What problem does MCP solve that function-calling alone does not?

**Function calling** lets a model ask your code to run a tool, like a weather lookup or a database query. The problem is that each provider defines tool schemas in its own format. You register your tools one way for OpenAI and a different way for Anthropic. If you have ten tools and want to support three models, you write the integration many times over. Switch the underlying model and you redo the wiring. This is the M times N problem: M models multiplied by N tools. The **Model Context Protocol** (MCP) fixes this with one shared standard. A tool is published once as an MCP server. Any model with an MCP client can discover and call it. Now you only build N tool servers and M model clients, which add up instead of multiplying. MCP also lets a tool run as a separate process, even on another machine. The protocol carries tool definitions and calls over a defined transport, so the tool implementation is decoupled from which model you happen to use. The key idea is a common contract. Build the tool once, and any compliant model can use it without rewiring.

**Function calling** is provider-specific. You describe each tool in OpenAI's schema, then describe it again in Anthropic's slightly different schema. With M models and N tools you face up to M times N bespoke integrations, and switching the model forces a rewrite. The Model Context Protocol turns that product into a sum: $$M \times N \;\longrightarrow\; M + N$$ MCP defines one standard contract between an LLM client and a tool server. A tool is published once as a server. Any model with an MCP client can discover and invoke it. You build N servers and M clients, and they compose instead of multiplying. The protocol exposes three primitives. Tools are model-invoked actions. Resources are read-only context the host can pull in. Prompts are reusable templates the server offers. The transport layer is the wire format that carries this traffic. Stdio is used for local servers spawned as child processes. HTTP with Server-Sent Events is used for remote servers. Because servers run out of process, the tool implementation is fully decoupled from the model choice.