How we secure the stack.
A short read on the architecture, the controls, and what we'll show you on a security review.
Architecture
- Layer 01 · Context (chanio): user-owned context graph. The hard drive metaphor is literal — the data is yours, on your device, federated through your own infrastructure.
- Layer 02 · Runtime (harnesshealth.ai): agentic orchestration on Anthropic's Claude. A proprietary harness, designed to be model-agnostic, with physician review built into the design.
- Layer 03 · Proof (hashcare): public, hash-only attestation ledger. When a physician attests a document, a hash of that attestation is recorded there so a later change is detectable. No PHI is stored on it.
Controls
- Physician sign-off (design) — the harness is designed so no agent-written clinical content reaches a patient or third party without a named physician's signature. It runs in the sandbox today; it is not yet enforced in production.
- Encryption at rest and in transit, as provided by our hosting and database vendors (Vercel, Google Cloud).
- Access via Firebase Authentication (Google Cloud) sign-in, checked on our servers. SSO planned for enterprise.
- Attestation records carry the signing physician's NPI and a timestamp. We have not yet published a retention schedule.
- Secrets stored in Vercel encrypted environment variables.
- No third-party trackers on any production surface. No Facebook pixel, no Google ads, no Hotjar, no session replay tools.
The AI piece
The chat assistant on our sites runs on Anthropic's Claude through Anthropic's commercial API. Under Anthropic's commercial terms, conversations are not used to train its models by default; Anthropic keeps them for a limited period under its own retention policy. The instructions each site's assistant follows are public at /kb/registry.json.
What we'll show you on a review
- Sample attestation records (test data — no PHI flows before a Business Associate Agreement is signed).
- Vercel's and Google Cloud's own SOC 2 reports, which those vendors provide on request.
- Anthropic's commercial terms covering the chat assistant.
Reporting a vulnerability
Found something? Report it through the contact page and say it is a security report. There is no bug-bounty program and no guaranteed response time — the founder reads what comes in. We will credit responsible disclosure by name if you want the credit.
For HIPAA-specific posture, see the HIPAA page.