MCP (Model Context Protocol) is a standard protocol that was created by Anthropic to allow AI models to connect with different sources of information and services through one unified interface. Rather than having to create customized integrations for each of the tools that you want to use, developers can create a single, standard way to connect to these various tools using the MCP protocol.
The protocol uses a client-server architecture. An AI-like assistant or development platform will provide a client to use with the MCP, and will connect that client to an MCP server which will expose the following three primitives: Tools (the executable functions that can be called by the AI), Resources (the data that can be accessed by the AI to provide context for the AI), and Prompts (reusable templates which structure the workflow). All communication over MCP is done using JSON-RPC 2.0, which provides a lightweight and language-agnostic interface to use for communication between the client and server.
MCP was originally released in approximately November 2024, and the Linux Foundation has now taken over the management of this open-source protocol. The MCP protocol has been adopted across the AI community and is considered to be the standard for connecting LLMs with the real world.