Marrow Health: your Apple Health data over MCP, from the free iOS app. Reads, and writes on demand.
tech.skeletonarmy/marrow — Model Context Protocol (MCP) Server
This server provides read-only MCP access to Apple Health data that is mirrored from the Marrow iOS app. It is a self-hosted companion that keeps a SQLite mirror of your data on hardware you own, enabling agents to query it via MCP.
🛠️ Key Features
Read-only MCP access to Apple Health data
Mirrors data from the Marrow iOS app into a local SQLite database
Uses a self-hosted setup (no cloud described)
Standard library only; one Python file
🚀 Use Cases
Let AI agents query your own health metrics via MCP
Support quantified-self workflows using Apple Health data
Include Marrow’s additional tracked series alongside Apple Health types
⚡ Developer Benefits
Agents can access data 24/7 through MCP
Local data flow: phone → your machine
Pre-existing dataset scope: 178 health data series (172 Apple Health types plus Marrow meal diary and set-by-set weightlifting log)
⚠️ Limitations
Read-only access only
Access depends on data being mirrored from the Marrow iOS app
Let your AI agents read your own Apple Health data.
Marrow is a free iOS app that tracks 178
health data series - 172 Apple Health types plus its own meal diary and
set-by-set weightlifting log. This repo is
its self-hosted companion server: one Python file, standard library only,
that keeps a SQLite mirror of your data on hardware you own and serves it over
the Model Context Protocol so agents can query it 24/7, phone in your pocket.
Nothing here talks to a cloud. Data flows phone → your machine, full stop.
Run it
sh
pip install "marrow-mcp[qr]"# [qr] adds segno, which draws the pairing QR
marrow-server --port 8800 --data ./data # prints a pairing URL
Without [qr] the pairing page shows the pairing JSON instead of a code.
or from a clone
sh
python3 marrow_server.py # prints a pairing URL
python3 marrow_server.py --port 8800 --data ./data
or with Docker
sh
docker compose up -d
Open the pairing URL on any screen, then in Marrow: Export, Server MCP,
Pair with a server, and scan. From then on the app's background pushes keep
the mirror current with the app closed. The app pairs over plain http only on
your own networks (a private or tailnet IPv4 address, a .local or .ts.net
name, a short LAN name). A public address needs https (--url https://...).
Upgrading to 0.3.0
Stop the old server and start the new one on the same --data directory.
An existing mirror.sqlite gains the new columns on first start, and the
tokens in --data carry over, so the phone stays paired. Sets logged by
Marrow 1.4 carry a set_type (work,
warmup, failure or drop). Warm-ups are still listed but no longer count toward
volume or records, and rows mirrored before the upgrade read as work sets.
Connect an agent
The server speaks MCP over Streamable HTTP (JSON-RPC 2.0, protocol version
2025-06-18), authenticated with a bearer token it generates on first run and
prints alongside the pairing link.
Download marrow.mcpb,
double-click it, and Claude Desktop installs the extension and asks for the
URL and token. No terminal involved. (The bundle is a tiny zero-dependency
bridge in mcpb/. It forwards Claude's stdio to your Marrow URL and nothing
else.)
Claude Desktop (manual), Cursor, or anything using an mcp.json
Thirteen tools, all of them reads. The server also carries write tools (log a
meal, a workout and so on), but it lists them only after a phone registers it
for write-back, and today only Marrow Connect does that. A self-hosted server
paired from the app serves the thirteen reads and nothing an agent can call
writes to your health record.
Four of them answer in the app's own terms, scored against your own trailing
normal with the same math the app's Home page uses (the day itself excluded),
so an agent starts from a read that is already true for you instead of
judging a number against the population:
Tool
Returns
daily_brief
One day: each key metric with its value and a phrase (normal for you, above your normal, well below your normal), food and water against your targets, last night's sleep, training done, mood and symptoms, and flags worth mentioning
baselines
Your own normal per metric: mean and spread over the window, the latest value's z and phrase, the 7-day mean and the 90-day trend per month
what_changed
The recent days against the period before them: what moved and how notable it is for you, food, training, sleep and bedtime shifts, and streams that started or stopped arriving
data_quality
What the mirror holds and how fresh it is: per metric the latest day, coverage, cadence, sources and whether it has gone stale, plus diary, sleep and training coverage and when the phone last synced
The nine raw reads:
Tool
Returns
health_summary
Recent days across activity, heart, sleep and nutrition
list_metrics
Every metric in the mirror, with units and coverage
metric_daily
Daily values for any metric, up to 400 days
metric_samples
Raw records with timestamps and source devices
workouts
Workouts from every source: activity, start, end, duration and source
nutrition_log
The meal diary: every logged food with full nutrient detail, by day
strength_log
Every lifting set with its set type and estimated 1RM, per-exercise progression, and volume per muscle group (warm-ups left out)
sleep_log
One night per wake day: bedtime, wake time, hours asleep, minutes per stage
symptoms_and_mood
Symptoms with their severity, and State of Mind entries with valence, feelings and what they were about
Every read takes days back from the phone's today, or from and to on
the phone's calendar. metric_samples pages with before (the start of the
last record you have). Tools with an output schema also answer with
structuredContent.
