What is MCP? The Model Context Protocol explained

MCP (Model Context Protocol) is an open-source standard for connecting AI applications to external systems: data sources such as files and databases, tools such as search engines, and workflows such as prepared prompts. The official site compares it to a USB-C port for AI apps: one standard plug instead of a custom integration for every tool (modelcontextprotocol.io, read Nov 6, 2026).

How MCP works

• Host and client: the AI application (for example Claude, ChatGPT, VS Code or Cursor, which the MCP site lists as supporting it) runs an MCP client.
• Server: a program that exposes capabilities to that client. It can run locally or as a remote server over HTTP.
• What a server offers: tools the model can call, resources it can read and prompts it can use. The model decides when to call a tool; the host decides what it may access.

The specification is versioned by date; the documentation listed version 2026-07-28 as the latest on Nov 6, 2026.

What MCP is used for

The MCP site’s own examples include agents reading a calendar and notes, coding assistants building an app from a design file, and company chatbots querying several databases. In our directory, vendors ship MCP servers for their APIs too: Nansen lets Claude or Cursor connect to live onchain data over MCP, Fathom lists a public API and MCP for meeting data, and LLM gateways such as LiteLLM and Requesty list MCP gateways (see LLM routers and AI gateways).

MCP vs an API

MCP does not replace APIs; it wraps them in a form an AI client can discover and call. An MCP server usually calls the vendor’s normal REST API underneath, with the same pricing and rate limits. What changes is discovery: the client asks the server which tools exist and what arguments they take, so the model can use a new tool without custom code.

Paying for MCP tools

Most MCP servers use the vendor’s API key and subscription. A growing set charge per call in USDC with x402 instead, so an agent with a wallet can pay without an account: see paid MCP servers that accept USDC and how AI agents pay for things.

Security basics

• Install servers only from sources you trust; a server can run code and read data.
• Give each server the narrowest keys and scopes it needs, and cap spend on paid tools.
• Treat tool descriptions and results as untrusted input; see audit an agent’s MCP tool list and prompt injection scanning.

Sources: modelcontextprotocol.io (What is MCP?), nansen.ai/api, fathom.ai/pricing, docs.litellm.ai and requesty.ai/pricing, read Nov 6, 2026.