What is the difference between MCP and Function Call?

I have used Function Call extensively in previous projects. After MCP gained attention, I did not fully understand the difference between the two and the practical significance of MCP. After careful consideration of the essential difference between MCP and Function Call, I have some experience, which is summarized as follows.

Function Call

In a large language model (such as a large language model LLM), Function Call is a mechanism that enables the model to interact with external tools, APIs, or systems, allowing it to not only generate text but also perform specific tasks (such as querying a database, calling a computing interface, etc.). The following is its core workflow:

1. Defines the Tools that are called by Function Call.
2. When the user inputs, LLM matches the input with the definition of Tools and decides which Tool to execute. When calling, it dynamically generates parameters based on the user input information.
3. Tool can call external API, such as weather query; or read local files (such as enterprise jargon summary files)
4. LLM directly generates a reply based on the Tool call result, or decides whether to call other Tools based on the returned result.

MCP

The official MCP architecture is as follows:

This architecture definition is relatively abstract. Here, I combine the above Function Call, which I personally feel is easier to understand.

As can be seen from this figure, the biggest change is that the definition of Tool has been moved outside of Application. Specifically, the service provider who originally provided the API now defines Tools and encapsulates them into an MCP Server. For example, Amap encapsulates their geocoding/reverse geocoding/route planning APIs into Tools and provides an API. For details, please refer to https://lbs.amap.com/api/mcp-server/summary

In this way, application developers only need to implement the MCP Client in the original Application and call the corresponding MCP Server according to the instructions of LLM. The original Application is now also called MCP Host.

Advantages of MCP

Greatly simplify application development workload

Assume that an application needs to call 100 APIs, and the Function Call solution needs to implement 100 Tools. Now, this is no longer necessary because the Tools are implemented by the service provider. The application side only needs to implement one MCP client.

Easier to standardize

If an enterprise has multiple teams, each team must implement its own Tool for Function Call, and the implementation methods vary. Now such problems no longer exist.

Improve the reliability and reusability of tools

The tool is implemented by the service provider and provided to developers around the world for use. This greatly increases reusability compared to having each tool implemented by different developers around the world.

Improve the usability of the tool

If the application logic is not complicated, there is no need to develop your own MCP Host and MCP Client. With the help of existing tools such as Cursor and CLine, you can directly call the MCP Server to complete the work. For people without programming experience, the usability is greatly enhanced.