MCP 10 Questions and 10 Answers

Preface

During the communication process, I found that everyone didn’t know much about MCP, so I quickly compiled this article, hoping to give you an overall understanding.

Free Q&A: What is MCP?

MCP (Model Context Protocol) is an open protocol designed to standardize how large language models (LLMs) interact with external data sources, tools, and services.

Originally developed by Anthropic, it is now open source.

Question 1: What is the P in MCP, Protocol?

P is the abbreviation of Protocol, which means protocol. The protocol here generally refers to network protocols, such as the common TCP/IP, HTTP, etc., which are all network protocols.

The network protocol defines how the communicating parties establish a connection, how to send data, the definition of the data format, and error handling, etc.

MCP is a network protocol that defines the network communication data format for integrating AI models with external data sources and tools, including what transmission protocol to use, data format definition, authentication method, etc.

Question 2: What problem does MCP solve?

MCP solves the fragmentation problem of integrating large language models with external data sources and tools. Through standardized interfaces, MCP enables AI applications to connect to different data sources and tools more reliably and efficiently, avoiding the trouble of developing connectors for each data source separately.

Simply put, a tool can be used by both Claude and cursor through the MCP protocol, and neither the tool nor Claude needs to be developed a second time.

Question 3: What is the position of MCP?

It is located between the model and the tools/resources. The model calls the function through function calling, and the function interacts with the external tools/resources through the MCP protocol.

Question 4: MCP was developed by Anthropic, can other models be used?

Can.

Although MCP was developed by Anthropic, it is an open protocol that can be used by any model. It is model-independent. If the model does not have strong tool usage capabilities, using MCP will not enhance the tool usage capabilities.

Question 5: What is the difference between MCP and function calling?

They are not the same concept.

Function calling is the interaction between the big model and the external digital world. MCP is the interaction between the MCP host (chatbot or AI tool) and external tools/resources.

Generally speaking, the model first triggers a function call through function calling, and then triggers the MCP request in this function call.

Question 6: What is the difference between MCP and previous OpenAI plugins?

Very similar, but MCP is more standardized than previous plug-ins, defines clear data formats, transmission protocols, authentication methods, etc., and has stronger capabilities.

Question 7: What is the difference between MCP and GPTS?

GPTs is more like an application market, and currently it is an application market for people.

GPTs may be similar to the markplace of MCP server, but MCP server is mainly used by AI, not directly by humans.

Question 8: Will MCP become a standard?

In the short term, MCP has no rivals in terms of model integration of external resources and tools, and its ecosystem is growing.

But I am not sure whether OpenAI will support it or come up with its own specification. If they do, it must be better than MCP in design, otherwise they will be criticized for building a "small courtyard with high walls".

In the long run, there are some problems in the design of MCP itself, such as complexity, client and server coupling, distributed identity authentication, etc., which need to be solved.

Question 9: Do we need to access MCP?

From a technical point of view, if it is used internally, you need to balance flexibility and cost. You can do whatever you want for internal use, and MCP is not necessarily the best solution. However, if it involves connecting to external (such as public network) tools, it is best to use a protocol that everyone agrees on.

Question 10: Will domestic Internet platforms be connected to MCP?

Domestic platforms such as Meituan, Didi, Taobao, Pinduoduo, etc., I think it is highly unlikely that they will find an upstream for themselves in the short term.

Just like when Taobao cut off the traffic entrances of Baidu and WeChat.

Finally, a question: Is MCP applicable to intelligent agents?

MCP is not designed for intelligent agents, it is designed for models to connect to external resources and tools.

There are several unresolved issues, such as the centralized identity authentication technology of MCP, which is not suitable for identity authentication between agents. In MCP's CS architecture (client-server architecture), the server cannot actively connect to the client.

If you need to connect intelligent agents, you can try our ANP, which is specially designed for intelligent agents.