The promise and challenges of unification
The discourse surrounding AI interoperability is intensifying. Following Baidu’s announcement of its comprehensive MCP services at its developer conference last week, major Chinese tech firms like Alibaba, ByteDance, and Tencent have all embarked on the MCP journey.
MCP, or Model Context Protocol, is envisioned as a unifying standard that enables AI to interface seamlessly with a multitude of applications and services. It can be likened to the ubiquitous USB interface found in computers and smartphones, allowing for the plug-and-play integration of diverse external devices. In essence, MCP aims to provide AI with a universal ‘USB port’ to access tools and execute tasks.
In November 2024, Anthropic, an American AI company, introduced the MCP standard, which was quickly embraced by competitors such as OpenAI and Google, signaling a departure from the conventional competitive practice of proprietary ecosystems. Starting in April, leading Chinese tech companies, including Alibaba Cloud’s Bailian, Tencent Cloud’s Knowledge Engine, ByteDance’s Kouzi Space, and Baidu AI Cloud, launched their own comprehensive MCP services.
The primary objective of MCP is to foster unification, but this endeavor faces significant challenges. According to multiple developers and researchers, while MCP is effective for accessing local enterprise data, it encounters obstacles when attempting to integrate with internet applications for tasks such as booking flights, checking prices, and creating travel guides. These challenges stem from the immaturity of AI’s invocation processes and the limited availability of internet tools, with many platforms only offering access to peripheral functionalities.
Not all internet platforms are equally enthusiastic about adopting this common standard and joining the MCP service provider network. The closed nature of the Chinese internet ecosystem, coupled with heightened sensitivity to data privacy, has made many platforms cautious. They prefer to assess the viability and development of the MCP ecosystem before fully committing to it.
The AI landscape is known for its rapidly evolving terminology and concepts. When Anthropic initially open-sourced the MCP protocol late last year, the industry largely adopted a wait-and-see approach. However, the explosive popularity of Manus has since fueled interest in MCP within China.
MCP as a catalyst for AI agency
According to Hou Xinyi of Huazhong University of Science and Technology, the crucial step in transcending the limitations of ‘chatbots’ lies in enabling AI to interact with external data and tools, which is precisely what MCP seeks to facilitate.
Prior to MCP, alternative approaches were explored to address the perceived lack of ‘AI agency.’ In late 2023, OpenAI introduced the concept of an app store (GPT Store), allowing ChatGPT to leverage external tools through plugins based on a defined set of standards. Similar AI app stores, such as ByteDance’s Kouzi, Baidu’s Qianfan, and Alibaba’s Bailian, followed suit.
However, these approaches eventually reached their limits. Plugins and app stores shared a common problem: siloization. Each tool possessed its own unique development documentation, parameter formats, and interface specifications. This meant that developers had to reinvent the wheel each time they integrated a new tool into AI, resulting in inefficiencies.
Over time, the number of new tools added to app stores declined, and the quality of plugins varied significantly, hindering the ability to tackle complex tasks. This indicated that the existing approaches were approaching their limits.
MCP as a unifying solution
MCP is viewed as a promising solution due to its emphasis on unification. In its official documentation, Anthropic likens MCP to a universal USB-C interface for the AI world. Hou Xinyi prefers to describe it as a ‘docking station’—a versatile adapter that allows AI to connect to multiple external tools simultaneously, eliminating the need for format conversions.
Many anticipate that MCP will have a transformative impact, akin to Qin Shi Huang’s standardization of weights and measures, which facilitated trade and communication among the previously fragmented states of the Spring and Autumn period.
According to a technical lead at a major tech company’s intelligent interconnection working group, MCP also optimizes AI’s language interactions. Previously, AI required users to precisely state ‘I want to navigate’ to utilize a navigation service’s API. Even a slight deviation could cause the AI to fail. Now, each tool must provide standardized names, parameters, and functional descriptions. As a result, AI only needs to understand the user’s intent and then match it with the most appropriate MCP server based on the descriptions.
This approach aligns more closely with the inherent capabilities of large language models, enabling users to invoke services with a single sentence, moving away from the previous requirement for direct interface-to-interface communication.
MCP’s current adoption and limitations
Despite its perceived potential, MCP has not yet achieved widespread adoption, and its practical applications remain limited. Currently, MCP is most popular among enterprise technical personnel and independent developers.
