What Is MCP?

MCP meaning: what the name stands for

The model is the AI model, such as Claude or GPT. Context is the information and actions that model needs to do useful work. A protocol is a set of rules both sides follow.

The MCP docs reach for hardware to explain it: "Think of MCP like a USB-C port for AI applications" (MCP docs). USB-C is one connector shape that works across many devices. MCP aims to be that for AI apps and tools.

What problem does MCP solve?

An AI model cannot see this morning's orders or your CRM unless something connects it. Before MCP, each AI app needed its own connector for each tool. Anthropic's launch post described the result: every new data source needed its own custom integration, which made connected systems hard to scale (Anthropic).

Five AI apps and twenty tools can mean up to 100 custom integrations. With a shared protocol it is five apps that speak MCP and twenty MCP servers, and each side builds once.

The spec says MCP takes inspiration from the Language Server Protocol, which did the same job for code editors and programming languages (spec).

How MCP works, briefly

There are three roles (architecture overview). The host is the AI app you use. For each tool it connects to, the host creates a client. At the other end is an MCP server, a program that offers tools and data from one system.

When you ask a question, the app already holds each server's tool list. The model picks a tool. The app calls it, the server does the work, usually through the product's own API, and the result comes back for the model to use.

Most people who use MCP through an AI app never need more detail than that. If you do:

• What is an MCP server? covers servers, their types and real examples.
• Model Context Protocol: how it works is the technical reference for messages, transports, sign-in and spec versions.
• MCP vs API explains how MCP sits on top of ordinary APIs.

A short history of MCP

The July 2026 revision is the current one. It removed the opening handshake and sessions, so each request stands alone and any copy of a remote server can answer it behind an ordinary load balancer. It also deprecated three older features, roots, sampling and logging, with at least twelve months before removal. The technical guide goes through the changes.

Who runs MCP now?

MCP is a Linux Foundation project. The Agentic AI Foundation launched on 9 December 2025 with three founding projects: MCP from Anthropic, goose from Block and AGENTS.md from OpenAI. Its platinum members are Amazon Web Services, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI (Linux Foundation).

The foundation does not set the technical direction. The MCP project said its existing governance carries on and that decisions stay with the maintainers (MCP blog).

Who uses MCP?

Anthropic's December 2025 post lists ChatGPT, Cursor, Gemini, Microsoft Copilot and Visual Studio Code as products that had adopted MCP, alongside its own Claude apps (Anthropic).

MCP in Claude

On claude.ai and Claude Desktop, you add a remote MCP server as a custom connector under Customize > Connectors. Custom connectors are available on Free, Pro, Max, Team and Enterprise plans, and Free is limited to one. On Team and Enterprise, an owner adds the connector for the organization first (Claude help center).

Claude Desktop can also start local servers from a config file (MCP docs). Claude Code adds servers with the claude mcp add command (Claude Code docs).

MCP in ChatGPT

ChatGPT's developer mode is a full MCP client for remote servers, covering both read and write tools. OpenAI lists it for Pro, Plus, Business, Enterprise and Education accounts on the web (OpenAI). OpenAI's API accepts remote MCP servers too (OpenAI MCP guide).

Adoption numbers

Two dated figures, one from Anthropic and one from the MCP project:

• December 2025: more than 10,000 active public MCP servers, and over 97 million monthly SDK downloads across Python and TypeScript (Anthropic).
• July 2026: close to half a billion SDK downloads a month across the Tier 1 SDKs, with the TypeScript and Python SDKs each past 1 billion total downloads (MCP blog).

We found no official server count newer than December 2025.

What people use MCP for

A store owner asks, "Which SKUs have fewer than 20 units left and sold more than 50 last month?" A marketer asks for ad spend against revenue by campaign for September. A sales lead asks which open deals have had no activity in 14 days.

Each is a question in plain words, answered from live data in the system that holds it. The use-case guides go further: ecommerce, marketing and CRM.

The MCP docs give examples outside business tools too, such as an agent that works with your Google Calendar and Notion, and Claude Code building a web app from a Figma design (MCP docs).

MCP gives a model live data. The model can still pick the wrong tool or misread a field, so check any number that matters. For a fixed job that runs on a schedule, a script that calls the API directly is the better fit, as MCP vs API explains.

What MCP is not

"MCP is an AI model"

MCP is a protocol. It carries requests between an AI app and tools. It does no reasoning of its own.

"MCP is an agent framework"

MCP does not plan tasks or run loops. The MCP docs say it covers only the exchange of context and leaves how an app uses its model to the app (architecture overview). Agent frameworks and AI apps do the planning and use MCP to reach tools.

"MCP is secure by default"

The spec sets principles, such as asking the user before a tool runs. It also says the protocol cannot enforce them, so the AI app and the server have to (spec).

Trusting the server's author is not the whole story either. OpenAI's guide warns that content reached through an MCP server can carry a prompt injection even when the developer is trustworthy (OpenAI MCP guide). MCP security best practices covers what to lock down.

How to get started with MCP

If you use AI at work, pick an app that supports MCP and decide which tools it should reach. Add a server for each, from the vendor or from a hosted provider. Try a read before anything that changes data. The step-by-step guides are Claude with your business tools and ChatGPT with your business tools.

If you are a developer, start with the architecture overview, then pick an official SDK and build a small server.

With PopMCP, you connect a provider account once, add https://app.popmcp.com/mcp to your AI app, sign in and tick the connectors that app may reach. On the Free plan every tool is read-only, so the model cannot change anything in your accounts while you try it.

PopMCP covers business tools from its MCP server catalog. It is not a workflow builder, and it does not run jobs on a schedule.