What is an MCP server?

An MCP server sits between an AI application and a system the business already runs. It tells the application what the system can do and carries out the requests that come back. Common examples are servers for a file system, a database, a calendar, GitHub or Slack.

Each server offers three kinds of building block. Tools are functions the model can call, such as writing to a database, calling an API or changing a file. Resources give read-only context, such as the contents of a file or a database schema. Prompts are ready-made instructions for a specific task, and the user picks them.

The Model Context Protocol is the shared rulebook behind every MCP server. Anthropic introduced it in November 2024 and released it as open source. Its documentation compares it to a USB-C port: one standard way to connect AI applications to outside systems.

Before MCP, each new data source needed its own custom integration, which made connected systems hard to scale. With MCP, the owner of a system writes one server. Claude, ChatGPT, Visual Studio Code, Cursor and other applications that support the protocol can then reach it in the same way.

In December 2025 Anthropic donated MCP to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded by Anthropic, Block and OpenAI. Anthropic said the governance model would remain unchanged.

There are three parties. The host is the AI application a person works in, such as Claude Desktop or Visual Studio Code. The client is the part of the host that keeps a connection open to one server. The server is the program that supplies context. A host opens a separate client for every server it connects to.

When the connection starts, the client asks the server what it offers. The server lists its tools with a name, a description and the inputs each one expects. During a conversation the model picks a suitable tool, the host sends the call to the right server, and the result is added to the conversation.

Messages use JSON-RPC 2.0. A server on the same machine talks over standard input and output (stdio). A remote server is hosted on the internet and uses a transport called Streamable HTTP, and the protocol recommends OAuth for obtaining access tokens. In Claude, a remote MCP server shows up as a connector.

An API is the set of rules one piece of software uses to reach the data and functions of another, and every service has its own. Connecting an AI application to an API means writing code for that specific API. Connecting a second AI application to the same system means doing that work again.

An MCP server can sit in front of an existing API. It describes the API's functions as tools, and when a tool is called the API still does the work behind it. What changes is the description: any application that supports MCP can read the tool list and call the tools the same way.

So MCP does not replace the API. When one AI application needs one system, a direct API integration may be enough. When the same system should be reachable from several AI applications, or an agent needs many systems, MCP reduces the work of building and maintaining the connections.

The first point is permission. The MCP specification recommends that a person always stays involved, with the ability to refuse a tool call. The application should show which tools the model can see, make each call visible and ask for confirmation before sensitive operations. In Claude, every tool must declare whether it only reads data and whether it can change or delete it.

The second point is sign-in. On a remote server each user signs in with their own account, and Claude's documentation states that their permissions mirror their access on the external service. A server must accept only access tokens issued to it. Scopes start as narrow as possible and widen only when a task needs more.

The third point is where the server comes from. A local server runs with the same privileges as the application that starts it, so one from an untrusted source can reach files on the machine. Every tool call should be logged, and for GDPR purposes it is written down which personal data goes where.

ALTAI Digital adds the model as a layer on top of existing systems and limits it to the sources it is given. Sending, paying and deleting each keep a separate human approval step.