Introduction of Building MCP

• Build an MCP Server that exposes your backend as AI tools for LLM to call to use.
• MCP server should:
  • NOT contain heavy logic
  • Just orchestrate services

Frontend (React)
     ↓
Backend (Spring Boot / Node)
     ↓
MCP Server Layer
     ↓
Microservices (AI, DB, Redis, Storage)

Build MCP Steps (Python FastMCP):

1. Install MCP SDK

   npm init -y
   npm install @modelcontextprotocol/sdk
   npm install zod
   

2. Basic MCP server

   import os
   import httpx
   from mcp.server.fastmcp import FastMCP
   from backend.core.drone_registry import registry
   
   mcp = FastMCP(
       name="rescue-drone-mcp",
   	  version: "1.0.0",
       instructions=(
           "You are the MCP interface for a rescue drone fleet operating in an earthquake disaster zone. "
           "Call get_map_info and list_active_drones first. "
           "Use start_search (not move_to + thermal_scan) to begin an autonomous sweep — it returns immediately. "
           "Then loop on wait_for_event: act on survivor_found with delivery_aid, on battery_low with return_to_base, "
           "and exit the loop when all drones have sent search_complete."
       ),
   )
   

3. Build a tool

   • Name must be **clear + action-oriented (**Verb)
   • Parameters should define what AI must provide (strict + validated)
   • Return Value must be structured + predictable (json friendly)

   @mcp.tool()
   async def get_battery_status(drone_id: str) -> dict:
       """
       Return the battery percentage for a specific drone.
   
       When battery is below 20%, call request_backup immediately.
       Battery charges gradually at 5%/s — poll this to track charging progress.
       """
       drone = registry.get(drone_id)
       if not drone:
           return {"error": f"Drone '{drone_id}' not found."}
   
       url = f"http://{drone.host}:{drone.port}"
       async with httpx.AsyncClient() as client:
           resp = await client.get(f"{url}/drones/{drone_id}/battery", timeout=3.0)
       return resp.json()
   

4. Start the server

   mcp.run()
   

Types of MCP Tools

1. Read Tools

• No side effects. (safe)