What is MCP (Model Context Protocol)?

Host, client and server

MCP has three roles. The host is the AI application you work in, such as Claude or ChatGPT. The server is the program that opens up a system: your CMS, your CRM or a database. The client sits inside the host and keeps the connection to one server open. Work with three servers, and the host has three clients.

A server offers three kinds of things: tools (actions the assistant can take), resources (data it can read) and prompts (ready-made templates for a task). The assistant first asks what’s available, then picks what it needs. A server can run on your own computer or remotely, for instance at the vendor of the system.

MCP, an API and an AI agent

MCP doesn’t replace an API. An MCP server usually calls the API of the system it opens up. The difference is who it talks to. An API integration is written for software that knows in advance what to do. An MCP server describes its actions in a way that lets a language model find and choose them itself.

An AI agent is something else again: the system that gets a goal and takes steps on its own. MCP is one of the ways such an agent reaches your systems. The agent decides what it wants to do. The MCP server decides what it can do.

Why it matters in practice

MCP defines how an assistant and a server talk to each other, not what’s sensible. The specification asks that users consent before a tool is called, but also says the protocol can’t enforce that itself. That’s up to whoever builds and configures the server. Three choices make the difference:

• Only the actions you need: if your team should only prepare pages, don’t build an action to publish or delete.
• Minimal rights: in the underlying system, the server only gets the permissions its actions require.
• Draft with preview: whatever the assistant creates or changes is saved as a draft first. A person reviews it and publishes.

Watch out for prompt injection, too. If the assistant reads outside text, such as an email or a web page, instructions can be hidden in it. The more actions a server offers, the more damage such an instruction can do. And don’t install servers from makers you don’t know: according to the specification, a tool’s description can’t be trusted unless the server comes from a trusted source.

How we use it

With a website, we deliver a dedicated MCP server, so the team can create and edit pages through Claude or ChatGPT using the building blocks of their own design system. Changes are saved as a draft with a preview first.

At Sprints & Sneakers, the team manages the website this way through Claude: pages, blogs, case studies and events, built from the blocks of their design system. Everything is saved as a draft with a preview first, and publishing is a separate step. How it fits together is covered in managing your CMS with your own MCP server.