Explain the Model Context Protocol (MCP) as simply as possible

How can a large language model (LLM) operate the real world like humans? The MCP protocol (Model Context Protocol) released by Anthropic three months ago is providing the answer.

This technology, known in the industry as the " AI version of USB ", is reshaping the way LLM connects to the digital world at a speed visible to the naked eye - just like when USB unified the interface standards of electronic devices, MCP is allowing AI to evolve from "being able to chat" to "being able to do things".

1. Background and Concept of MCP

You must be familiar with the USB interface, which is a standardized and universal interface that allows different devices to connect and interact through the same interface. The emergence of MCP is to solve the problem of complex interfaces between large language models (LLMs) and external services.

In the traditional software ecosystem, when you want to connect LLM (such as GPT-4) with external services, you often need to design different interfaces and calling schemes for each service, which is both complicated and time-consuming. The core goal of the MCP protocol is to provide a unified standard to simplify the connection and calling between different tools.

With the MCP protocol, all LLM tools, whether enterprise APIs or simple command-line tools, can be easily integrated into applications as long as they follow the MCP standard. This "plug-and-play" approach can significantly reduce development and maintenance costs and facilitate the construction of complex LLM applications.

2. Working Mechanism of MCP

1. MCP Server

In the MCP ecosystem, each external service or tool needs to implement an MCP server. This server acts like a "bridge" that connects external tools with LLM. Specifically, it provides the following functions:

• •  Tool List Query : When a query is received tools/list When requested, the MCP server will return a description of the tools and functions it supports.
• •  Tool call execution : When a tool call is received tools/call When requested, the MCP server performs the operations of the specified tool and returns the results.

In this way, developers do not need to worry about the interfaces and documentation of each platform, but only need to focus on the actual functional implementation of the tool.

2. MCP Client

The MCP client is an “integrator” that connects to various MCP servers. In the application, developers only need to configure a list containing all MCP server addresses, and the client can automatically complete the following tasks:

• • Automatically access each MCP server to obtain tool information.
• • Integrate these tools into LLM in a unified way for model calling.

This modular and decoupled design concept makes the application more scalable and also enhances interoperability between different services.

3. Practical Example: Domino's Pizza Ordering

Suppose Domino's Pizza wants to provide ordering services through LLM. In the traditional method, developers need to write ordering code and design interfaces for different platforms. With MCP, Domino's only needs to build an MCP server and perform the following steps:

1. 1.  Tool exposure : Return a list of pizza ordering tools through the MCP server.
2. 2.  Tool call : Integrate the MCP client in the LLM application to automatically call the pizza ordering tool.

This approach not only simplifies the tool docking process, but also ensures the consistency of the interface. For example, developers only need to run simple commands such as npm install @dominos/pizza-mcp-server, you can easily integrate the ordering function.

3. Practical Example: Domino's Pizza Ordering

Take Domino's Pizza as an example, suppose it wants to provide meal subscription service through LLM.

In the traditional method, developers need to write special pizza ordering code and write different calling interfaces for different model platforms. With MCP, Domino's only needs to build an MCP server to achieve the following two steps:

1. 1.  Tool exposure : A list of tools including pizza ordering functionality is returned through the MCP server.
2. 2.  Tool invocation : Integrate the MCP client in the LLM application to automatically invoke the pizza ordering tool without writing additional code for each model.

This approach not only allows the pizza ordering function to be quickly connected to various LLM platforms, but also ensures the consistency and ease of use of the interface. For example, simply executing something like npm install @dominos/pizza-mcp-server command to integrate pizza ordering functionality into your app.

4. Extended Functionality: Prompts and Resources

In addition to tools, the MCP protocol supports two additional primitives - prompts and resources :

• •  Prompts : Similar to the prompt directory, it allows the MCP server to provide preset prompt words or templates. Although this feature is not yet widely used, it may provide more customized interaction methods for LLM in the future.
• •  Resources : References to specific documents or resources, commonly used in enhanced retrieval (RAG) or to provide additional contextual information for LLM.

Currently, the public code base of the MCP protocol is mostly focused on tool functions, but with development, prompts and resources are expected to be used more widely.

V. Summary and Outlook

The MCP protocol marks the development of the LLM tool ecosystem towards standardization and modularization. Through unified interface standards, both large companies and individual developers can integrate various tools into LLM at low cost. This " plug and play " experience will greatly promote the popularization and application of LLM technology.

With the increase of MCP-compatible applications and the further expansion of functions such as prompts and resources, LLM will have richer interactive capabilities in the future. For developers, understanding and mastering the MCP protocol will become an important skill for building efficient and flexible LLM applications.