Wrap an EHR (FHIR R4) for a care agent

Context

A care coordination agent needs an EHR. The MCP server exposes read tools over encounters, labs, and meds, but the only mutation is a propose_order that returns pending clinician approval, PHI stays behind the server and nothing writes to the chart autonomously.

The decision

The design decision is which resources to expose and where the write approval boundary sits: reads are broad, the single write is a proposal a human ratifies.

What most miss

Teams expose a write_order tool 'for convenience' and quietly hand an agent chart write authority. The boundary is the whole safety design, not a feature flag.

Stakes

An autonomous chart write is a patient safety and liability event; a proposal queued for a clinician is not.