MCP often gets confused with related ideas. Here's how they fit together.
Function Calling
Function calling (or tool use) is a model capability: given tool definitions, the model can output a structured request to call one. It says nothing about where tools come from.
MCP
MCP is a protocol for delivering tools, resources and prompts to AI applications from separate servers. A host connects to MCP servers, collects their tool definitions and passes them to the model using its function-calling capability.
So MCP builds on function calling rather than replacing it.
Framework Integrations
Agent frameworks have their own tool abstractions, which usually only work inside that framework. MCP servers work with any compatible host.
Direct API Integration
You can always write custom code calling an API directly. That's simplest for a single application with one integration. MCP pays off when many hosts need the same integration, or one host needs many integrations.
Choosing
- One app, one integration → direct code may be enough.
- Many apps or many integrations → MCP servers are reusable.
- Tools for agents inside a specific product → consider offering an MCP server so users' AI tools can reach it.