MCP — Model Context Protocol for Healthcare

Model Context Protocol is Anthropic's open standard for connecting LLM agents to external tools, data, and resources. For healthcare AI builders, it's the cleanest way to give Claude (or any MCP-aware model) live, scoped, auditable access to FHIR servers, terminology services, EHRs, and clinical knowledge bases.

Why MCP fits healthcare

Three properties matter:

1. Tool boundaries are explicit. Every MCP server declares its tools by name + schema. You can audit what the agent could call, not just what it did call.
2. Resources are typed. A fhir://Patient/123 resource is different from a free-text blob. Logging is auditable down to the resource type.
3. Servers are composable. Run one MCP server per backend (Medplum, terminology, FHIRBuilders sandbox, GitHub) and let the agent pick the right tool for the question.

Reference setup for a Cohort 00 build

Claude Code (agent)
    ├── MCP: FHIR REST (Medplum sandbox)
    │     · patient_search, observation_read, condition_create
    ├── MCP: Terminology
    │     · code_lookup, valueset_expand
    ├── MCP: Vector DB (clinical guidelines)
    │     · semantic_search
    └── MCP: GitHub
          · file_read, pr_create

Each server runs locally during dev (stdio transport) and as a hosted service in prod (SSE or HTTP).

Patterns to copy

• Allow-list tools per session. Even if a server exposes 20 tools, your launcher should only expose the ones this session needs.
• PHI redaction at the boundary. See HealthClaw — it sits between the agent and FHIR, redacting PII before context reaches the model and re-injecting on tool calls.
• Tool-call audit log. Persist every tool invocation + arguments + return summary. This is your compliance + debugging trail.
• Structured output for clinical claims. When the agent generates anything clinical-facing, use the model's tool_use to force a structured Recommendation schema instead of free text.

Where to learn more

• The MCP spec at modelcontextprotocol.io — start here
• The MCP server registry at github.com/modelcontextprotocol/servers — reference implementations
• HealthClaw (healthclaw.io) — Eugene's open security layer between agents and clinical data, MCP-native
• The Cohort 00 Session 2 workshop ("MCP servers + vector DB") is the cohort's hands-on intro

The honest caveat

MCP is the right abstraction for healthcare agents — but the standard is young and the ecosystem is still settling on conventions for auth, multi-tenancy, and observability. Expect to write some glue. The bet is that this glue is the right glue.