Langcare MCP FHIR

LangCare MCP FHIR Server

Enterprise-grade MCP Server for FHIR-based EMRs. Fully written in Go with enterprise-grade security and 4 generic FHIR operations that work with any FHIR R4 resource type. Supports EPIC, Cerner, OpenEMR, GCP Healthcare API, and any generic FHIR R4 server.

Ships with a [40+ Clinical Skills Library](skills/README.md) — agent-agnostic workflow guides covering medication management, lab interpretation, clinical decision support, documentation, population health, and more. Extend with ✨ New: [Claude Managed Agents](cma/README.md) (9 production-ready clinical AI agents on the Anthropic Managed Agents API), [MCP Apps](apps/README.md) (interactive clinical UIs inside Claude Desktop), a [Healthcare Voice Agent](pipecat-agent/README.md) (real-time voice AI over FHIR), and a [LangCare CLI](cli/README.md) for agent frameworks that don't speak MCP natively.

langcare.ai

Installation

Install via npm:

npm install -g @langcare/langcare-mcp-fhir

Or use directly without installation:

npx @langcare/langcare-mcp-fhir -config /path/to/config.yaml

Quick Configuration

LangCare MCP FHIR connects Claude to your FHIR-based EMR system. You need a YAML configuration file pointing to your backend.

1. Get a Config Template

Choose your backend:

• EPIC: config.epic.example.yaml
• Cerner: config.cerner.example.yaml
• OpenEMR: config.openemr.example.yaml
• GCP Healthcare API: config.gcp.example.yaml
• Any FHIR R4 Server: config.base.example.yaml

2. Configure Claude Desktop

Add to your Claude Desktop config file (~/.config/Claude/claude_desktop_config.json):

{

"mcpServers": { "langcare-mcp-fhir": { "command": "langcare-mcp-fhir", "args": ["-config", "/path/to/your/config.yaml"] } } }

On macOS, the config is typically at:

~/Library/Application\ Support/Claude/claude_desktop_config.json

3. Restart Claude Desktop

Close and reopen Claude Desktop. The FHIR tools will now be available.

Need detailed setup help? See the Local Testing Guide.

Architecture

This MCP server acts as an intelligent proxy between AI agents and FHIR R4 servers. It exposes 4 generic FHIR operations through the Model Context Protocol (MCP), enabling AI-powered workflows for any FHIR resource type.

Key Design:

• MCP SDK: Official github.com/modelcontextprotocol/go-sdk (Anthropic/Google maintained)
• FHIR Client: Generic HTTP client working with any FHIR R4 server
• Transport: stdio and Streamable HTTP
• Backend: Proxy to existing FHIR server (no database)
• Language: 100% Go for high performance and reliability

4 Generic MCP Tools

All tools work with any FHIR resource type (Patient, Observation, Medication, etc.):

1. fhir_read

Read a FHIR resource by type and ID.

{

"resourceType": "Patient", "id": "example-123" }

2. fhir_search

Search FHIR resources with query parameters.

{

"resourceType": "Patient", "queryParams": "name=John&birthdate=gt1990-01-01" }

3. fhir_create

Create a new FHIR resource.

{

"resourceType": "Observation", "resource": { "resourceType": "Observation", "status": "final", "code": { ... }, "subject": { "reference": "Patient/123" } } }

4. fhir_update

Update an existing FHIR resource.

{

"resourceType": "Patient", "id": "example-123", "resource": { "resourceType": "Patient", "id": "example-123", "name": [{ "family": "Smith" }] } }

Security Architecture

LangCare MCP FHIR implements a two-layer security model for HIPAA-compliant healthcare data access:

┌─────────────┐         ┌──────────────┐         ┌─────────────┐

│ Claude │ Auth1 │ MCP Server │ Auth2 │ FHIR API │ │ Client │────────▶│ (Go) │────────▶│ (EMR) │ └─────────────┘ └──────────────┘ └─────────────┘

Auth1: MCP Client Authentication (Bearer Token/API Key) Auth2: FHIR Backend Authentication (Bearer/OAuth2/SMART on FHIR)

Security Features

• ✅ TLS 1.3 encryption for HTTP transport
• ✅ PHI Scrubbing in logs (enabled by default)
• ✅ HIPAA-compliant audit logging
• ✅ No persistent PHI storage (stateless proxy)
• ✅ Secrets via environment variables (never in config files)
• ✅ OAuth 2.0 with automatic token refresh
• ✅ mTLS support for service-to-service communication
• ✅ Rate limiting per client

