"Isn't MCP just an API?" Not quite — and the difference is the whole point. An API is how two programs talk. MCP is how an AI model discovers and uses tools at runtime, and it usually sits on top of an API.

The key differences

	API	MCP
Audience	A developer writing code	A model choosing tools mid-conversation
Discovery	You read docs, then hard-code calls	The client asks the server what it offers
Shape	Endpoints, params, schemas you learn	Named tools with descriptions the model reads
Coupling	Tight — you wire each integration	Loose — any MCP client works with any server

Put simply: with an API, a developer decides in advance exactly which calls to make. With MCP, the model decides at runtime, choosing from tools a server describes in plain language.

They work together

Most MCP servers are thin wrappers over an existing API: the server calls the API, and MCP is the standard envelope that lets any assistant use it without bespoke glue. You still need the API — MCP just makes it AI-native.

Which should you build?

Building an integration for other programs? Ship an API. Want an AI assistant to use your system without someone hand-wiring it? Add an MCP server in front. Truffle does both — a REST API for programmatic access and an MCP server for assistants — because they serve different callers. New to MCP itself? Start with What is MCP?