MCP (Model Context Protocol)

MCP stands for Model Context Protocol. It is an open standard that defines how an AI application talks to other software. Think of Claude, ChatGPT, Gemini or an AI agentAn AI application that doesn't just answer, but takes steps itself to complete a task. It chooses the tools it needs, calls them, looks at the result and decides what the next step is. you have built for yourself. Through MCP, such an application first asks a system which tools it offers, and then calls one of those tools with the right data. Technically, those two steps are called tools/list and tools/call.

MCP and APIs are not rivals

People reading about MCP often wonder whether it replaces the API. It does not. An MCP server usually sits on top of your existing API. The API remains the place where your system offers its data and functions, and the MCP server translates them into a form an AI model understands. The chain looks like this: the AI agent talks to the MCP server, the MCP server calls your API, and your API works with the data in your systems.

The difference lies in who asks the question. With a regular API, a developer writes out every call in advance. With MCP, the AI model decides in the moment which tool it needs, based on the description the server provides for each tool.

What to look out for when you have an MCP server built

Is your software partner building an MCP server for your systems, or does one come with a package you already use? These four rules of thumb help you ask the right questions.

1. Keep authentication and business logic in your API. The MCP server is a thin layer that translates. Who may see what, and which business rules apply, such as discounts or approvals, stays handled where it already is today.
2. Make one tool per task, not one per endpoint (each separate entry point of your API). A tool that looks up a customer's outstanding invoices is far more useful to an AI model than ten separate technical calls it has to string together itself.
3. Expose less. Every tool the server offers takes up room in the model's working memory (the context) and therefore costs tokens. The more tools, the more expensive each question becomes and the greater the chance that the model picks the wrong one.
4. Label what reads and what writes. For each tool you can indicate whether it only reads data (readOnlyHint) or can delete or overwrite something (destructiveHint). That way the AI application knows when it should ask for your confirmation first.