MCP has been very popular these days, so let’s come out and explain.

Imagine you have a very smart friend. This friend is knowledgeable and can answer all kinds of questions, but he is confined to a closed room and cannot directly use external tools and services. Whenever you need him to help you book a restaurant, check the real-time weather, or analyze the latest data, you have to act as a "middleman" - bring external information to him, and then take his instructions outside to execute. Such a communication process is cumbersome and inefficient.

In the AI ​​world, Large Language Models (LLMs) are like this smart friend. They have powerful understanding and generation capabilities, but they themselves cannot directly interact with external tools and services.

The Model Context Protocol (MCP) launched by Anthropic in November 2024 is an innovative solution to this problem. Although it was launched last year, it became popular after the new year, proving that good things also need time to ferment.

1. What is MCP?

The Model Context Protocol (MCP) is a standardized protocol framework designed to enable AI models to access external tools and services at a lower cost and in a more universal way. In simple terms, MCP is like a "universal translator" and "intelligent connector" between AI and the outside world.

If we explain it with an example from everyday life, MCP is like:

1. A universal power adapter: No matter where you are in the world, you can use the same adapter to connect various electrical appliances

2. A multilingual translator: able to translate the "languages" between different systems in real time to ensure smooth communication

3. An intelligent control center: helps AI send requests in the correct format to the outside world and converts external results into a form that AI can understand

2. How does MCP work?

The architecture of MCP mainly includes three core components. A simple working principle diagram is drawn for reference only.

1. MCP Client

Imagine you are using a remote control to control various smart devices in your home. The MCP client is like this remote control. It is embedded in the AI ​​model or its interface layer and is responsible for:

• Convert AI model generated requests into standardized MCP format
• Passing these requests to external tools and services
• Receive responses from the outside and return them to the AI ​​model
  
  

2. MCP server

Continuing with the analogy of a smart home, the MCP server is like a smart hub that connects various home appliances. It is responsible for:

• Receive standardized requests from MCP clients
• Convert these requests into a format that a specific tool or service can understand
• Call the corresponding tool or API
• Convert the execution result back to a standardized format and return it to the client
  
  

3. Tools and Service Adapters

These adapters act like receivers on various smart home appliances and can:

• Understand the commands sent from the MCP server
• Perform specific functions (such as querying data, controlling equipment, performing calculations, etc.)
• Return the execution result to the MCP server
  
  

Let’s understand the workflow of MCP through a practical example: suppose you ask the AI ​​assistant to check today’s weather and recommend clothing suggestions based on the weather conditions.

1. User makes a request : You ask the AI ​​assistant "What's the weather like today? What should I wear?"

2. AI understands and generates intent : The AI ​​model understands that you need weather information and clothing suggestions

3. MCP client processing :

• Convert AI intent into standardized requests: "Need to query {today's} weather data for {current location}"
• Send this standardized request to the MCP server
  
  

• Receives the request and recognizes that a call to the weather API is required
• Convert the standardized request to the format required by the specific weather API
• Call the weather API to get data

  
  

• The weather API returns data such as temperature, humidity, and precipitation probability for the day
• The data is converted into a standard format by the MCP server.
  
  

• The MCP client returns the weather data in a standard format to the AI
• AI generates dressing suggestions based on these real-time data

  
  

The whole process is seamless and users do not feel the complicated interaction process behind it.

Wait, isn’t this the function of real-time query of information on large models connected to the Internet that we can experience now?

You're right. So let's explain.

4. Why is MCP so important?

There are three things to clarify here: Model Network Query vs Model Connection API vs Model Connection MCP Protocol

Model online query: All information can be found, but it is only information acquisition and cannot issue instructions to the corresponding business provider. The technology is currently mature and universal;

The model is connected to a single specific business API: it can be queried through the API, and instructions can be issued to the business through the API. The cost and customization of a single API connection are very high;

The model docking MCP protocol has three real advantages.

1. Significantly reduce integration costs

Before MCP, whenever developers wanted AI to use a new tool or service, they needed to perform customized development and create specialized interfaces and processing logic for each tool.

With MCP, developers only need to develop an adapter once according to a unified standard, and AI can access the tool. Just as the popularity of USB interfaces has eliminated the need to prepare different cables for each device, MCP makes the connection between AI and tools simple and efficient.

2. Achieving true universality

Under traditional methods, different AI models may use different interface standards, which requires tool developers to customize solutions for each AI platform.

MCP provides a unified standard, allowing the same tool to be used by different AI models, while allowing an AI model to use a variety of tools that conform to the MCP standard. This greatly improves the compatibility and openness of the ecosystem.

3. Enhance AI capabilities without retraining

Imagine upgrading your phone — you don’t have to relearn how to use it, you can just install new apps to get new features.

MCP enables AI models to similarly “install” new tools, thereby acquiring new capabilities, without the need to retrain or fine-tune the AI ​​itself. This means that AI can continuously upgrade its capabilities to keep up with rapidly changing needs.

5. Why is the ecosystem willing to connect to the MCP protocol?

This is a question that may arise from an ecological perspective.

It is equivalent to the main data and service assets for users operated by a company (such as Taobao's product search or Meituan's takeaway products). Why are they willing to access the MCP protocol? After access, these tools and assets can be used not only in their own business scenarios.

Tool service providers’ access to MCP does involve core assets, and they have several key motivations for doing so:

1. Expand user reach channels

• By using AI assistant as a new entry point, more potential users can be reached
• When users ask AI "What delicious takeaways are nearby?", takeaway platforms connected to MCP can be naturally recommended.
  
