In plain language
AI applications such as Claude or Codex can be capable without automatically having access to your GitHub repositories, Notion workspace, database or browser.
MCP defines a common way for an AI application to connect to those outside capabilities. Its architecture describes roles such as hosts, clients and servers. A tool provider can offer an MCP server, and a compatible AI application can use the protocol to communicate with it instead of every product inventing a completely different integration. (Model Context Protocol)
MCP is the connection standard. It is not GitHub, a browser or a database by itself.
An everyday example
An analogy
USB is not a mouse or a hard drive. It is a standard that helps devices connect.
In the same way, MCP is not the external service. It defines how an AI application and that service can connect. This is an editorial analogy, not the protocol's formal definition.
Where you will encounter it
You may see MCP in Codex, Claude Code, Cursor, AI-agent settings and tool installation pages. It often appears when an AI application needs access to development tools, documents, browsers or business systems.
Should you care
- Chat-only users: knowing the basic idea is enough.
- Codex / Claude Code / Cursor users: worth understanding before you add tools.
- People who install AI integrations regularly: especially useful to understand.
- Developers: understand the host, client and server roles rather than treating every MCP integration as the same thing.
How it differs
MCP vs : An API is an interface through which software communicates. MCP standardizes how AI applications discover and use connected capabilities. An MCP server may still call APIs behind the scenes.
MCP vs : MCP focuses on connecting capabilities; a Skill focuses more on reusable guidance for how an AI should handle a kind of task. They can work together.
MCP vs : A CLI lets an agent operate software through commands. MCP exposes capabilities through a protocol. If a good command-line tool already solves the problem, you may not need an additional MCP server.
What to read next
See whether Playwright MCP is worth installing, compare Playwright MCP with the CLI, or follow the guide to add an MCP server to Codex.
Related terms
Continue with these published explanations.
Sources & last checked
Sources supplied by the reviewed editorial pack. The check date records our pack cross-check, not a new live verification of these websites. Examples illustrate the concepts.
Official documentation
01Last checked: September 3, 2026