MCP Pharma Assistant: An Intelligent MCP Agent for Healthcare

An open-source, modular AI assistant for the pharmaceutical sector built using the Model Context Protocol (MCP).

This project transforms generic Large Language Models (like Claude) into deterministic, hallucination-free clinical agents by seamlessly connecting them to official healthcare databases, live APIs, and advanced chemoinformatics tools.

The Problem & The Solution

The Problem: Doctors and pharmacists need access to highly heterogeneous data (drug availability, reimbursement rules, molecule interactions, clinical trials). Generic LLMs suffer from stochastic hallucinations on specific numerical data and lack real-time access to official government or scientific databases.

The Solution: By leveraging the Model Context Protocol, this architecture provides an LLM with external "Tools" to fetch deterministic data. The agent is strictly grounded in official sources, ensuring:

Zero Hallucinations on prices, dosages, and supply chain shortages.
Real-Time Data via live APIs for clinical trials and pharmacovigilance.
Advanced Scientific Capabilities like molecular cross-referencing and dynamic visual rendering.

System Architecture

The ecosystem relies on an Orchestrator (Claude Desktop) connected via stdio to three specialized micro-servers built with FastMCP:

1. Italian Pharma Server (Local Sovereignty)

Manages official Italian supply chain data (AIFA - Agenzia Italiana del Farmaco).

Tech: DuckDB, SQL, @lru_cache, Pydantic.
Features: Instant lookups for commercial status, hospital shortages (Class H), public pricing, and generic alternatives.
Performance: Complex view queries resolved in ≈23ms, dropping to <0.001ms for cached repeat queries.

2. Global Knowledge Server (Big Data & Chemoinformatics)

Acts as the analytical engine for molecular data and drug-drug interactions (DDI) based on DrugBank.

Tech: Polars, RDKit, Pillow.
Features: Resolves entity synonyms, detects interaction conflicts, parses SMILES strings, and generates dynamic 2D molecule structures & LogP (lipophilicity) heatmaps natively in the chat.

3. External API Server (Live Knowledge)

An asynchronous gateway to external health authorities.

Tech: httpx, asyncio.
Features: Non-blocking API calls to OpenFDA (for real-world adverse event tracking) and ClinicalTrials.gov (to geolocate actively recruiting experimental trials).

Data Engineering & ETL Pipelines

A significant part of this project involves robust ETL pipelines to clean and prepare raw medical data for the MCP servers.

Parallel XML Parsing (DrugBank): Bypassed the Python GIL using ProcessPoolExecutor and memory mapping (mmap) to process a highly nested 1.5GB XML file. Converted the data into analytical Parquet lakes using Polars for zero-copy memory efficiency, achieving a 97.5% disk space reduction.
Smart CSV Ingestion (AIFA): Automated encoding detection (chardet), regex-based header cleaning, and relational mapping (using AIC codes as universal Foreign Keys) to build a unified in-process DuckDB data warehouse.

Installation & Setup

Prerequisites

Python 3.10+
Claude Desktop App (for MCP integration)

1. Clone & Install

git clone https://github.com/yourusername/mcp-pharma-assistant.git
cd mcp-pharma-assistant
pip install -r requirements.txt

2. Build the Databases (ETL)

Ensure you have the raw CSVs/XMLs in the 0_raw_files folder, then run:

# Build the DuckDB AIFA database
python build_aifa_database.py

# Parse DrugBank XML into Parquet files
python process_drugbank_parallel.py

3. Configure Claude Desktop

Add the FastMCP servers to your claude_desktop_config.json:

{
  "mcpServers": {
    "ItalianPharmaServer": {
      "command": "python",
      "args": ["/absolute/path/to/italian_pharma_server.py"]
    },
    "GlobalKnowledgeServer": {
      "command": "python",
      "args": ["/absolute/path/to/global_knowledge_server.py"]
    },
    "ExternalAPIServer": {
      "command": "python",
      "args": ["/absolute/path/to/external_api_server.py"]
    }
  }
}

MCP Servers