Advanced MCP (Server & Client) Practice Task

Python implementation for building AI Agent with MCP (Model Context Protocol) tools and MCP server/client architecture.

🎯 Task Overview

Create and run an MCP server with custom tools, then implement an AI Agent with MCP Client that utilizes tools from the created server. This task demonstrates the full MCP workflow from server implementation to client integration.

🎓 Learning Goals

By completing this project, you will learn:

• MCP Protocol Implementation: Understand the Model Context Protocol specification and JSON-RPC communication
• Server-Side Tool Development: Create custom tools that follow MCP standards
• Client Integration: Connect AI agents to MCP servers and handle tool execution
• Session Management: Implement proper session handling and state management
• Streaming Responses: Work with Server-Sent Events (SSE) for real-time communication
• Error Handling: Implement robust error handling in distributed systems

🏗️ Architecture

├── agent/                        # MCP Client Implementation
│   ├── clients/
│   │   ├── custom_mcp_client.py    🚧 TODO: Pure Python MCP client
│   │   ├── mcp_client.py           ✅ Complete: Framework-based client
│   │   └── dial_client.py          ✅ Complete: AI model integration
│   ├── models/           
│   │   └── message.py              ✅ Complete: Message structures
│   └── app.py                      🚧 TODO: Test it with MCPClient and CustomMCPClient
└── mcp_server/                   # MCP Server Implementation
    ├── models/
    │   ├── request.py              ✅ Complete: Request model
    │   └── response.py             ✅ Complete: Response model
    ├── services/
    │   └── mcp_server.py           🚧 TODO: Implement core server logic
    ├── tools/
    │   ├── base.py                 ✅ Complete: Abstract tool interface
    │   ├── create_user_tool.py     🚧 TODO: Implement web search tool
    │   ├── delete_user_tool.py     🚧 TODO: Implement web search tool
    │   ├── update_user_tool.py     🚧 TODO: Implement web search tool
    │   ├── get_user_by_id_tool.py  🚧 TODO: Implement web search tool
    │   └── search_users.py         🚧 TODO: Implement web search tool
    └── server.py                   🚧 TODO: Implement FastAPI server

📋 Requirements

• Python: 3.11 or higher
• Dependencies: Listed in requirements.txt
• API Access: DIAL API key with appropriate permissions
• Network: EPAM VPN connection for internal API access
• Optional: Postman for API testing

🔧 Setup Instructions

1. Create virtual environment

   python -m venv .venv
   

2. Install Dependencies

   pip install -r requirements.txt
   

3. Environment Variables

DIAL_API_KEY=your_dial_api_key

Getting DIAL API Key:

1. Connect to EPAM VPN
2. Visit: https://support.epam.com/ess?id=sc_cat_item&table=sc_cat_item&sys_id=910603f1c3789e907509583bb001310c
3. Follow the instructions to obtain your API key

---

🚀 Task:

If the task in the main branch is hard for you, then switch to the with-detailed-description branch

Create MCP Server:

0. Run docker desctop with UMS
1. Open mcp_server and review mcp server structure:
   • in models persist implemented request and response models, details about request and response official documentation
   • in services/mcp_server.py you need to implement parts described in TODO sections
   • in tools you will find simple tools
   • lastly, in server.py provide implementations described in TODO sections
2. Run MCP server locally
3. Test it with Postman. Import mcp.postman_collection.json into postman. (init -> init-notification -> tools/list -> tools/call)
4. Open agent/app.py and run it locally with MCPClient and implement it
5. Test agent with queries below 👇
6. Provide implementations described in TODO sections for custom_mcp_client.py
7. Test again agent with queries below 👇

Check if Arkadiy Dobkin present as a user, if not then search info about him in the web and add him

---

🔍 MCP Protocol Details

JSON-RPC Structure

Request Format:

{
  "jsonrpc": "2.0",
  "id": "unique-request-id",
  "method": "method_name",
  "params": {
    "parameter": "value"
  }
}

Response Format:

{
  "jsonrpc": "2.0",
  "id": "matching-request-id",
  "result": {
    "data": "response_data"
  }
}

MCP Session Flow

1. Initialize: Client sends initialize request
2. Notification: Client sends notifications/initialized
3. Discovery: Client calls tools/list to get available tools
4. Operation: Client calls tools/call with specific tool and arguments
5. Shutdown: DELETE, {host}, Mcp-Session-Id: {Mcp-Session-Id}, shutdown is not covered in this practice, but it’s simple REST request

Headers

• Content-Type: application/json
• Accept: application/json, text/event-stream
• Mcp-Session-Id: Session identifier (after initialization)

🎯 Implementation Tips

Custom MCP Client Implementation

1. Error Handling: Always check for HTTP session initialization
2. Session Management: Store and reuse session IDs properly
3. SSE Parsing: Look for data: prefixed lines, ignore [DONE]
4. JSON-RPC Errors: Check for error field in responses
5. Content Extraction: Tool results are in result.content[0].text

Common Issues

• Missing Accept Header: Server requires both JSON and SSE accept types
• Session ID Missing: Most operations require a valid session ID
• Tool Arguments: Arguments must be properly formatted as per tool schema
• Async Context: Use proper async/await patterns for HTTP requests

📚 Additional Resources

• MCP Specification
• JSON-RPC 2.0 Specification
• MCP Error Codes
• OpenAI Function Calling

---