What is MCP? The Model Context Protocol explained

Looking for the short version without technical terms? Read What is MCP? in plain words.

What is MCP? The Model Context Protocol is an open standard that lets an AI application call tools and read data in other systems through one common interface. Anthropic published it in November 2024. Instead of a custom integration per app, a system exposes an MCP server once and every MCP client can use it.

What is MCP in plain terms?

The short answer to what is MCP: a contract between two programs, one that wants to use a capability (the AI application) and one that offers it (the server). The protocol defines how they find each other’s capabilities, how a request is shaped and how a result comes back. The messages are JSON-RPC 2.0. The model never talks to your database or inbox directly; it asks the client, the client asks the server, and the server does the work.

What is MCP made of?

An MCP setup has three roles. Knowing them makes the rest of the documentation easier to follow.

• Host: the application you work in, for example Claude Desktop or Claude Code. The host runs the model and decides when a tool is needed.
• Client: the connector inside the host. A host runs one client per server it is connected to.
• Server: the program that exposes capabilities. It knows how to reach the real system (an API, a database, a folder of files) and translates MCP requests into calls to that system.

What can an MCP server expose?

The specification defines three kinds of capabilities a server can offer:

• Tools: functions the model can call, such as “list conversations” or “get order”. Each tool has a name, a description and a JSON Schema for its input.
• Resources: data the application can read, addressed by a URI.
• Prompts: reusable prompt templates a user can pick.

Tools are the part most people mean when they ask what MCP is for. The model reads the tool descriptions, picks one, fills in the arguments and gets structured output back.

How does a client call a tool?

A client calls a tool in a fixed sequence. First it sends an initialize request; client and server agree on a protocol version and on which capabilities each side supports. Then the client sends tools/list and receives the available tools with their input schemas. When the model decides to use one, the client sends tools/call with the tool name and arguments. The server runs the call against the real system and returns the result as content the model can read.

Which transports does MCP use?

MCP defines two standard transports. stdio runs the server as a local child process of the host and exchanges messages over standard input and output. Streamable HTTP runs the server as a normal HTTP service, so it can live in a container or on another machine. Streamable HTTP replaced the older HTTP+SSE transport in the 2025-03-26 version of the specification. Many clients still support the SSE variant, which is why servers often offer both.

What is MCP used for with your own systems?

It gives your systems one gateway that every compatible AI application can use. Without a shared protocol, every AI application needs its own plugin for every system. With MCP, a system gets one server and any compatible client can use it. For a team that means one place to decide which systems Claude can reach, with which login details and with which permissions. Answering “what is MCP” for your organisation comes down to that: one controlled gateway instead of a pile of one-off integrations.

Where does MCP Self-hosted fit in?

For an AddoCore customer, what is MCP has a concrete answer: MCP Self-hosted, an MCP server that you run yourself in Docker. It exposes tools over Streamable HTTP at /mcp and keeps legacy SSE at /sse for older clients. Claude Desktop and Claude Code connect to it, as does any MCP client that supports Streamable HTTP or SSE. Each tool add-on (Missive, WooCommerce, Database, Brain, WordPress, Lightspeed eCom) is a set of tools inside the same server image.

Three rules shape what a client sees after tools/list:

• Only tools your license is entitled to are registered. A tool you have not bought does not appear at all.
• Every client authenticates with the token of its own seat, and each seat has its own subset of tools.
• Write tools only appear when that seat’s mode for the tool has Can make changes on. Read-only is the default for every tool except Brain.

MCP Self-hosted also ships one built-in tool, mcp_status, that is always available. Ask Claude to run it and you get the license state, your seat’s tools and modes, and version information. The MCP Self-hosted add-on model explains how the license and add-ons fit together.

What is MCP not?

MCP is not a model and it does not store your data. It does not give the AI more rights than the server grants: if a server only registers read tools, the model cannot write, however the prompt is phrased. It is also not a hosted service by definition. Where the server runs is your choice; with MCP Self-hosted it runs on your own infrastructure, next to the systems it connects to.

Where do you go from here?

Now that the question “what is MCP” is answered, the next step is a running server. Follow Install MCP Self-hosted to start MCP Self-hosted with ./mcp-update.sh, then connect Claude to the server and pick a tool from the setup guides.