Supported Authentication Methods

• Bearer Token - Simple API key authentication
• OAuth2 - Full OAuth2 flow with token refresh
• SMART on FHIR Backend Services - private_key_jwt (RS384) for EPIC, OpenEMR, and other SMART-conformant EMRs
• SMART on FHIR - EPIC, Cerner, OpenEMR, and other EMR standards
• Basic Auth - Username/password authentication
• Custom - Extensible for additional auth methods

For complete security documentation, see [Security Guide](https://github.com/langcare/langcare-mcp-fhir/blob/main/docs/SECURITY.md):

• HIPAA compliance checklist
• OAuth configuration for EPIC/Cerner/GCP
• Kubernetes security manifests
• Credential management procedures
• Audit logging implementation

MCP Apps (Interactive UIs)

LangCare MCP FHIR ships with built-in MCP Apps — interactive, rich UI views that run directly inside MCP-capable hosts like Claude Desktop. Unlike traditional chat-based tool output, MCP Apps render full React-based interfaces with charts, tables, and interactive controls while using the same underlying FHIR tools.

How it works: Each app is a single-file HTML bundle (React + TypeScript, compiled with Vite) that gets embedded into the Go binary at compile time via go:embed. At runtime the MCP server registers each app as both an MCP Resource (text/html;profile=mcp-app) and a dedicated MCP Tool linked via _meta.ui.resourceUri. When an MCP host calls the tool, it fetches the resource and renders the UI. The app calls back into the server's generic FHIR tools (fhir_search, fhir_read, etc.) through app.callServerTool() — no LLM round-trips for data fetching.

Advantages over plain tool output:

• Rich visualization — SVG charts, color-coded cards, expandable detail panels
• Interactive controls — search fields, date range pickers, click-to-expand rows
• Deterministic data fetching — apps call FHIR tools directly, no LLM involvement in data retrieval
• Zero external dependencies — everything inlines into a single HTML file, embedded in the binary
• Works offline — no CDN, no external scripts, no network requests beyond FHIR API calls

Built-in Apps

Both apps are reference implementations demonstrating the MCP Apps pattern. See [apps/README.md](apps/README.md) for architecture details and how to build new apps.

Patient Chart Review running inside Claude Desktop

Agent Usage

AI agents use LangCare MCP FHIR Server to help healthcare professionals access and manage patient health records through 4 FHIR tools. The server handles EMR authentication, allowing agents to focus on clinical workflows while maintaining strict privacy and accuracy standards.

Agent capabilities:

• Search, Read, Create, Update - Any FHIR R4 resource (Patient, Observation, Medication, etc.)
• Patient privacy - Use partial identifiers, confirm identity before updates
• Clinical accuracy - Verify data, use standard codes (LOINC, SNOMED, RxNorm)
• Professional communication - Structure responses with context, findings, and next steps

Common workflows:

• Patient lookup: Search by name/DOB → verify identity → read full details
• Clinical review: Retrieve labs, vitals, medications → present with reference ranges
• Documentation: Extract structured data → map to FHIR resources → confirm → create
• Updates: Verify existing resource → modify → confirm changes → update

System support:

• Works with any FHIR R4 resource type (60+ types including DocumentReference, Binary, Media)
• Automatic authentication and token refresh to EPIC, Cerner, OpenEMR, GCP Healthcare API
• HIPAA-compliant PHI handling with audit logging
• Comprehensive OAuth2 scopes for clinical data access

📖 Complete guide: Agent Prompt Guide - System prompt, tool examples, workflows, and error handling

Clinical Skills Library (Optional)

40+ agent-agnostic clinical workflow guides that teach AI agents how to perform complex healthcare tasks using the MCP server's 4 FHIR tools (fhir_search, fhir_read, fhir_create, fhir_update).

• Optional - The MCP server works without them
• Portable - Work with Claude, ChatGPT, Gemini, or any AI agent
• Evidence-based - Built on USPSTF, ADA, ACC/AHA, CDC, ACOG, KDIGO, and other society guidelines
• Copy-paste ready - Add a skill's SKILL.md to your agent's system prompt or custom instructions

Skill Categories (40 Skills)

| **Population Healt