Clinical Reasoner MCP server

fhir-mcp-suite

Three coherent MCP servers for clinical AI — FHIR R4, terminologies, and clinical reasoning.

What's in the suite

Why this suite is different

Every FHIR MCP server available today (June 2026) is a read proxy — they retrieve resources but never tell you whether the resource is valid. mcp-fhir adds HAPI profile validation as a first-class MCP tool. Composing fhirread → validateagainst_profile in one Claude session enables clinical AI pipelines that are actually safe.

Three sharp differentiators:

1. Profile validation built into mcp-fhir — HAPI validator sidecar, US Core + IPS profiles supported out of the box
2. Composable suite — three coherent servers sharing one install, one config convention, one eval harness
3. Production rigor — latency benchmarks, golden-query eval suite, structured JSON logs, /health + LangFuse traces

Quick start — mcp-fhir

# 1-command install (requires Python 3.12+)
uvx mcp-fhir

# Validate a Patient against US Core
# (requires HAPI validator sidecar — see docker-compose.yml)
uvx mcp-fhir --transport sse  # or set MCP_TRANSPORT=sse

Claude Desktop

Add to ~/Library/Application Support/Claude/claudedesktopconfig.json (Windows: %APPDATA%\Claude\claudedesktopconfig.json):

{
  "mcpServers": {
    "fhir": {
      "command": "uvx",
      "args": ["mcp-fhir"],
      "env": {
        "FHIR_BASE_URL": "https://hapi.fhir.org/baseR4"
      }
    },
    "terminology": {
      "command": "uvx",
      "args": ["mcp-terminology"]
    },
    "clinical-reasoner": {
      "command": "uvx",
      "args": ["mcp-clinical-reasoner"]
    }
  }
}

Local dev stack

# Start HAPI FHIR + validator + Postgres
docker compose up hapi-fhir hapi-validator postgres

# Install workspace
uv sync

# Run unit tests
uv run pytest -m "not integration and not eval"

# Run mcp-fhir locally (stdio, points at local HAPI)
FHIR_BASE_URL=http://localhost:8081/fhir \
HAPI_VALIDATOR_URL=http://localhost:8082 \
  uv run mcp-fhir

Repo layout

Shortened. The full README is on GitHub.