Four prompts come with the server (how_am_i_doing, weekly_review,
doctor_prep, log_from_text) and four resources (marrow://today,
marrow://metrics, marrow://targets, marrow://quality), for the clients
that show them.
Metric keys: list_metrics shows Apple's short names (active_energy_burned),
and the tools also accept the app's own keys (active_energy,
heart_rate_variability), so an agent moved over from the phone keeps its
vocabulary.
Then ask your agent things like "how did my sleep change once I started
training in the mornings?" and let it go and look.
Kits, recipes and the Claude Code plugin
kits/: copy-and-paste setup for Claude Code, Claude Desktop,
Cursor, OpenClaw and Hermes Agent.
recipes/: receivers for Marrow's Automations (the app POSTs
your data to any URL on a schedule, with the app closed): Home Assistant,
n8n, Google Sheets through Apps Script, InfluxDB.
Built with Marrow:
complete projects people run on these connections, starting with Body
Ops, a Grafana dashboard for every kind of data Marrow sends. Submit
yours there.
The Claude Code plugin: /plugin marketplace add MarrowHealth/marrow-mcp
then /plugin install marrow-health@marrow. It adds the Marrow Connect
server and four commands (/marrow-health:how-am-i-doing, weekly-review,
doctor-prep, log) plus a skill that knows which tool answers what.
Why not the built-in Claude or ChatGPT connector?
Both Claude (Pro/Max, US) and ChatGPT now read Apple Health natively. Use them
if that is all you need. They stop where this server starts:
Claude / ChatGPT connector
Marrow + this server
Where it works
The iOS app only
Claude Code, Claude Desktop, Cursor, OpenClaw, Hermes, any MCP client, on any machine
When it works
While you are in the chat on the phone
24/7, phone in your pocket, from your own box
What the model sees
Raw HealthKit numbers
Daily values reconciled against Apple's own totals, meal diary, set-by-set lifts
Where the data lives
Their cloud, per conversation
SQLite on hardware you own
Cost
Subscription
Free, MIT
OpenClaw / Hermes / any agent on a home machine. Point the agent's MCP
config at the server URL with the bearer token, exactly as in the mcp.json
block above. Reach it from outside the LAN with tailscale serve --bg --https=8800 http://127.0.0.1:8800 and use the tailnet URL.
The other ways in
This server is one of the doors on Marrow's Export tab, and you do not need it
to give agents access:
Local MCP. Marrow runs an MCP server on the iPhone itself at
http://<phone-ip>:21212/mcp, answering while the app is running. Seven of
the raw reads above, zero extra hardware.
Server MCP. This package: an always-on mirror on your own box.
Automated. Scheduled pushes to any URL (Home Assistant, n8n, your own
endpoint).
Manual. CSV, JSON and GPX files out of the share sheet.
Marrow Connect is the hosted Pro path: Marrow runs the mirror, and
ChatGPT, Claude or Gemini read it with the phone asleep and nothing to run.
It is the only door that can log for you, once you turn on Allow AI write in
the app. https://marrowhealthapp.com/connect
Security
Both servers are built for a LAN or a tailnet. Do not port-forward this to
the open internet. If you need it remotely, front it with
tailscale serve, Caddy, or an equivalent so the transport is HTTPS and the
device is authenticated before the bearer token is ever presented.
Two separate tokens live in --data: .app-token for the phone's pushes and
.mcp-token for agents. Rotating one does not disturb the other.
Ingest is idempotent and UUID-keyed, so a replayed batch cannot double-count.