What is MCP, and does your business need an MCP server?

MCP sounds like another acronym you'll be asked to pay for. Sometimes it is. But it's also one of the more useful ideas to come out of AI in the last couple of years, and it's worth ten minutes to understand what it does.

I build MCP servers for clients as part of my agentic systems work, so you'd expect me to tell you that you need one. Often you don't. Here's what MCP is, the situations where an MCP server genuinely earns its place, the situations where it's overkill, and the security questions to ask either way.

What is the Model Context Protocol?

The official site describes MCP as "an open-source standard for connecting AI applications to external systems". It gives a handy analogy: "Think of MCP like a USB-C port for AI applications."

The analogy works. Before USB-C, every device had its own cable. Before MCP, every AI tool needed its own custom connection to every system you wanted it to use. When Anthropic introduced MCP in November 2024, it described exactly that problem: "every new data source requires its own custom implementation, making truly connected systems difficult to scale."

MCP fixes the shape of the plug. If your system speaks MCP, any AI application that speaks MCP can connect to it. According to the official documentation, that includes AI assistants like Claude and ChatGPT and development tools like Visual Studio Code and Cursor, among many others.

The pieces, without the jargon

There are three parts worth knowing, as the MCP architecture overview lays them out:

• The host is the AI application your team uses, such as Claude Desktop or an AI coding tool.
• The client is the part inside that application that holds a connection to one server.
• The server is a program that gives the AI access to something: a database, a file store, your CRM, your internal tools.

A server can offer three kinds of thing:

• Tools, which are actions the AI can take, like running a database query or calling an API.
• Resources, which are information the AI can read, like file contents or database records.
• Prompts, which are reusable templates that help structure how the AI works with that system.

The documentation gives a simple example. A server for a database could offer tools for querying it, a resource containing the database's structure, and a prompt with examples of how to use the tools well.

When an MCP server is worth building

MCP pays off when you want one integration to work in more than one place.

• Your team uses several AI tools. If some staff use Claude, some use ChatGPT and your developers use Cursor, one MCP server can give all of them the same access to the same system. You build the connection once.
• Staff want to ask questions of internal systems. The official site gives the example of company chatbots connected to several databases at once, so staff can analyse data by simply asking. A salesperson asking their AI assistant "which of my accounts haven't ordered in 60 days?" and getting an answer from your real data is a good fit.
• You sell software and your customers use AI assistants. If your customers want to use your product from inside Claude or ChatGPT, an MCP server is the standard way to offer that.

In each case, the value comes from the "build once, use everywhere" part. That's the whole point of the protocol.

When you don't need an MCP server

If you're building one application that talks to one AI model, you probably don't need MCP at all.

Most AI model APIs let you give the model a list of tools directly, often called tool calling or function calling. Your code describes what the tool does, the model asks to use it, your code runs it. For a single product feature, that's simpler, and there's less to secure.

Not everyone is convinced MCP is needed even in bigger setups. A Hacker News thread titled "Why do we need MCP?" has people arguing that command-line tools and ordinary APIs work well enough for many jobs. Others in the same thread defend it as a common protocol every agent can use, whatever sits behind it. Both sides have a point. The deciding question is how many AI applications need to reach the same system.

A quick way to tell:

• One app, one model, your own code: use plain tool calling.
• Several AI tools or several teams needing the same access: consider an MCP server.
• Customers wanting to use your product from their own AI assistant: an MCP server is probably the right answer.

Is an MCP server secure?

It can be, but only if it's built carefully. An MCP server is a door into your systems, and the protocol's own security best practices are frank about how that door can be misused. In plain English, the main risks are:

• Accepting the wrong keys. The guidance is unambiguous: "MCP servers MUST NOT accept any tokens that were not explicitly issued for the MCP server." Passing someone's login token straight through to another system breaks the checks that would normally stop misuse, and makes it hard to trace who did what.
• Being tricked into acting for the wrong person. The "confused deputy" problem is when a server that connects to another service can be manipulated into granting an attacker access without the user properly agreeing.
• Being tricked into fetching the wrong address. Server-side request forgery is when an attacker gets the system to make requests to places it shouldn't, such as internal services or cloud credential endpoints.
• Running untrusted code locally. Servers installed on someone's own computer may have direct access to that machine, so a malicious one can do real damage.
• Asking for too much access. The guidance recommends starting with a minimal set of permissions and asking for more only when a specific action needs it. A stolen key with broad access gives an attacker far more to work with.

For a business owner, that boils down to a few questions to put to whoever builds it: what exactly can this server read and change, whose credentials does it use, is it read-only by default, and where are the logs? I go deeper on this in the post on giving AI your business data.

Who controls MCP now?

A fair worry with any standard is that one company owns it. On 9 December 2025, MCP joined the Agentic AI Foundation, a directed fund under the Linux Foundation. The foundation was co-founded by Anthropic, Block and OpenAI, with support from Google, Microsoft, AWS, Cloudflare and Bloomberg. At that point it had over 97 million monthly SDK downloads and 10,000 active servers.

The announcement frames the move as protecting "MCP's vendor-neutrality and long-term independence". For you, that means building on MCP doesn't tie you to one AI provider. If you switch from one assistant to another, your MCP server should keep working, as long as the new one supports MCP.

Before you commission one

Run through this first:

• How many different AI applications need to reach this system?
• Would plain tool calling inside one app do the job?
• What is the smallest set of actions the server needs to offer?
• Can it start read-only?
• Who will maintain it when the systems behind it change?

If you answered "several", "no" and "a clear, short list", an MCP server is probably worth it. If you'd like a second opinion on your setup, send me what you're building.