MAQPNA

Healthcare and life sciences

Clinical agents that keep patient data at home.

Health data is the most sensitive data you hold, and it must stay in the country. MAQPNA runs agents next to your clinical systems, redacts patient data before it leaves the gateway, routes model calls in-country and records each action.

Why clinical agents are hard to approve

Privacy, residency and safety reviews come first.

  • Patient data in promptsNames, identifiers and diagnoses can leak into model prompts, tool arguments and logs.
  • Residency rulesHealth data often may not be processed outside the country, including by a model endpoint.
  • Safety-relevant actionsAn agent that writes to a record or schedules care needs a clinician in the loop.
  • Proof for review boardsEthics and data-protection reviews ask what the agent accessed and why.

How MAQPNA solves it

Controls for sensitive data

  • DLP and redaction

    Find and redact personal data in arguments, results and prompts before they leave the gateway.

  • In-country model routing

    Model routes with allow-lists per agent, model-for-data-class rules and residency checks.

  • Clinician approval

    Writes to clinical systems are held for a qualified approver, with the context they need.

  • Sandboxed tools

    Each session runs in an isolated sandbox with default-deny egress, so tools reach only what policy allows.

  • Governed memory

    Memory stores scoped per agent and user, with DLP on writes, retention and erasure with a signed certificate.

  • Access records

    Every access is a record with the session, the person the agent acted for, the rule and the decision.

Architecture

A clinical-summary agent inside the hospital network

The agent reads records through the gateway, which redacts identifiers before the model route sees them and holds writes for a clinician.

Hospital network · in-country · your keys

Sandboxes

tier-1 (microVM) or tier-2 (confidential VM)

  • clinical-summary
  • trial-ops
  • coding-assist

MAQPNA gateway

identity · policy · DLP · budgets · approval · audit

  • Allowedehr.read_encounterread-only-tools + DLP redact
  • Held for approvalehr.create_noteclinician approval
  • Deniedmodel route outside countryresidency

Your systems

  • MCPEHR adapter
  • MCPtrial database
  • modelin-country model route
Hash-chained audit ledger of every access → write-once storage (WORM)
Illustrative names. You write your own policies.

Evidence

Support for privacy and oversight work

Controls and evidence, mapped. Not a certification.

  • HIPAA Security Rule

    per-agent access control at the gateway and the hash-chained ledger help produce evidence for the access control and audit controls standards (45 CFR 164.312(a) and (b)); DLP keeps PHI out of prompts and logs.

  • HIPAA Security Rule update

    the overhaul proposed in January 2025 has slipped, with final action now expected in July 2027 (a provisional date); the current rule stays in force until then.

  • GDPR

    minimisation through DLP redaction, erasure with a signed certificate and argument capture off by default.

  • EU AI Act Art. 12 and 14

    record-keeping and human oversight controls. Under Regulation (EU) 2026/1744, Annex III high-risk duties apply from 2 December 2027 and Annex I products, such as medical devices, from 2 August 2028. How the EU AI Act applies to agents.

  • Residency

    sovereignty policy rules for model locations, registries and egress.

  • Honest scope

    MAQPNA limits what an agent can do; clinical validation of the model's output is yours.

FAQ

Common questions

What does MAQPNA do for healthcare and life sciences?

Health data is the most sensitive data you hold, and it must stay in the country. MAQPNA runs agents next to your clinical systems, redacts patient data before it leaves the gateway, routes model calls in-country and records each action.

What stops healthcare and life sciences from putting AI agents into production?

Patient data in prompts: Names, identifiers and diagnoses can leak into model prompts, tool arguments and logs. Residency rules: Health data often may not be processed outside the country, including by a model endpoint. Safety-relevant actions: An agent that writes to a record or schedules care needs a clinician in the loop. Proof for review boards: Ethics and data-protection reviews ask what the agent accessed and why.

Which controls does MAQPNA enforce?

DLP and redaction: Find and redact personal data in arguments, results and prompts before they leave the gateway. In-country model routing: Model routes with allow-lists per agent, model-for-data-class rules and residency checks. Clinician approval: Writes to clinical systems are held for a qualified approver, with the context they need. Sandboxed tools: Each session runs in an isolated sandbox with default-deny egress, so tools reach only what policy allows. Governed memory: Memory stores scoped per agent and user, with DLP on writes, retention and erasure with a signed certificate. Access records: Every access is a record with the session, the person the agent acted for, the rule and the decision.

Which regulations does it supply evidence for?

HIPAA Security Rule: per-agent access control at the gateway and the hash-chained ledger help produce evidence for the access control and audit controls standards (45 CFR 164.312(a) and (b)); DLP keeps PHI out of prompts and logs. HIPAA Security Rule update: the overhaul proposed in January 2025 has slipped, with final action now expected in July 2027 (a provisional date); the current rule stays in force until then. GDPR: minimisation through DLP redaction, erasure with a signed certificate and argument capture off by default. EU AI Act Art. 12 and 14: record-keeping and human oversight controls. Under Regulation (EU) 2026/1744, Annex III high-risk duties apply from 2 December 2027 and Annex I products, such as medical devices, from 2 August 2028. How the EU AI Act applies to agents. Residency: sovereignty policy rules for model locations, registries and egress. Honest scope: MAQPNA limits what an agent can do; clinical validation of the model's output is yours. MAQPNA supplies controls and evidence; it does not certify compliance.

Where does it run?

The agent reads records through the gateway, which redacts identifiers before the model route sees them and holds writes for a clinician. The same runtime runs on a laptop, in your cloud account, on Kubernetes or air-gapped.

Can AI agents meet HIPAA Security Rule access and audit controls?

Per-agent access control at the gateway and the hash-chained ledger help produce evidence for the access control and audit controls standards (45 CFR 164.312(a) and (b)); DLP keeps PHI out of prompts and logs.

Bring an agent to a clinical workflow.

Policy, approval flow and evidence, walked through with you.

More install options
curl -fsSL https://maqpna.com/install.sh | sh