"Is MCP just a new kind of API?" Not quite. An API is how two programs talk; MCP is a standard for how an AI model discovers and uses tools — often APIs underneath. They are layers, not rivals.
What each one is
- API — a contract a service publishes so other software can call it. REST, GraphQL, gRPC. It defines endpoints, parameters and responses. Built for developers writing code.
- MCP (Model Context Protocol) — a standard for exposing tools and data to an AI model in a way it can discover and call at runtime, without bespoke integration code per app.
The key difference: who is the consumer
An API is consumed by code a developer wrote. The developer reads the docs, hardcodes the calls, handles auth and parsing.
MCP is consumed by the model itself. The server advertises its tools; the model decides which to call and with what arguments, on the fly. You did not write code to wire each tool into the app — you just added the server.
Why MCP exists
Before MCP, every AI app built custom integrations for every service: N apps × M services = chaos. MCP makes it N + M: each service ships one MCP server, each app speaks MCP once. That standardisation is why the ecosystem grew from dozens of servers to thousands in a year.
They work together
Most MCP servers are thin wrappers over existing APIs. The GitHub MCP server calls the GitHub REST API; the Stripe server calls Stripe's API. MCP is the AI-facing standard; the API is still doing the work underneath. You can even generate an MCP server from an OpenAPI spec.
When to think in which
- Building traditional software integration → API.
- Giving an AI model new abilities across many apps → MCP server over that API.
Going further
See what is an MCP server for the building block, what is MCP for the protocol, and MCP vs OpenAI functions vs LangChain for how it compares to other tool-calling approaches.