Definition of the MCP protocol
The MCP protocol, for Model Context Protocol, is an open standard that describes how an AI application connects to external systems: data sources, tools, prompt templates. This definition is the one given by the protocol’s official website (modelcontextprotocol.io), consulted on 12 November 2026. The website compares it to a USB-C port, a common socket for a variety of devices.
Anthropic published the protocol as open source on 25 November 2024. Its announcement credits David Soria Parra and Justin Spahr-Summers with its creation. At the date of consultation, the website’s governance page states that the project is constituted as “Model Context Protocol a Series of LF Projects, LLC”. Versions of the specification are identified by a date: the current version displayed that day is 2026-07-28.
Host, client, server: who talks to whom
The architecture, of the client-server type, distinguishes three roles.
- The host is the AI application. Among compatible applications, the website cites the Claude and ChatGPT assistants, and the development tools Visual Studio Code and Cursor.
- The client is the component that the host creates for each server it connects to, and which maintains that connection.
- The server is the program that provides the context, on the host’s machine (local server) or remotely, over HTTP.
Exchanges follow the JSON-RPC 2.0 format. For remote servers, the documentation recommends OAuth.
What a server exposes: tools, resources, prompts
A server offers three primitives:
- tools, functions the model can call in order to act: query a database, call an application programming interface;
- resources, data provided as context: the content of a file, records;
- prompts, reusable prompt templates.
On the client side, elicitation allows a server to ask the user for clarification or confirmation. The list evolves: version 2026-07-28 deprecates the primitive known as sampling.
MCP protocol, API, agent, RAG: not to be confused
The protocol does not replace an application’s programming interface (API): an MCP server can rely on it. Having clean interfaces remains a matter of information system architecture.
It is not an AI agent: the agent decides and chains actions, while the protocol gives it what it can read and call. The documentation specifies that MCP does not dictate how the application uses the language model.
It is not to be confused with RAG, an architecture that searches for passages before writing. Such a search can be exposed by an MCP server.
Permissions granted and content returned: two points of vigilance
According to its specification, the protocol cannot enforce its security principles by itself. Two subjects fall to those who implement it.
Permissions, first. The security documentation points out that a local server runs with the privileges of the client application. For remote servers, it advocates minimal permissions, extended as needed.
Content, second. What a tool returns enters the model’s context and may carry instructions intended to hijack it. The specification recommends that applications validate a tool’s results before passing them to the model, and requires them to treat as untrusted any tool annotations that do not come from a trusted server.
A use case: internal IT support
Imagine the IT support desk of a transport company. Its internal assistant, the host, is connected to two MCP servers. The first provides the workstation inventory in read-only mode, as resources. The second exposes two tools: search for a ticket, create one.
A technician asks for a workstation’s history. The assistant reads the inventory, then offers to create a ticket. The application displays the call’s parameters and waits for the technician’s approval. The account used by the second server can neither close nor delete a ticket.
Choosing the tools opened to an agent, its permissions and its human oversight is covered on our page devoted to AI agents.