MCP Demo - Learning Project

This repository documents my exploration of Model Context Protocol (MCP) implementation. I built this project to understand how MCP enables AI models to interact with external tools and services.

Project Structure:

• Part 1: Basic MCP mechanics and transport methods
• Part 2: Real LLM integration with OpenAI API

This hands-on project helped me understand MCP fundamentals through building a calculator tool that works with different transport mechanisms.

How MCP Works

The diagram above illustrates the Model Context Protocol architecture:

• MCP Client (hosted in Claude, IDEs, or other tools) connects to multiple MCP servers
• MCP Servers act as bridges between the client and various data sources
• Data Sources can be local (databases, files) or remote (web APIs, cloud services)
• All communication happens via the MCP Protocol, enabling standardized tool access

Part 1: Basic MCP Implementation

What I Built:

• Calculator tool with basic arithmetic operations
• MCP server that exposes the calculator via @mcp.tool() decorator
• Two different client implementations demonstrating transport methods

Key Learning:

• How MCP servers register and expose tools
• Client-server communication patterns
• Transport mechanisms: stdio vs SSE

Running the Code:

# STDIO Transport - client manages server lifecycle
python client-stdio.py

# SSE Transport - manual server startup required
python basic_mcp_server.py  # Terminal 1
python client-sse.py        # Terminal 2

Part 2: LLM Integration

What I Built:

• MCPOpenAIClient class that integrates MCP with OpenAI API
• Automatic tool discovery and execution by AI models
• Query processing system where AI decides when to use tools

Key Learning:

• How to bridge MCP tools with LLM APIs
• Tool format conversion between MCP and OpenAI schemas
• Managing conversation context with tool results
• Resource cleanup and error handling patterns

Running the Code:

cd 2_MCP_OpenAI_Integration/
# Add OPENAI_API_KEY to .env file
python client.py

What Happens:

1. AI receives user query + available MCP tools
2. AI automatically decides if tools are needed
3. Tools execute via MCP, results integrated into conversation
4. AI provides natural response using tool results

Key Differences

Part 1	Part 2
Manual tool calls	AI decides automatically
Learning MCP mechanics	Real AI tool integration
Technical demonstration	Natural conversation

What I Learned

This project helped me understand:

• MCP's role in extending AI capabilities with external tools
• Different transport mechanisms and their use cases
• How to integrate MCP with popular LLM APIs
• The difference between manual tool calling and AI-driven tool usage
• Resource management in async Python applications