Security
Last reviewed 29 October 2026. For security reviewers, integrators and anyone who wants to check rather than trust.
MetricBridge is local-first and read-only: there is no account and no developer server in the health-data path, so the developer is never in a position to leak it. This page lists what is hardened, how to verify each claim yourself, and how to report a problem. The plain-language version, one row per data flow, is the privacy receipt.
In one screen
- Health data: read on the iPhone, written only to the destination you choose. The app holds no permission to write to Apple Health.
- AI: every answer's numbers are computed on the iPhone; the optional Your Mac mode words answers on your own Mac. No cloud AI SDK or API is in the binary.
- MCP server: open source, zero dependencies, no outbound network calls, 21 tools: 19 read-only, plus two that write only an events file beside the export, never Apple Health. npm releases carry Sigstore provenance; the site mirror is minisign-signed.
- Site: strict CSP, HSTS preloaded, DNSSEC, DMARC
p=reject, CAA, TLS 1.2+. - Reporting: [email protected] or a private GitHub advisory; security.txt per RFC 9116.
Inside the app
- HealthKit access is read-only and category-scoped. The set of types the app asks to write
is empty in code (
Sources/HealthKitProvider.swift), and a unit test fails the build if it ever grows. - Ask answers, the Weekly Health Brief and reports are computed on the device: the statistics are deterministic Swift; the model that phrases them runs on the device too (Apple's on-device model, or an optional downloaded model), unless you turn on "Your Mac", which words answers with Ollama on your own Mac. A grounding gate rejects any sentence whose numbers are not in the computed facts.
- There is no cloud AI provider in the app: no OpenAI, Google or Anthropic SDK or API. The app's only third-party packages are Meta's SDK (advertising measurement, below) and the on-device model runtime.
- Sensitive categories (reproductive health, sexual activity, alcohol) stay off unless you explicitly opt in.
- Exports to your Wi-Fi or a webhook carry a token you set (
X-Health-Token); a public webhook must be HTTPS, plain HTTP is allowed only for local and loopback addresses. The in-app history shows what was sent, when, and whether it succeeded. - Optional export encryption: AES-256-GCM on the iPhone with a key derived from your passphrase, before the file leaves the device. The key lives in the iPhone Keychain and never moves; the MCP server decrypts on your own computer.
- Agent pairing: if you turn it on, your agent reads your export only after you give it a pairing code from the app. Regenerating the code is the revoke: the old code stops working.
- "Your Mac" (Ollama on your own Mac over Wi-Fi) shipped in version 1.20 and is off until you turn it on. It sends one request per answer, to your Mac only, after you have read the
exact request in a consent sheet. iOS allows it only to
.localnames, never raw public addresses. - The app includes Meta's advertising SDK for install, trial-start and purchase attribution. It is
gated by the App Tracking Transparency prompt: until you tap Allow, the SDK is never started, so no Meta code runs
and nothing is queued. It is walled off from HealthKit:
Sources/AdAttribution.swiftis the only file permitted to import it, it does not import HealthKit, and its API accepts only an amount and a currency. - No crash-reporting or analytics SDK. Crash logs reach us only through Apple, and only if you opted in to sharing with developers in iOS Settings.
Supply-chain integrity
npm. health-export-mcp 1.8.0 was published by the repository's GitHub Actions release
workflow with npm publish --provenance. The registry holds a Sigstore-signed
SLSA provenance attestation
(predicate https://slsa.dev/provenance/v1) that links the tarball to the exact commit and workflow run that built it,
and the npm package page shows the provenance badge. Versions 1.6.0 and earlier were published by hand and have registry
signatures but no provenance.
GitHub releases. Each release carries the tarball, the .mcpb bundle, a CycloneDX SBOM
(health-export-mcp-1.8.0.cdx.json), a SHA256SUMS file and SHA256SUMS.minisig, a
minisign signature made offline with key
951BA54856154FBA. Signed checksums start with v1.6.0.
This site's mirror (/mcp/*.mjs, used by the agent setup skill) is
covered by the same key: SHA256SUMS lists every file and
SHA256SUMS.minisig signs the list (trusted comment: "healthexport.dev/mcp artifacts, health-export-mcp 1.8.0").
The public key is published from a second, independent source (the
GitHub repository and SKILL.md),
so a compromise of this origin alone cannot swap both an artifact and its key. Pin it from there, not from here.
CI re-verifies the signature on every change (scripts/verify-minisig.mjs), so a rebuilt checksum file that
could not be re-signed cannot be published.
The MCP server itself is about 4,000 lines of plain ES modules with zero runtime dependencies and no build step: what is
on npm is what is in the repository, which you can read in one sitting. It makes no outbound network calls; a grep for
fetch(, http. and net. in server.mjs and healthstore.mjs comes back empty.
Checks you can run
Every claim above and below has a command. Fail closed on any mismatch.
