simple Health Export iOS app
2. Skill calls health_import_csv(folder_path)
3. Same idempotency as XML

Mode C: Health Auto Export TCP-live (paid iOS app, real-time)

1. User has Health Auto Export installed on iPhone with TCP server enabled
2. Skill calls health_live_query(metric, host, port, start, end) for the metric the user asks about
3. v1 returns the raw response. v0.2 will normalize into the same DuckDB schema as A and B.

Voice rules

• No em dashes in any user-facing prose
• No exclamation marks
• Direct, no fluff
• Voice rules apply to the summary the skill returns, NOT to docstrings inside health-mcp source

Invocation

The skill is a thin orchestrator. Actual import runs in Python at ${SKILL_ROOT}/ingest.py. The skill assembles the health-mcp tool calls and hands the work to the script.

When invoked:

1. Parse arguments. Accept any of:

• <path> (positional): infer mode from path (.zip or .xml -> A, directory -> B)
• --mode xml|csv|tcp
• --force (re-import even if SHA matches)
• --host HOST --port PORT --metric METRIC --start DATE --end DATE (TCP mode)

1. Resolve mode + paths. If ambiguous, ask the user.

1. Call the appropriate health-mcp tool:

• Mode A: health_import_xml(zip_or_xml_path, force=...)
• Mode B: health_import_csv(folder_path, force=...)
• Mode C: health_live_query(metric, host=..., port=..., start=..., end=...)

1. Surface the response: row counts, skipped/imported flag, top metric types.

1. After successful import (A or B), suggest a sanity check call:

• health_status() for stats
• health_recovery_score("<recent date>") for end-to-end smoke

Output

The skill returns a brief summary. No file writes. Example:

Imported 482,193 records / 18 workouts / 1,247 sleep segments from
/Users/<you>/Downloads/export.zip in 6.4s.
 
Top types:
  HKQuantityTypeIdentifierStepCount: 358,201 records (2018-09-01 to 2026-05-09)
  HKQuantityTypeIdentifierHeartRate: 92,047 records (2019-12-15 to 2026-05-09)
  HKQuantityTypeIdentifierHeartRateVariabilitySDNN: 4,182 records
  ...
 
Try: health_recovery_score("2026-05-09") for an end-to-end smoke check.

Composition

After ingestion, these skills can use the data:

• daily-journal: calls health_journal_context(date) to prompt with last-night sleep + HRV
• patterns: calls health_floor_correlation(metric, days, vault_root) to surface biometric-vs-Floor patterns
• coaching: calls health_coaching_context(start, end, vault_root) to surface recovery markers
• advisory-panel: calls health_panel_context(date, vault_root) to inform deferral suggestions
• insights: calls health_weekly_rollup(week_start) for the weekly review

The wrapper skill health-context auto-fires these when the relevant trigger keywords appear.