# How to Create an MCP Server: Tutorial > Source: > Published: 2026-08-23 05:27:07+00:00 **Model Context Protocol (MCP)** allows an AI assistant such as Kiro, Codex, Claude, or another MCP-compatible agent to interact with external systems in a structured way. A useful mental model is: ``` php AI Agent --> MCP Client --> MCP Server --> External API / Database / Application ``` For this example, assume we have an internal Todo Management API called **TodoHub**. TodoHub provides REST APIs such as: ``` GET /todos/123 POST /todos PUT /todos/123 POST /todos/123/comments ``` We want an AI agent to understand requests such as: Show todo 123 or: Create a high-priority todo for fixing the login issue. Our MCP server acts as the bridge between the AI agent and the TodoHub API. Our architecture will look like this: ``` php User ----> "Show todo 123" AI Agent (Kiro / Codex / Claude) ----> MCP tool call TodoHub MCP Server ----> HTTP REST call ---> TodoHub API ``` The MCP server exposes tools such as: ``` get_todo create_todo update_todo add_comment ``` These are **MCP tools**. The AI does not need to know exactly how the underlying REST API works. It only needs to understand the tool and its input: ``` Tool: get_todo Input: todo_id ``` The MCP server handles the actual API communication. For example: ``` php AI ----> get_todo(todo_id=123) | |-----> MCP Server ---> GET /api/todos/123 ----> TodoHub ``` This distinction is important. You might wonder: Why don't we simply give the AI our REST API? The reason is that an MCP server provides the AI with a **cleaner, AI-friendly abstraction** over the underlying API. Your REST API might require something like: ``` POST /api/v2/workitems ``` with a request body: ``` { "subject": "...", "type_id": 7, "priority_id": 3, "workspace_id": 19, "creator": 758 } ``` However, exposing all these internal implementation details to the AI is unnecessary. Instead, the MCP tool could expose a much simpler interface: ``` create_todo( title, description, priority ) ``` The MCP server translates the AI-friendly parameters into the parameters required by the internal application. For example: ``` priority = "high" │ ▼ MCP Server │ ▼ priority_id = 3 ``` So the architecture becomes: ``` AI-friendly parameters │ ▼ MCP Server │ ▼ Internal application parameters │ ▼ REST API ``` This keeps implementation details away from the AI and gives the AI a simpler interface to work with. An MCP server can expose different tools for different operations. For our TodoHub example: | MCP Tool | Purpose | |---|---| `get_todo` | Retrieve a todo | `create_todo` | Create a new todo | `update_todo` | Update an existing todo | `add_comment` | Add a comment to a todo | For example: `get_todo` ``` Input: todo_id: integer ``` `create_todo` ``` Input: title: string description: string priority: string ``` The AI can then select the appropriate tool based on the user's request. The AI agent needs to know how to start and communicate with the MCP server. For example, we can create an `mcp.json` configuration file: ``` { "$schema": "https://agent-plugins.org/schemas/1.0.0/mcp.schema.json", "mcpServers": { "todohub": { "type": "stdio", "command": "uvx", "args": [ "--from", "mcp-todohub", "mcp-todohub" ], "env": { "TODOHUB_URL": "https://todos.example.com", "TODOHUB_API_KEY": "xxxxx" } } } } ``` The important parts are: ``` mcpServers │ └── todohub │ |___ SKILL.md ├── type ├── command ├── args └── env ``` The configuration tells the AI client: `todohub` is available.`stdio` .`uvx` is used to start the server.Let's follow one complete request. Show me todo 123. The AI determines that the user wants information about a todo. ``` Intent: Retrieve todo information Todo ID: 123 ``` The MCP server has advertised tools such as: ``` get_todo(todo_id: int) create_todo(...) update_todo(...) add_comment(...) ``` The AI chooses: ``` get_todo ``` with: ``` todo_id = 123 ``` Conceptually, the request looks like: ``` { "name": "get_todo", "arguments": { "todo_id": 123 } } ``` The MCP server receives the request and executes something equivalent to: ``` get_todo(123) ``` The MCP server then communicates with TodoHub: ``` GET https://todos.example.com/api/todos/123 ``` TodoHub returns: ``` { "id": 123, "title": "Payment timeout", "status": "In Progress" } ``` The MCP server sends the result back to the AI: ``` TodoHub ↓ MCP Server ↓ AI ``` The AI can now respond to the user: Task #123 is "Payment timeout"and is currentlyIn Progress. Putting everything together: ``` User │ │ "Show me todo 123" ▼ AI Agent │ │ Understands intent ▼ Selects MCP Tool │ │ get_todo(todo_id=123) ▼ MCP Server │ │ Translates tool input ▼ REST API │ │ GET /api/todos/123 ▼ TodoHub │ │ Returns JSON ▼ MCP Server │ │ Returns structured result ▼ AI Agent │ │ Generates natural-language response ▼ User ``` The key idea is: MCP provides a standardized bridge between an AI agent and external systems. The AI works with meaningful tools such as `get_todo` and `create_todo` , while the MCP server takes care of authentication, API calls, parameter translation, and other implementation details. That is the complete MCP cycle. `SKILL.md` Additionally, we can have a `SKILL.md` file under the MCP project directory structure mentioned above. This becomes particularly useful when the MCP tool needs **business context or parameter-building guidance that cannot be expressed cleanly through the tool schema alone**. `SKILL.md` vs MCP Server An important distinction is: | Component | Purpose | |---|---| MCP Tool Definition | Tells the AI what the tool does and what parameters it accepts. | `SKILL.md` | Provides additional instructions, context, rules, examples, and parameter-building guidance for the agent. | MCP Server Code | Validates and translates the parameters before making the actual REST API call. | A `SKILL.md` generally contains: ``` → Business context → How to construct parameters → Business rules → Examples ``` For example, the MCP tool might simply define: ``` create_task( title, description, priority ) ``` While `SKILL.md` can explain **how the AI should derive those parameters from the user's request**, including business rules and examples. `SKILL.md` Maintainable If `SKILL.md` becomes too large, we can split the content into multiple Markdown files and organize them under a `references` directory. For example: ``` taskhub-mcp/ ├── SKILL.md ├── server.py ├── tools/ │ ├── get_task.py │ ├── create_task.py │ ├── update_task.py │ └── add_comment.py └── references/ ├── task-creation.md ├── priority-rules.md └── business-rules.md ``` This keeps the main `SKILL.md` concise while allowing more detailed business context to be maintained separately. **Happy reading!**