# 1. Verify the site mirror's signature and checksums (needs minisign: brew install minisign)
PUBKEY='RWS6TxVWSKUblYkx7Db6ZpmvHALwHpznZpjaED/FlZj+PpxSlel0MxHZ' # pin from the GitHub repo
curl -fsSLO https://www.healthexport.dev/mcp/SHA256SUMS
curl -fsSLO https://www.healthexport.dev/mcp/SHA256SUMS.minisig
minisign -Vm SHA256SUMS -P "$PUBKEY" || { echo "signature INVALID: do not run"; exit 1; }
curl -fsSLO https://www.healthexport.dev/mcp/server.mjs
shasum -a 256 --ignore-missing -c SHA256SUMS || { echo "checksum mismatch: do not run"; exit 1; }
# 2. Verify the npm package's registry signature and provenance attestation
mkdir verify-hx && cd verify-hx && npm init -y >/dev/null
npm install health-export-mcp@1.8.0 && npm audit signatures
# 3. Confirm the server makes no outbound network calls
grep -nE 'fetch\(|https?\.|net\.|dns\.' node_modules/health-export-mcp/server.mjs node_modules/health-export-mcp/healthstore.mjs
# 4. Site headers, DNS and TLS
curl -sI https://www.healthexport.dev/ | grep -iE 'content-security-policy|strict-transport|x-frame|cross-origin|permissions-policy'
dig +short healthexport.dev CAA
dig +short _dmarc.healthexport.dev TXT
dig @1.1.1.1 +dnssec healthexport.dev A | grep -E 'flags:.* ad|RRSIG'
openssl s_client -connect www.healthexport.dev:443 -tls1_1 </dev/null 2>&1 | grep -i 'no protocols'
curl -s https://hstspreload.org/api/v2/status?domain=healthexport.dev
What runs on this site
Styles, fonts and images are served from this origin. Two things measure traffic, and we would rather name them than let you find them in the network tab.
- A self-hosted visitor counter (Umami): no cookies, nothing written to your device, no cross-site
identifier. Your browser talks only to this domain; a same-origin relay (
/u/) forwards the beacon to our collector server-side. It runs without asking because it stores nothing on your device to ask about, and the "Don't count this browser" switch turns it off. - A Meta advertising pixel: a third party that sets a cookie, so it does not load unless you say yes to the banner. We ask once, with the same sized button either way, and remember the answer.
Neither runs anywhere near your health data. The MCP server carries no telemetry at all. The iOS app carries no health telemetry; it does carry Meta's SDK for advert attribution, as set out above and in full on the privacy page.
Transport and domain hardening
Configured in site/_headers and at the DNS; each line is checkable with the commands above.
- Content-Security-Policy:
default-src 'self',object-src 'none',frame-ancestors 'none',base-uri 'self',form-action 'self',upgrade-insecure-requests. Two named exceptions, and only two:connect.facebook.net/www.facebook.comfor the opt-in pixel, andchallenges.cloudflare.comfor the Turnstile bot check on the sign-up and partner forms. No inline scripts (script-src 'self'); every behaviour on the site is a separate file you can read. - HSTS
max-age=63072000; includeSubDomains; preload, and the domain is on the browser preload list. - TLS 1.2 and 1.3 only; TLS 1.0 and 1.1 are refused.
- DNSSEC signed (DS record at the registry; validating resolvers return
ad). - DMARC
p=reject; sp=reject; adkim=s; aspf=swith SPF: mail claiming to be from this domain that is not from us is rejected, not quarantined. - CAA records restrict which authorities may issue certificates, with an
iodefcontact. - COOP and CORP
same-origin,X-Frame-Options: DENY,X-Content-Type-Options: nosniff,Referrer-Policy: strict-origin-when-cross-origin, and a Permissions-Policy that denies geolocation, camera and microphone.
Report a vulnerability
Please do not open a public issue for a security problem. Either route reaches the developer directly:
- Email [email protected].
- Open a private report on GitHub: Report a vulnerability (the repository's Security tab). The app is not in that repository; app issues go to the same email address.
Include the version (npx -y health-export-mcp --doctor prints the server's; the app's is under Settings),
what you observed and how to reproduce it. Use synthetic data (--demo serves a synthetic dataset): never
send real health data, pairing codes or export passphrases. Fixes ship as a new version on top of the latest release;
older versions are not patched. Machine-readable contact details, with an expiry date so you can tell the file is
maintained, are at /.well-known/security.txt (RFC 9116). The full policy for the
server is SECURITY.md.
What this page does not claim
- No independent penetration test or audit has been commissioned. Everything here is verifiable by you, not certified by a third party.
- Not "HIPAA compliant": HIPAA applies to health-care providers, insurers and their contractors, not to a consumer app you install yourself. The privacy page explains which rules do apply.
- Not "100% private": the export leaves the phone by your choice, and the advertising SDK sends what the privacy receipt says it sends, if you allow it.