ResMed-Data-MCP

Archives raw data from a ResMed AirSense CPAP's SD card — read over Wi-Fi via a Toshiba FlashAir card — into a local SQLite database on a Mac, and serves it read-only to any MCP client (Claude, Codex, and others) over the LAN via bearer-token-authenticated Streamable HTTP.

This is a personal quantified-self data plane, not a clinical tool: it archives every field the device reports (AHI, leak rate, pressure, usage hours, and everything else in the nightly summary, plus raw signal detail as it becomes available) without computing or interpreting anything. Synthesis — trends, "how did I sleep", correlations — is left to whatever AI client queries it.

How it works

The AirSense writes to its SD card in the EDF format: STR.edf is a rolling ~365-day nightly summary (one data record per night, one signal per column — AHI, leak, pressure, usage minutes, etc.), and DATALOG/<YYYYMMDD>/ holds per-night detail files at higher resolution once the device has flushed them. A generic EDF/EDF+ parser (src/edf/parse.ts) reads both without any ResMed-specific field list — it just reads whatever signals the header declares, so it also works on future firmware and other ResMed models without changes.

Related MCP server: Apple Health AI Bridge MCP Server

Architecture

Both run as separate LaunchAgents (see examples/) so a slow FlashAir sync never blocks queries and a client bug never touches the data. The MCP server listens on 0.0.0.0:8420 by default, so it's reachable identically whether a client is on the same Mac or elsewhere on the LAN — every client just points at http://<host>:8420/mcp and sends Authorization: Bearer <token> (see Security).

Data model

• nightly_summary — one row per calendar night, one column per signal STR.edf reports (columns are discovered from the device automatically, not hardcoded — so AHI, Leak_95, Duration, S_C_Press, etc. all show up as real, queryable REAL columns). This is the ergonomic table for "how did I sleep" questions.

• edf_signals / edf_signal_records — the fully generic archive: every signal, from every file ever seen (STR.edf and every DATALOG detail file), with raw digital samples preserved as-is plus enough calibration metadata to decode them. This is what "as much raw data as possible" means in practice — nightly_summary is a convenience view over a subset of it.

• source_files, sync_runs, devices — provenance and operational history: what file came from where, when, and whether syncing is healthy.

Event files (the EDF+ annotation channel in DATALOG _EVE / _CSL files) are decoded into session_events. Each apnea or hypopnea row includes the pressure and leak at that minute and whether pressure rose over the next two minutes. minute_stats is one row per minute of pressure, leak, flow limitation, tidal volume, respiratory rate, snore, oxygen, and event counts. night_therapy rolls those minutes up against the night's minimum and maximum pressure.

No vector database: this data is fully structured and numeric (nightly metrics, calibrated signal samples), which SQL aggregation handles better than similarity search. If a future signal turns out to carry free text worth searching semantically, that's a reason to reconsider, not a default.

MCP tools

The connection backing the MCP server is opened SQLite-read-only, so query_raw can't write regardless of what SQL text it's given — that's enforced by the database, not just by string validation.

Install

npm install -g resmed-data-mcp

Requires Node.js 22.5+ (uses the built-in node:sqlite module — no native dependencies, no npm rebuild across machines/architectures).

Develop from source

git clone https://github.com/sfls1397/ResMed-Data-MCP.git
cd ResMed-Data-MCP
npm install   # runs `tsc` via the prepare script

Configure

Copy examples/config.json.example to ~/.resmed-data-mcp/config.json and set flashAirBaseUrl to your FlashAir's address (find it with your router's DHCP client list, or the FlashAir's own config page). Everything else has a sane default. Configure the address once; later package upgrades reuse the existing local file:

Env vars override the file: RESMED_DATA_FLASHAIR_URL, RESMED_DATA_POLL_INTERVAL, RESMED_DATA_SERVER_HOST, RESMED_DATA_SERVER_PORT, RESMED_DATA_DB_PATH. RESMED_DATA_HOME relocates the whole ~/.resmed-data-mcp/ directory (test/dev only).

Backfill once, then install the LaunchAgents

resmed-data-mcp backfill     # one-shot: pulls everything currently on the card
resmed-data-mcp http-token   # creates the HTTP bearer token on first use and prints it

Copy examples/com.resmed-data-mcp.indexer.plist and examples/com.resmed-data-mcp.server.plist into ~/Library/LaunchAgents/, replacing REPLACE_ME_WITH_NPM_BIN_DIR and REPLACE_ME (your macOS username) with real values, then:

launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/com.resmed-data-mcp.indexer.plist
launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/com.resmed-data-mcp.server.plist

Point clients at it

See examples/mcp-client.json. Any MCP client that supports an HTTP server entry with headers connects the same way whether it runs on the Mac that hosts the server or another machine on the same LAN:

{
  "mcpServers": {
    "resmed-data": {
      "url": "http://<host>:8420/mcp",
      "headers": { "Authorization": "Bearer <token from resmed-data-mcp http-token>" }
    }
  }
}

Paths

All under ~/.resmed-data-mcp/ (config cannot override this directory):

Security

Every HTTP request, including /health, needs Authorization: Bearer <token>; anything else gets 401. The token is a random secret created on first use and kept in the login Keychain (service resmed-data-mcp-http, account http-auth-token), never in config or logs. resmed-data-mcp http-token prints it. It is still meant for a trusted home LAN: don't expose port 8420 to the internet (no port forwarding, no tunnel). The MCP server's database connection is read-only at the SQLite level, so even a buggy or malicious client can't write through it.

License

MIT

Privacy gates and synthetic fixtures

EDF test buffers are constructed in code with generic patient headers and fixed samples. SQL fixtures use in-memory databases and synthetic therapy values; no patient export, card image, or local SQLite database is tracked. Keep actual device addresses and patient data in local settings and storage.

Install Gitleaks 8.30.1 and enable the staged gate once per clone:

git config core.hooksPath .githooks
gitleaks git --log-opts=--all --config .gitleaks.toml --redact=100 --no-banner .
npm run build && npm run scan:package

Before a pull request is merged, python3 .security/local-gate.py <PR number> runs Gitleaks, the production npm audit, offline Semgrep and the tests on the exact PR head and records the result as the local-gate status. Push and pull-request CI also scans full fetched history and package contents. The prepublishOnly lifecycle rebuilds and scans the exact npm file list. After a history rewrite, discard old clones and re-clone; merging old history can restore removed data. Releases use GitHub Release → npm OIDC.