MCPHealthContextOctopus

MCPHealthContextOctopus AI: MCP-Powered Patient Insight Engine:

Inspiration:

Healthcare data is often scattered across multiple systems—lab results, vitals, and clinical records are not unified. This project was inspired by the need to: create a single source of truth for patient health context enable dynamic tool discovery via MCP integrate FHIR-compliant healthcare data with AI agents reduce manual effort in interpreting patient data

What it does

Retrieves patient data using FHIR APIs Exposes clinical tools via MCP server Aggregates: Blood Pressure Cholesterol Triglycerides Builds a structured patient health context Enables A2A + MCP hybrid architecture Uses LLM (Grok / Gemini) for reasoning and summarization

How we built it

Core Architecture Custom Agent → MCP Server → Tools → FHIR API pipeline

Pipeline Components Agent Layer PatientHealthContextAgent System prompt-driven context builder MCP Server (mcp-health-context-server) Tool mapping layer Exposes: GetPatientBloodPressure GetPatientCholesterol GetPatientTriglycerides FHIR Layer fhir_client.py API-based retrieval using secure token Utilities mcp_utilities.py → response formatting mcp_instance.py → MCP server initialization API Layer FastAPI ngrok for external exposure

Context Handling

FHIR context injected via middleware / environment

Challenges we ran into

1. Unable to fetch latest data uploaded FHIR returned older observations instead of latest _sort was unreliable across server implementations Required manual sorting using effectiveDateTime
   1. MCP Tool Mapping Complexity Mapping tools dynamically to MCP server required careful design Ensuring consistent input/output contracts
   2. Data Consistency Issues Multiple records for same metric (e.g., triglycerides) Needed validation and filtering (final vs preliminary)
   3. Context Propagation Passing FHIR context correctly across agent → MCP → tools

Accomplishments that we're proud of

Built a custom MCP server for healthcare tools Successfully integrated FHIR APIs with AI agents Enabled dynamic tool usage via MCP skills layer Created a scalable multi-agent + MCP architecture Solved real-world issue of clinical data inconsistency

Built With

• backend
• configuration
• cutomagent
• geminiflash
• llm
• logic
• mcpserver
• mcptools
• ngrok
• promtpopinion
• python