How to Create an MCP Server: Tutorial

Model Context Protocol (MCP) enables an AI assistant to connect with external systems like APIs or databases through a structured format. In this tutorial, we'll explore how to build an MCP server that acts as a bridge between an AI agent and a Todo Management API called TodoHub.

Our architecture consists of a User sending a request to an AI Agent (Kiro, Codex, Claude), which then calls an MCP tool provided by the TodoHub MCP Server. The server communicates with the REST API of TodoHub to fulfill the request.

The MCP server provides AI-friendly tools, such as "get_todo" and "create_todo", which simplify the API interaction for the AI agent. These tools translate the AI-friendly input parameters into the internal application parameters required by the REST API.

To set up the MCP server, we create an mcp.json configuration file that specifies the available tools, their commands, arguments, and necessary environment variables. For our example, we configure the "todohub" MCP server to use stdio communication and provide the required environment variables such as the TodoHub URL and API key.

When a user sends a request like "Show me todo 123", the AI agent understands the intent and identifies the required tool, "get_todo", with the input parameter "todo_id = 123". The MCP server receives this request and executes the equivalent GET request to the TodoHub REST API: GET https://todos.example.com/api/todos/123.

TodoHub responds with the requested todo data, which is then returned to the user by the MCP server. This end-to-end flow demonstrates how the MCP server abstracts the complexity of the underlying REST API, providing a simpler and more AI-friendly interface for the assistant to interact with.