atlas_mcp

An MCP server that brings AI-powered search and conversation to your FHIR clinical documents.

What It Does

atlas_mcp is a developer-focused MCP server for working with FHIR data. It lets you embed FHIR resources, search them with semantic retrieval, and talk to an AI agent that can answer questions with citations from your clinical documents.

• AI agent that understands and queries FHIR documents

• Semantic search with cross-encoder reranking for accuracy

• Multi-turn conversations with session memory

• Local-first LLM support (Ollama), plus cloud options (OpenAI, Anthropic, Bedrock)

Key Features

• Native FHIR resource handling and metadata extraction

• Vector embeddings plus access to full documents

• Built-in validation and HIPAA-aware prompts

• YAML + environment configuration for easy setup

Quick Start (5 Minutes)

1) Clone and Install

git clone https://github.com/rsanandres/atlas_mcp.git
cd atlas_mcp
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate
pip install -r requirements.txt

2) Set Up PostgreSQL + pgvector

createdb hc_ai
psql -U postgres -d hc_ai -f scripts/setup_db.sql

3) Configure Environment

cp env.example .env
# Edit .env and set DB_PASSWORD at minimum

4) Start Ollama (Local-First)

ollama pull mxbai-embed-large:latest
ollama pull llama3
ollama serve

5) Run the Server

# stdio transport (Claude Desktop, Cursor)
python server.py

# HTTP transport
python server.py --transport streamable-http --port 8000

Architecture

flowchart TB
    subgraph clients [MCPClients]
        Claude[Claude Desktop]
        Cursor[Cursor IDE]
        Custom[Custom Client]
    end

    subgraph server [AtlasMcpServer]
        MCP[MCP Protocol Layer]
        Tools[Tool Registry]
        Agent[LangGraph Agent]
        Reranker[Cross-Encoder Reranker]
        Session[Session Store]
    end

    subgraph backends [Backends]
        LLM[LLM Provider]
        PG[(PostgreSQL + pgvector)]
        Embed[Embedding Service]
    end

    clients --> MCP
    MCP --> Tools
    Tools --> Agent
    Tools --> Reranker
    Tools --> Session
    Agent --> LLM
    Agent --> PG
    Reranker --> PG
    Embed --> PG

Available Tools

Agent tools

• agent_query, agent_clear_session, agent_health

Retrieval tools

• rerank, rerank_with_context, batch_rerank

Session tools

• session_append_turn, session_get, session_update_summary, session_clear

Embeddings tools

• ingest, embeddings_health, db_stats, db_queue, db_errors

Example Use Cases

• Querying patient records: “What medications is patient P123 taking?”

• Lab results analysis: “Show abnormal lab values from the last 30 days.”

• Clinical notes search: “Find notes mentioning diabetes management.”

• Medication history: “Has this patient been prescribed blood thinners?”

Configuration

Tool Configuration

Enable/disable tools in config.yaml:

tools:
  agent_query:
    enabled: true
  rerank:
    enabled: true
  ingest:
    enabled: false

LLM Providers (Local-First)

• Ollama: LLM_PROVIDER=ollama, LLM_MODEL=llama3

• OpenAI: LLM_PROVIDER=openai, OPENAI_API_KEY, OPENAI_MODEL=gpt-4o-mini

• Anthropic: LLM_PROVIDER=anthropic, ANTHROPIC_API_KEY, ANTHROPIC_MODEL=claude-3-5-sonnet-20241022

• Bedrock: LLM_PROVIDER=bedrock, AWS_REGION, LLM_MODEL=haiku|sonnet|opus

Environment Variables

See env.example for all options. Core requirements:

Debug Logging

HC_AI_DEBUG=true

Timeouts

AGENT_TIMEOUT=60
RERANK_TIMEOUT=30

Connecting to MCP Clients

Claude Desktop

{
  "mcpServers": {
    "atlas": {
      "command": "python",
      "args": ["/path/to/atlas_mcp/server.py"],
      "env": {}
    }
  }
}

Cursor IDE

{
  "atlas": {
    "command": "python",
    "args": ["/path/to/atlas_mcp/server.py"]
  }
}

Example Usage

result = await client.call_tool("agent_query", {
    "query": "What medications is patient P123 currently taking?",
    "session_id": "session-001",
    "patient_id": "P123"
})

Requirements

• Python 3.11+

• PostgreSQL 14+ with pgvector

• Ollama (or cloud LLM credentials)

Disclaimer

This project is HIPAA-aware, but it is not HIPAA-certified. It is intended for development and testing only. You are responsible for compliance and security if you use it in production.

Author

Created by @rsanandres. Issues and feedback welcome. Pull requests are reviewed.

License

MIT License. See LICENSE.