What Is Model Context Protocol (MCP)?

Model Context Protocol (MCP) is an open-source standard for connecting AI applications to outside systems. It gives an AI assistant a common way to use another product through a set of named tools, so the assistant can read that product’s data and, if you allow it, take actions in it. You connect an assistant to a product once, and you can then ask for answers or work in plain language.

How does MCP work?

MCP has three parts.

The client is the part of an AI application that holds the connection to a server. The application it belongs to, such as a chat assistant or a coding tool, is called the host. In everyday talk, people call the whole thing the assistant or the MCP client.

The server is a program that offers context to clients. It can run on your own machine or be run by a company, such as a software vendor. A product that has an MCP server lets any MCP client connect to it.

Tools are the functions a server offers. Each tool has a name, a description and a list of inputs. The assistant asks the server which tools exist, picks one when your request needs it, and the server runs it and sends the result back. A tool can only read, or it can change something. MCP servers can also offer resources (data to read) and prompts (reusable templates).

“Connecting” an assistant to a product means adding that product’s MCP server to the assistant. For a server run by a vendor, you usually sign in to the product and approve the connection. From then on, the assistant can use the tools you allowed. The standard recommends OAuth for the sign-in step on remote servers.

Why does MCP matter for security awareness training?

Most security awareness training lives in its own console. You open it to check who is overdue, pull click rates for a report or send a few reminders, then go back to the tools you work in all day. That is one more dashboard for a small IT team to remember.

With MCP, an admin can ask their own assistant instead. “Which departments had the most phishing clicks last quarter?” or “Who is overdue on the password workshop?” goes to the training product’s server, and the answer comes back where the admin is already working. For an MSP, the same request can cover every client organization in one pass. This matters for training programs that depend on follow-up, because the follow-up gets easier to do.

What should you ask before connecting an MCP server?

An MCP server is a new way into your data, so treat it like any other integration. Ask these six questions of any server before you connect it.

What can it read? List the data the tools can return. Names, email addresses and individual results are personal data, and the assistant will see them.

What can it change? Find out whether the connection is read-only or can also create, edit and delete. Choose the narrowest access that does the job.

How does sign-in work? Prefer signing in with your own account and approving the connection to pasting a long-lived key into a config file. The connection should act as you and never see or do more than you can.

Are actions confirmed? A good server shows what will happen before it happens, such as who will be reached and what will be sent, and waits for you to confirm. Confirmation matters most for deletes and anything that sends messages.

Is there an audit trail? Every action should be recorded with who did it, which assistant did it and what changed, so you can review it later.

Where does the data go? Whatever a tool returns goes to the assistant you chose, and so to the company behind it. Check your agreement with that company. Also remember that text inside the data can try to steer the assistant, which is called prompt injection.

You also want a way out: disconnect one assistant, or switch the server off for the whole account.

How does Kinds use MCP?

Kinds has an MCP server for its security awareness training. Any MCP client can connect, including an app you build. Sign-in is OAuth with your usual Kinds login, so no API keys are needed, and you choose Read or Read, create and delete for each connection. It uses the same set of more than 50 tools as Clay, the assistant inside the Kinds app. You confirm before anything runs, each action goes into an audit trail, and the server cannot launch phishing simulations or edit templates. Read the full details in Kinds now speaks MCP, connect an assistant with the MCP setup guide, or see how other vendors’ servers differ in Security awareness training MCP servers compared.

FAQs

What is MCP in simple terms?

MCP is a shared way for AI assistants to talk to other products. Its documentation compares it to a USB-C port for AI applications. A product that offers an MCP server can be used by any assistant that speaks MCP, without a custom integration for each one.

Who created MCP?

Anthropic announced MCP on November 25, 2024, and described it as an open standard for connecting AI assistants to the systems where data lives. Anthropic’s announcement says it was created at Anthropic by David Soria Parra and Justin Spahr-Summers.

Is an MCP server the same as an API?

No. An API is a general way for software to talk to software, and you usually call it from code. An MCP server offers a described set of tools that an AI assistant can find and use on its own. The two can sit side by side in one product.

Is it safe to connect an assistant to an MCP server?

It can be, if the server limits what the assistant can reach. Check what the server can read and change, how you sign in, whether actions are confirmed, whether there is an audit trail and where the data goes. Start with read-only access and widen it only if you need to.

Does the assistant see my data?

Yes. Whatever the tools return is passed to the assistant you connected, and so to the AI company behind it, under your own agreement with that company. It does not stay inside the product that holds it.

Do I need to write code to use an MCP server?

No. For a vendor-run server, you add it in your assistant, sign in and approve the connection. You only write code if you want to build your own MCP client or server.