EN

AI assistants and MCPExplainer

What is the Model Context Protocol (MCP)? A plain explanation

Dave stands in the snow in a thinking pose; on the left, a lantern, a cube of ice and a glass sphere each send out a thread of light, and the threads meet in one small glowing connector, from which a single thread leads on.

The Model Context Protocol (MCP) is an open standard that lets AI applications such as Claude or ChatGPT connect to files, databases and other services in a common language. Anthropic announced it on November 25, 2024; since December 2025, it has been developed at the Agentic AI Foundation, under the Linux Foundation. Its own documentation compares MCP to a USB-C port for AI applications: a connector is written once and works in every application that supports it.

What problem does MCP solve?

It removes the need for every AI application to build a separate integration for every data source. A source is written once as an MCP server, and every application that supports MCP can use it.

On its own, a language model knows only what it was trained on and what you type into the chat. For it to look at your calendar, your company's documents or your health data, it needs a bridge. Before MCP, these bridges were built one at a time: an integration written for one assistant didn't work in another. Ten assistants and ten data sources meant a hundred separate connections. With a shared protocol, each side does its part once: the data source writes a server, and the assistant writes a client.

How does MCP work?

There are three roles: the AI application you use (the host), the client it opens for each connection, and the program that offers the data or the tool (the server). The messages between them use the JSON-RPC 2.0 format.

Say you add two MCP servers to a desktop assistant: one for the files on your computer, the other for a calendar service. The assistant opens a separate connection to each server and first asks it, “What do you offer?” The server lists what it can do, with names, descriptions and the inputs it expects. When you ask a question, the model picks the right item from this list, the application passes the request to the relevant server, the response is added to the chat, and the model writes its answer based on it.

The protocol is updated in dated versions; the official site says which version of the specification is current.

What can an MCP server offer?

Three basic things: tools, resources and prompts.

  • Tools are functions the model can call: running a query on a database, getting information from a weather service, reading a file.
  • Resources are data that can be read as context: the contents of a file, the details of a record.
  • Prompts are reusable question templates, prepared for a particular task.

A server doesn't have to offer all of these. A server that offers only read tools also makes sure the model can't change anything.

What's the difference between a local and a remote server?

A local server runs on your computer and talks to the application directly; a remote server sits at an address on the internet and is reached over HTTP.

MCP defines two transports. Standard input and output (stdio) is for programs on the same machine to talk to each other. Streamable HTTP is for connecting to a server somewhere else, and it lets one service be open to many users at once. When an assistant needs to reach the data of an app on your phone, it's a remote server that does the job: the assistant can't reach your phone directly, but it can reach a server where the data is kept safely.

How do you give permission for a connection?

For remote servers, MCP recommends using OAuth for access. In practice, this means that when you add a connector in your assistant, you're sent to the service's own page, where you sign in and approve access.

It's similar to the flow you see when you tap “Sign in with Google” in an app: your password isn't given to the assistant; the assistant only gets a key for the access you approved. You can revoke this key on the service's side and remove the connector from your assistant. Some assistants may need a paid plan for you to add your own connector; this varies from one assistant to another.

Which applications support MCP?

MCP's official documentation lists assistants such as Claude and ChatGPT, as well as developer tools such as Visual Studio Code and Cursor, among the applications that support it. In its December 2025 announcement, the Linux Foundation said the protocol has been integrated into platforms such as Claude, Microsoft Copilot, Gemini and ChatGPT. Support isn't equally deep in every product; it's best to check an assistant's own documentation to see which features it offers.

Is MCP safe?

The protocol provides tools for setting up secure connections, but it doesn't guarantee safety on its own. What really matters is what the server you connect can do, and who runs it.

  • Add connectors only from services you trust, where it's clear who runs them.
  • Don't grant broader permissions than needed. If read-only access is enough, be wary of a connector that asks to write.
  • Content coming from a tool can carry hidden instructions aimed at the model. That's why it's a good sign when your assistant asks for confirmation before it takes an action.
  • Remove connections you don't use.

What does MCP mean for health data?

With MCP, you don't have to upload your health data to a chat as a file; the assistant asks for the part it needs to answer your question, when it needs it. Dave Connect does exactly this: with your consent, it uploads your Apple Health or Health Connect data to Dave and lets the assistant you connect read that data, read-only, through MCP. The data your assistant asks for goes to the company that runs it and is handled under that company's terms; you can remove the connection whenever you like. For more on this, see our article on connecting your health data to an assistant.

Sources

Claude, ChatGPT, Gemini, Microsoft Copilot, Visual Studio Code and Cursor are trademarks of their owners. Dave is not affiliated with these companies and is not endorsed by them.