  
  

• Users can complete the entire process from inquiry to ordering in one conversation interface
• Reduce the friction for users switching between multiple apps
  

• New business models and revenue sources based on AI conversations can be established
• Potential traffic conversion rate may be higher than traditional channels
  
  
  

• As AI assistants become more common, not having access could lead to a competitive disadvantage
• If competitors access the network but you do not, you may lose market share.

• 
  

5. Controllable asset opening

• MCP allows service providers to customize the degree and scope of interface openness
• Only some functions can be opened to retain core differentiation capabilities
• Access control and usage restrictions can be implemented, and the tool service remains on the provider's server

• 
  

5. Data Insights

• Gain new data insights into user interactions through AI
• Understand the real needs expressed by users in the conversation
  
  

In order to protect core assets, service providers usually adopt a sophisticated API design strategy, open only necessary functions, and limit excessive use of data through contracts and technical means. This is a balance between openness and protection.

6. What is the industry value and application scenarios of MCP?




Just because a new concept becomes popular doesn’t mean it will continue to be popular. This is especially true when it comes to building a new set of standards and requires a lot of ecosystem compatibility. But if this works, the future is promising.




1. Enterprise internal system integration: For enterprises with complex internal systems, MCP allows AI assistants to easily connect to various enterprise resources. This allows enterprises to quickly build powerful customized AI assistants without a lot of engineering resources.




• Connect to CRM system to query customer information
• Access internal knowledge base to provide accurate answers
• Interact with project management tools to update task status
  
  

2. A thriving developer ecosystem: MCP creates new opportunities for developers, allowing companies to quickly build powerful customized AI assistants without requiring extensive engineering resources.




• Tool developers can create once and serve multiple AI platforms
• AI platforms can more easily expand their function repertoire
• Third-party developers can create specialized toolkits for specific areas
  
  

3. Customized solutions for vertical industries: Different industries have specific tool and service requirements. MCP makes it easier to customize AI solutions for these industries, and also allows consumer AI assistants to connect to various daily services through MCP. Typical examples include the following three industries.




• Healthcare: Connecting electronic medical record systems, medical knowledge bases, and diagnostic tools
• Financial industry: access to market data, risk assessment tools and trading platforms
• Education: Connecting learning management systems, assessment tools, and educational resource libraries
  
  

4. Expansion of capabilities of all personal AI assistants on the market: This allows AI assistants to truly become powerful assistants in daily life, rather than just a conversational robot.




• Memo/schedule/task management application: For various high-frequency information + task writing and reading, you can directly close the loop in the AI ​​dialogue
• Smart home control system: The various protocols of the previously scattered IOT business can also be used together under the smart home hub
• Online shopping platforms and reservation services: AI can better assist with daily consumption scenarios such as food, clothing, housing and transportation




7. Is there such a thing as MCP in China?




There are indeed many companies in China that are already developing protocol standards and platforms similar to MCP, but the biggest difference is:




Domestic shopping malls are still mainly self-use closed ecosystems, connecting the ecosystem to their own platforms; but few have the ability to build an intermediate layer, because this requires a design approach of " taking from the ecosystem and giving back to the ecosystem ."




This type of business is mainly distributed in two camps: AI large model manufacturers and mobile phone hardware manufacturers:




AI big model manufacturers:




• Baidu: Its Qianfan big model platform provides plug-ins and agent frameworks, and the Wenxin Yiyan toolset allows developers to access their own tool services, using a mechanism similar to Function Calling, but with its own protocol specifications.
• Alibaba Cloud: The Tongyi Qianwen tool calling system includes function calling and Agent capabilities. The Lingji platform provides a tool and service access framework that supports plug-in ecosystem construction and knowledge base connection.
• Tencent: The Skylark Big Model Platform provides a tool calling API, and the Button Platform is its application and tool access framework.
• iFlytek: The Spark Cognitive Service Platform provides tool calling capabilities, and the iFlytek AI Capability Open Platform provides tool access specifications.
• Zhipu AI: Zhipu GLM open platform supports APIs called by tools.
  
  

Mobile phone hardware manufacturers:

• Xiaomi: XiaoAi Super Brain is its AI tool ecosystem access platform, supporting the interconnection between XiaoAi and mobile phone applications.
• Huawei: Hongmeng's intelligent assistant Xiaoyi allows applications to serve its intent framework, supporting the distribution of application traffic at various Hongmeng entrances and flexible service calls.
• OPPO: Xiaobu Assistant Open Platform provides an application and service access framework, and the AILink framework supports application capability connection.
• vivo: Jovi smart assistant provides service access capabilities, and the Blue Heart model provides an application tool calling framework.
  
  

The main differences between these platforms are:

• Access method: Large model manufacturers mostly adopt API-based access methods, and the standards are closer to the international MCP; mobile phone manufacturers mostly emphasize deep integration with the operating system and pay more attention to device-side capabilities.
• Degree of openness: Large model platforms are usually more open to third-party developers; mobile phone manufacturer platforms sometimes focus more on access to applications within the ecosystem.
• Technical route: Some platforms emphasize cloud calls, while others focus on device-side capabilities.
• Authentication and security mechanisms vary.
  
  

The interface specifications and capabilities of each platform are also constantly being updated and improved, but if they want to be comparable to MCP in the future, the overall trend must be to develop in a more standardized and open direction.




To summarize:

MCP represents an important breakthrough in the interconnection between AI and the outside world. It is not only a technical framework, but also a bridge connecting AI capabilities with actual application scenarios, turning AI from a "brain that understands the world" into a "hand that can operate the world."

As MCP becomes standardized and mature, we can expect the AI ​​application ecosystem to flourish, bringing unprecedented convenience and value. This will not only enhance the practicality of AI, but also create new opportunities and possibilities for all walks of life.

PS: Part of this article was written by AI, thanks to Claude 3.7, who can both write and generate pictures.