Can everything be MCP? MCP needs to be reduced

In November 2024, the Model Context Protocol (MCP) released by Anthropic,  known as the "USB interface of the AI ​​world", swept the world like a hurricane.

In a live broadcast held by Open Source China with the theme of "What is the MCP that is so popular on the entire Internet?", we saw some industry consensus and heard many different voices.

MCP has become a de facto standard and will not be replaced in the short term

MCP has been widely used in a variety of AI tools, including Claude, ChatGPT, Replit, and Cursor, and supports SDKs in multiple programming languages ​​(such as Python, TypeScript, Java, Kotlin, and C#), further expanding its application scope. According to public information from MCP.so, the world's largest  MCP server collection website  , the number of MCP servers included has exceeded  13,000 . Judging from the current development trend, MCP has become a  recognized de facto standard in the industry.

" The essence of the so-called protocol standard lies not in technological advancement but in industry consensus. When enough participants recognize its value, it will naturally become a standard. " Liu Kang, a technical expert at ByteDance, believes that the success of MCP is due to the right time and place : backed by core manufacturers with industry influence, as of now, leading companies such as OpenAI and Google have clearly stated that they will connect to its ecosystem, which directly consolidates its market position.

" In the short term, it is unlikely that similar protocols will emerge to challenge the status of MCP. " Luo Yaxin, head of Gitee Public Cloud Technology , said that the MCP protocol was originally created to solve the fragmentation problem when AI Agents call various tools - because different tools and different software now have their own interface standards, and even the manual format of each tool may be completely different, so each tool must be adapted separately, which is equivalent to duplication of work. Therefore, the development team initially referred to the proven standard LSP - just like programming language tools integrate various compilers through a unified protocol, they migrated this idea to the AI ​​field.

With MCP, a set of standard formats can be defined for all tool calls, allowing Agents to obtain resources and call functions more smoothly. Therefore, even if new protocols may appear in the future, they are essentially along the main line of "expanding Agent capabilities". At most, they will be optimized in implementation details (such as stronger compatibility or better performance), but the underlying logic will not change.

"This standardization is an inevitable trend," Luo Yaxin pointed out, "just like the development tools that went from working independently to unified interfaces through LSPs. When the system complexity increases and the number of tools that Agents need to connect to explodes, this kind of protocol is necessary to break the chaos. This is a natural result of technological evolution."

The overall design of MCP is good, but the structure is complicated

In this booming development trend, MCP is also being deified. Therefore, there is a popular saying that everything can be MCP.

" The MCP protocol has been mentioned a lot, and people's expectations of it go beyond the scope of its definition. " Liu Kang believes that there is a clear deviation between the public's perception and the technical circle's understanding : the technical community knows that MCP is essentially a protocol that defines the interaction between agents and external tool resources, but the public discussion gives it expectations that go beyond the design scope . In fact, the protocol only focuses on standardized connection interfaces, but it is often misunderstood in the public opinion field as being able to solve all problems in the intelligent agent ecosystem.

In Chang Gaowei's opinion, although the overall design of MCP is good - for example, the three core concepts of the server (resources, tools, and prompt words) are reasonably designed, the protocol structure is indeed relatively complex . For example, the client supports root permissions while opening up sampling capabilities and giving the client AI capabilities to the server .

" I don't fully understand the design logic . First, the sampling function tends to be over-designed due to the lack of typical application scenarios; second, the degree of implementation is also questionable - even the sampling support of the Cloud client has not yet been implemented, let alone the complete server-side capability migration. I guess they may have imagined how to use this function when designing it ." Chang Gaowei is the author of the intelligent agent protocol ANP. When MCP was first released, there was no identity authentication mechanism. Chang Gaowei and others also provided MCP with their own DID proposal.

Recently, MCP's SSE mode has been adjusted, and Chang Gaowei thinks this improvement is good. The new solution adopts the " Streamable HTTP" mechanism to improve compatibility and achieve statelessness . " In short, the essence of protocol design is to use simple solutions to solve specific problems. Overly complex designs will affect its life cycle . "

“ MCP needs to do subtraction ”

After thoroughly studying the MCP protocol documentation and conducting practical verification - developing the universal database gateway DBHub that implements the MCP server interface, Bytebase CEO Chen Tianzhou became confused. " What it does is not pure enough - on the one hand, it involves functions at the data acquisition level, and on the other hand, it intervenes in the implementation of the application layer . There is also controversy in the technical community about MCP: should it be positioned as an agent development framework or as a universal data interface? "

This problem has been exposed especially after Google launched the intelligent communication protocol A2A. Although the A2A protocol focuses more on application layer specifications, some of its functions do overlap with MCP.

"MCP needs to do subtraction and should focus on becoming a pure atomic capability interface for information acquisition, rather than the agent development direction mentioned in its roadmap."  Chen Tianzhou believes that the development prospects of MCP will depend on the strategic focus of the development team. If MCP insists on extending to the agent field, its development prospects are not optimistic. On the contrary, if it can focus on building a standardized information acquisition interface, MCP will have strong vitality. "Just like the current application system needs DAL (data access layer) when accessing the database, MCP can completely occupy a similar 'basic capability layer' ecological niche. But if it tries to develop to a higher level of abstraction, there is greater uncertainty."

Chang Gaowei also agrees with this view: " MCP will have better development if it focuses on model connection tools and resources . If you want to solve this problem and that problem at the same time, it may not be a good solution."

Chang Gaowei pointed out that the more complex the protocol is, the more difficult it will be to use. In addition, MCP has developed to a considerable scale of application and has become the de facto standard in the field of model connection tools and resources. "From a global perspective, there will not be a second MCP in the entire industry. Unless MCP wants to be involved in both the intelligent entity layer and the application layer, it may give up this advantageous position."

Lin Jingjing, product director of Gitee Private Cloud, believes that at this stage, MCP does have a first-mover advantage, but this does not mean that it will be smooth sailing. "As a protocol that connects agents and tools, the outside world currently has too high expectations for MCP. Whether it will go smoothly in the future depends on whether MCP can evolve quickly and make up for its shortcomings. A more likely development path is that MCP is forced to specialize in a certain niche by the new protocol, while other functions are divided up by the new protocol. "