As a front-end engineer, Gong Dian relies heavily on the AI programming assistant Cursor. However, Cursor has struggled to seamlessly integrate with his company’s internal project systems, requiring manual intervention. While plugins or function calls could be used previously, external AI could not access the company’s internal systems, and real-time invocation raised security concerns. MCP, on the other hand, can be initiated within the company’s internal network, making it more reliable and compliant.
Independent developer Zhu Mama recently instructed Cursor to learn MCP documentation and package Google Maps and Search APIs into an MCP server, which was then used to invoke Google’s Gemini large language model. The resulting MCP-equipped Gemini was transformed into a travel guide assistant. When asked about public transportation routes from Singapore Airport to various attractions, the assistant provided more detailed and accurate information compared to Doubao’s response.
Various travel assistants are emerging within the developer community. When ByteDance’s Kouzi Space launched its internal beta on April 19, the demonstration case was also a travel AI assistant, prompting some to joke about the industry’s obsession with travel.
Zhu Mama candidly admits that the focus on travel scenarios is primarily due to their relevance to everyday consumer needs. Another reason is the limited availability of MCP-compatible internet software in China, which restricts the market’s potential.
According to the latest statistics from the navigation platform MCP.so, there are over 11,028 MCP service providers worldwide, and the number is growing rapidly. However, within China, only a few major geographical location applications, such as AutoNavi, Baidu Maps, and Tencent Maps, currently function as large-scale MCP servers.
This limitation is why Zhu Mama’s plan to create a Chinese version of a travel assistant quickly stalled. To develop a Chinese travel guide, it would be ideal to utilize domestic map services. However, Zhu Mama discovered that the official MCP server provided by AutoNavi offered very limited information. While it could provide route queries between two locations, it lacked detailed information on landmarks, reviews, hotel ticket prices, and other essential details.
In contrast, Google Maps API provides detailed booking methods, hotel prices, hotel reviews, hotel facilities, and even price comparisons across multiple platforms, a level of detail that is difficult to imagine within the Chinese ecosystem.
While Tencent, Alibaba, ByteDance, and Baidu products are embracing MCP, their high-frequency applications have not yet formally joined the MCP service provider network. Platforms such as WeChat, Xiaohongshu, and Douyin, as well as lifestyle service platforms like Ele.me, Meituan, and Ctrip, are conspicuously absent.
Challenges in tool availability and AI scheduling
In addition to the limited availability of tools, the AI’s scheduling capabilities also pose a constraint. Zhu Mama packaged 6-8 API interfaces, including Google Hotels, Maps, and Search, into a single MCP server, which is far below the maximum limit (Cursor allows a maximum of 40 tools per agent). However, the AI was already struggling to determine which tool to invoke. When faced with complex requests, the AI was unable to break down the process and invoke MCP in stages, instead attempting to handle everything at once.
According to Gong Dian, MCP’s value hinges on the quality of both the client and server sides. Just as a USB port has no inherent capabilities and relies on the services behind it, MCP requires robust services to realize its potential.
MCP lays the foundation for AI agents, but it does not solve all the problems. A standard that remains unused is merely a piece of paper.
The aforementioned technical lead suggests that the widespread adoption of Anthropic’s MCP standard is due to its open-source, non-profit nature and the credibility of its creator. Other organizations are willing to follow a standard set by a reputable entity.
Currently, small and medium-sized companies and large internet companies seeking to diversify their revenue streams are the primary adopters of the MCP standard.
AI companionship company MiniMax recently launched an MCP server, with community manager Cai Jiaren stating that developers can use MCP to invoke MiniMax’s multi-modal capabilities for video generation, voice generation, and voice cloning. The MCP includes strict access control mechanisms to ensure compliance when enterprises access internal data. The overall invocation process is also simplified, without adding extra token costs.
MiniMax’s decision to launch an MCP server was driven by the desire to enable global developers to easily leverage MiniMax’s model capabilities and unlock more flexible and efficient creation.
Other startups share similar aspirations. Biu Technology mentioned in an interview that developers can use AutoNavi MCP to obtain transportation data and then use Biu’s products to generate a PPT. MCP lowers the barrier to entry by providing access to AutoNavi’s interface, which would otherwise be unavailable to them.
The aforementioned technical lead believes that MCP is essentially a story about service providers. By encapsulating their APIs according to the MCP standard, application service providers can make their services accessible to all AI.
Divergences and concerns among service providers
However, disagreements arise among service providers. Many companies are not fully committed to the idea. While major platforms like AutoNavi and Baidu Maps have launched MCP servers, they primarily repackage existing API interfaces, offering conventional functionalities while maintaining strict control over core user permissions and transaction data.
