fhir-r6-guardrails

HealthClaw Guardrails

A healthclaw.io open source project. Reference implementation of security and compliance patterns for AI agent access to FHIR data via MCP.

Supports FHIR R4 US Core v9 (stable) and FHIR R6 v6.0.0-ballot3 (experimental).

This is a runtime guardrail layer, not a knowledge skill. It sits between any AI agent and FHIR data (local or upstream), enforcing PHI redaction, audit trails, step-up authorization, and tenant isolation on every request.

When to Use This Skill

• You need to read, search, or write FHIR clinical resources through MCP
• You need PHI to be automatically redacted before the agent sees it
• You need an immutable audit trail of all agent access
• You need step-up authorization gates on write operations
• You need to evaluate R6 Permission resources for access control

MCP Tools Available (12)

Read Tools (no step-up required)

Write Tools (require step-up token)

Two-Phase Write Pattern

Writes always follow propose-then-commit:

1. Propose: Call fhir.propose_write with the resource and operation (create/update). This validates the resource and returns a preview. No data is changed.

2. Review: Check the proposal response:

   • proposal_status: "ready" means validation passed
   • requires_human_confirmation: true for clinical resource types
   • requires_step_up: true always for commits

3. Commit: Call fhir.commit_write with the same resource. Include:

   • X-Step-Up-Token header (HMAC-SHA256 signed, 5-minute TTL)
   • X-Human-Confirmed: true header for clinical resources

Clinical Resource Types (require human-in-the-loop)

Observation, Condition, MedicationRequest, DiagnosticReport, AllergyIntolerance, Procedure, CarePlan, Immunization, NutritionIntake, DeviceAlert.

Supported FHIR Resource Types

FHIR R4 US Core v9 (Stable)

Patient, Encounter, Observation, AuditEvent, Consent, Condition, Provenance, AllergyIntolerance, Immunization, MedicationRequest, Medication, MedicationDispense, Procedure, DiagnosticReport, CarePlan, CareTeam, Goal, DocumentReference, Location, Organization, Practitioner, PractitionerRole, RelatedPerson, Coverage, ServiceRequest, Specimen, FamilyMemberHistory.

FHIR R6 ballot3 (Experimental)

Permission, SubscriptionTopic, Subscription, NutritionIntake, NutritionProduct, DeviceAlert, DeviceAssociation, Requirements, ActorDefinition.

• Permission — Access control (separate from Consent). $evaluate operation.
• DeviceAlert — ISO/IEEE 11073 device alarms.
• NutritionIntake — Dietary consumption tracking.
• DeviceAssociation, NutritionProduct, Requirements, ActorDefinition — CRUD only.

Security Guardrails (Always Active)

PHI Redaction

Applied on every read path. Names truncated to initials, identifiers masked to last 4 characters, addresses stripped to city/state/country, birth dates truncated to year, photos removed, notes replaced with [Redacted].

Audit Trail

Append-only AuditEvent records for every resource access. Database-level immutability. Logs agent ID, tenant ID, resource accessed, and outcome.

Step-Up Authorization

HMAC-SHA256 tokens with 128-bit nonce, 5-minute TTL, tenant binding. Required for all write operations.

Tenant Isolation

X-Tenant-Id header enforced on every query at the database layer.

Search Parameters (Local Mode)

Supported: patient (reference), code (token), status (token), _lastUpdated (date with ge/le/gt/lt prefix), _count (1-200), _sort (_lastUpdated/-_lastUpdated), _summary (count).

Not supported in local mode: chaining, _include, _revinclude, modifiers.

In upstream proxy mode, all query parameters are forwarded to the upstream server.

Setup

# Docker Compose (recommended)
docker-compose up -d --build

# Manual
uv sync && python main.py &
cd services/agent-orchestrator && npm ci && npm start

Required environment variables:

• STEP_UP_SECRET — HMAC secret for step-up tokens

Optional:

• FHIR_UPSTREAM_URL — Connect to a real FHIR server (e.g., https://hapi.fhir.org/baseR4)
• FHIR_BASE_URL — Backend URL (default: http://localhost:5000/r6/fhir)
• MCP_PORT — MCP server port (default: 3001)

Known Limitations

• Local mode uses SQLite JSON blob storage (not indexed search)
• Structural validation only (no StructureDefinition or terminology binding)
• SubscriptionTopic stored but notifications not dispatched
• Human-in-the-loop is header-based, not cryptographic
• No historical versioning (version_id increments but old versions not retrievable)
• Upstream proxy: no response caching, no cross-version translation