What Is MCP? An Introduction to the 'USB-C Port' for AI Integrations

Phrases like “MCP support” and “MCP server now available” keep showing up in descriptions of AI tools and services. Claude, ChatGPT, Visual Studio Code, Cursor — the major AI applications and developer tools all support it¹. So what exactly is MCP (Model Context Protocol)?

This article draws on MCP’s official documentation and Anthropic’s announcement to explain why MCP was created and how it works, in terms accessible to readers who are not AI specialists.

What is MCP? A “USB-C port” for AI applications

According to the official site, MCP is “an open-source standard for connecting AI applications to external systems”¹. Using MCP, AI applications like Claude or ChatGPT can connect to data sources (such as local files and databases), tools (such as search engines and calculators), and workflows (such as specialized prompts), enabling them to access key information and perform tasks¹.

The official site describes it as being “like a USB-C port for AI applications”¹. Just as USB-C standardized how electronic devices connect, MCP standardizes how AI applications connect to external systems. The way any USB-C charger works with any USB-C device, anything that speaks MCP can be combined with an AI application — no custom integration work required.

Why it exists: the fragmentation problem

MCP is a protocol that Anthropic announced and open-sourced on November 25, 2024³. The problem Anthropic described at launch: connecting an AI assistant to external data and tools required a custom implementation for every single data source. With connection methods fragmented across services, genuinely connected systems were hard to scale. MCP’s answer was to let developers build against one common standard protocol instead of maintaining a separate connector for every destination³.

Consider ten AI applications and ten external services you want them to reach: with bespoke integrations, that is up to one hundred connections to build. With a shared protocol, each side implements the standard once. Block and Apollo integrated MCP early on, and developer-tool companies including Zed, Replit, Codeium, and Sourcegraph worked on support³. OpenAI has also reportedly adopted MCP⁵, and with ChatGPT among the major applications supporting it¹, MCP is becoming the de facto common standard.

How it works: hosts, clients, and servers

MCP follows a client-server architecture with three participants².

The MCP host is the AI application itself — the thing the user actually touches, such as Claude Code or Claude Desktop. The host establishes connections to one or more MCP servers and coordinates multiple MCP clients².

The MCP client is a component inside the host that maintains a dedicated one-to-one connection with a server. The host creates one client per server, and the client obtains context from its server on the host’s behalf².

The MCP server is a program that provides context and capabilities to the AI². For example, Visual Studio Code acts as an MCP host: when it connects to the MCP server for the error-monitoring service Sentry, it instantiates one client; when it then connects to a local filesystem server, it instantiates another².

Despite the name, an MCP server need not be heavyweight infrastructure — it can be a small program running on your own machine. Both local servers (communicating over standard input/output) and remote servers (communicating over HTTP) are supported². Under the hood, the messages follow JSON-RPC 2.0, a well-established format².

The three things a server can offer: tools, resources, and prompts

What an MCP server can provide to an AI application falls into three categories, which the official documentation calls primitives².

First, Tools: executable functions the AI can invoke to perform actions — file operations, API calls, database queries, and the like². Tools give the AI “hands.”

Second, Resources: data sources that provide contextual information to the AI, such as file contents, database records, and API responses². Resources hand the AI its “reference material.”

Third, Prompts: reusable templates that structure interactions with the language model, such as system prompts and few-shot examples².

To borrow the official documentation’s example, an MCP server for a database could expose tools for running queries, a resource containing the database schema, and a prompt with worked examples of how to use the tools². On connection, the AI application first lists what a server offers before using it, and when the server’s lineup changes, it can notify clients so they refresh the list² — a design that tolerates a dynamically shifting set of capabilities.

What MCP makes possible

The official site offers concrete examples: agents accessing your Google Calendar and Notion to act as a more personalized assistant, Claude Code generating an entire web app from a Figma design, enterprise chatbots connecting to multiple databases across an organization so users can analyze data through chat, and AI models creating 3D designs in Blender and sending them to a 3D printer¹.

The common thread: AI stops being something that merely answers from its trained knowledge, and becomes something that reads the data that exists right now and operates external tools to get work done. For AI agents that carry out tasks autonomously, MCP is the standardized wiring to the outside world.

We have covered specific MCP servers on this site before, including Figma Context MCP, which feeds design data to AI coding tools. For a tour of which servers are popular, see our MCP tools ranking — with the mechanics under your belt, browsing the actual ecosystem makes its breadth tangible.