In addition to map location services, a third-party developer’s Xiaohongshu auto-publisher, which automates the search and posting of content, is currently the most popular item on the Modeng community’s MCP plaza. Hou Xinyi suggests that this may have limited impact on social content platforms like Xiaohongshu, but data and permissions become particularly sensitive in transaction-intensive scenarios like food delivery platforms.
One of the primary concerns for service providers is the control of the user experience.
For example, opening up a complete food delivery service requires granting AI agents access to sensitive personal data such as prices, store information, and user addresses and contact information. Anthropic has acknowledged that MCP’s security system, including permission management and invocation auditing, is still under development. Consequently, some platforms are concerned about the risk of unauthorized invocation when connecting to MCP.
Some platforms are testing relatively safe transaction scenarios. For example, Alipay recently launched an MCP server, claiming to give AI agents ‘one-click access to payment capabilities.’ However, a closer look reveals that it primarily offers collection rather than payment services.
According to Hou Xinyi, Alipay’s approach focuses on facilitating merchants’ payment collection rather than allowing AI to make payments on behalf of consumers. This is a viable option, as allowing AI to control wallets and place orders freely is not yet secure enough for everyone’s comfort. This is also the key reason why transaction services cannot be widely promoted.
A deeper issue is that if AI freely participates in the transaction process – helping users compare prices or recommending the most cost-effective restaurant – it would undoubtedly provide significant convenience for users. However, it would also mean that service platforms would lose control over the user’s selection process, and their core algorithm advantages would be marginalized, reducing them to ordinary suppliers.
Addressing security and promoting universality
Multiple interviewees believe that MCP needs to address two key issues: security and universality.
First, security. Hou Xinyi points out that MCP faces two security challenges: a lack of centralized security supervision and an incomplete identity verification and data authorization mechanism. Currently, there is no official ‘discovery plaza’ for MCP. Many third-party navigation platforms collect MCP services by directly pulling code projects from GitHub, which is fast and straightforward but lacks a formal review process. Anthropic has stated that it will formally address the MCP hosting mechanism and discoverability issues this year. Anthropic’s recently updated protocol draft is working to address this shortcoming. In addition, domestic organizations such as IIFAA (Internet Trusted Authentication Alliance) are attempting to fill the security gap.
There are also long-standing issues in the AI agent field, such as prompt hijacking and tool combination attacks. However, the aforementioned technical lead believes that these are not MCP vulnerabilities but rather risks that exist for any AI agent. Currently, no obvious security vulnerabilities have been found in the MCP protocol itself, and the data transmission and interaction mechanisms are generally reliable.
Security is just the first hurdle. The real challenge is overcoming the interest defenses of manufacturers and persuading more manufacturers to become MCP servers.
According to Hou Xinyi, this is related to the understanding of the ‘walled garden’ nature of internet platforms. Data is an important competitive barrier for various platforms, so many manufacturers may only open up some peripheral functions as MCP servers for testing. Manufacturers may need to wait and see how much impact the MCP ecosystem will have.
The aforementioned person in charge said that if it is connected to AI as an MCP server, it can obtain more user data and habits, and give back to its own base model, which may become the biggest motivation for manufacturers to actively join.
When the MCP server market is truly abundant, more distant issues must be considered.
For example, how do smart bodies call different Apps on mobile phones? The person in charge mentioned that to wake up another App through the local AI smart body of the mobile phone, there will be an extra layer of application authorization and identity verification, which is not as simple as MCP calling cloud services, and there is currently no particularly suitable solution.
For another example, when the service supply is excessive, how do smart bodies make choices - call JD takeaway or Meituan takeaway? Use Gaode map or Baidu map? Multiple interviewees mentioned that today’s MCP invocation logic is still very basic, mainly determined by the ‘functional description’ of the service provider, and there is no sorting and optimization mechanism. If a service provider deliberately adds inductive language to the description, such as ‘most efficient’ and ‘must-choose’, AI may be misled and diverted to places it should not go.
As the person in charge of the aforementioned technology explained, ‘It’s like you can’t find the service you want in the search engine, but a bunch of messy information pops up. How to accurately match the service that users need most, the future MCP ecosystem will also face the same problem.’
Ultimately, the implementation process of any standard is full of challenges. Hou Xinyi said that to promote the popularization of MCP, a key opportunity similar to Manus may be needed to truly make the entire industry realize the power of MCP.