Function Calling and the Model Context Protocol

Ask any dev to pick between Function Calling and the Model Context Protocol and you may get a blank stare. They solve different problems, so lumping them together only muddies the water. Let’s clear it up.

Function Calling: the model decides, your app executes

Function Calling turns a natural‑language request into a tidy JSON payload, then hands it off to your code.

• The large language model (LLM) reads the user prompt, chooses the right function, and fills in the parameters.
• Your application owns every line of execution logic. The model never fetches data or flips a switch on its own.
• This pattern shines when tasks are well defined, the surface area is small, and you need tight control over side effects.

Example: A user types “What’s the weather in Paris?” The model returns

{  
  "name": "get_weather",  
  "arguments": { "location": "Paris" }  
}

Your back‑end calls the weather API, formats the answer, and sends it back to the user. The model never leaves its lane.

Model Context Protocol: a common language for every tool in the box

The Model Context Protocol (MCP) tackles a broader headache: wiring an AI client to many tools without writing a one‑off connector for each one.

• MCP defines a standard, open protocol (think JSON‑RPC over SSE in Copilot Studio today).
• An MCP server lists its “tools.” Any compliant client can call them without caring who wrote the server or where it runs.
• Real‑time data flows stay inside clear security boundaries because the protocol handles auth and transport details.

Need to juggle a dozen services from different vendors? Speak MCP once and skip the tangle of bespoke adapters.

Don’t miss MCP Sampling

Most people skim past Sampling, but it is MCP’s secret weapon. A server can ping the AI mid‑workflow, ask follow‑up questions, or even loop in another MCP server. The user sees each prompt and response, so nothing slips by unchecked. Picture a Figma plug‑in that needs new images. It asks the AI for ideas, shows them to you for approval, then calls a Midjourney MCP to generate the art before dropping the files back into Google Drive. All that choreography happens through Sampling with you in the driver’s seat.

When to reach for each approach

• Pick Function Calling when you only need the model to pick an action and your codebase can handle the rest. Simple, scoped tasks. Full control stays with you.
• Pick MCP when you integrate many tools, care about vendor neutrality, or want advanced hand‑offs like Sampling. It is about scale, interoperability, and future‑proofing, not single calls.

The quick take‑away

Function Calling lets the model decide what to do. MCP standardises how that decision travels across your stack. Use the first for tight, local tasks. Use the second when you are building a city, not a cabin.

Now you can nod with confidence the next time someone mixes the two up.