The “Universal Socket” of the AI Era
A new front has emerged in the ever-evolving landscape of Artificial Intelligence (AI), captivating the attention of tech giants and developers alike. This ‘new darling’ is the Model Context Protocol (MCP), a technology that promises to revolutionize how AI models interact with the external world.
On April 29th, Alibaba made waves by releasing and open-sourcing its next-generation Tongyi Qianwen model, Qwen3. This model not only boasts enhanced performance and reduced costs but also features improved support for MCP. This advancement allows developers to seamlessly integrate various external data sources and tools with the Qwen3 large language model, paving the way for more cost-effective and efficient development of AI Agents (intelligent agents).
Enthusiastic developers are already experimenting with Qwen3 to create innovative applications. Feng Lei, the founder of Mars Waves and former product leader at MiniMax Conch AI, developed a webpage that successfully utilized various MCPs, including images, audio, and maps, to quickly achieve desired prompt results. He shared his accomplishment on social media, emphasizing the critical importance of native MCP support.
Similarly, during the Create 2025 Baidu AI Developer Conference on April 25th, Baidu founder Li Yanhong stated that ‘MCP enables AI to better understand the external world, access information more easily, and utilize tools more freely. We believe MCP is a significant step forward for AI development, and developers should understand and embrace it as soon as possible.’ At the conference, Baidu Intelligent Cloud officially launched the first enterprise-level MCP service in China, encouraging developers to fully adopt MCP.
In essence, MCP can be viewed as a ‘universal socket’ for the AI era. It enables large language models to effortlessly access diverse external data sources and tools, achieving ‘one-click interconnection’ with the outside world. This will significantly enhance the development efficiency of AI applications and various AI Agents. In fact, prior to Baidu, several other internet giants, including Alibaba, Tencent, and ByteDance, had already implemented MCP support. Initially a niche term among developers, MCP has transformed into a mainstream concept, becoming a new battleground for various companies. As the MCP ecosystem continues to develop and thrive, AI agent applications are expected to flourish.
Song Jiaji, Deputy Director of the Guosheng Securities Research Institute and Chief Analyst of the Communications Industry, compares the emergence of MCP to the TCP/IP protocol in the realm of communication, asserting that it will foster the advent of AI-native applications. In the internet era, TCP/IP serves as the fundamental data communication protocol, enabling efficient data transmission and seamless connectivity between different devices. Similarly, in the AI era, MCP plays a comparable role, providing a ‘one-click interconnection’ for large language models to access external data and tools.
According to a senior AI professional, MCP is essentially a technical protocol, a set of collectively agreed-upon specifications for the development of AI Agents, akin to the standardization of writing and transportation during the Qin Dynasty. With a unified standard and specification, collaborative efficiency will be significantly improved. MCP is not a recent invention; it was initially released by Anthropic, a well-known US-based large language model startup, in November of last year, with the goal of reducing the cost of using external data and tools for large language models.
While MCP initially received a muted response, the emergence of Manus, a domestically developed general-purpose AI agent, in February of this year, sparked renewed interest. Manus, capable of autonomously executing complex tasks based on human instructions, from automatically booking tickets and generating travel guides to creating websites, quickly gained popularity for its ability to not only chat and think but also ‘get its hands dirty’ like humans. Although its founder stated that Manus was developed before the release of MCP and therefore did not use the protocol, instead utilizing other coding methods to call multiple tools, Manus nonetheless highlighted the value of intelligent agents and the significance of MCP.
Prior to the introduction of MCP, the cost of accessing external tools for large language models was relatively high. For example, if a user wanted to use a large language model to book flights and hotels and receive email confirmation, the model would need to call the APIs (Application Programming Interfaces) of the airline, hotel, and email applications separately. Each API integration would require writing separate code, documentation, authentication methods, error handling, and maintenance methods, essentially requiring different ‘keys’ to unlock these services. However, with MCP, only the MCP servers of the airline, hotel, and email services need to be connected or configured, similar to plugging a USB drive containing the airline, hotel, and email services into the user’s computer via a Type-C port.
The advantage of a unified standard is that it reduces redundant development and construction, avoids repetitive coding, and thus significantly improves development efficiency and reduces development costs. As long as it complies with and supports the MCP standard, all tools can achieve ‘plug-and-play’ functionality, allowing developers to rapidly build more powerful AI applications. Galaxy Securities’ research report points out that MCP is expected to drive the upgrade of AI agent applications from simple information consultation and recommendation capabilities to execution capabilities, fostering the construction of a richer and more complex application ecosystem for AI agents.
The Full-Scale Entry of Internet Giants
The year 2025 is being hailed as the ‘Year Zero’ for AI agents. As a standard protocol, MCP can significantly address the issues of high technical costs and low efficiency in calling external tools during intelligent agent development, making it a new focal point for internet giants.
On March 21st, Baidu Maps announced that its core APIs are fully compatible with MCP, becoming the first map service provider in China to do so. On April 9th, Alibaba Cloud’s Bailian platform launched the industry’s first full-lifecycle MCP service. On April 14th, Tencent Cloud announced that its large language model knowledge engine has been upgraded to support the MCP protocol. On April 18th, ByteDance’s AI application development platform, Kouzi Space, began internal testing, with the platform integrating an MCP extension system. The initial phase of the internal test supports the integration of high-frequency components such as Feishu multi-dimensional tables, Gaode Maps, and image tools.
Xu Zhiyuan, Senior Product Expert at Alibaba Cloud Bailian, stated that Alibaba Cloud is a leading large language model manufacturer in China, possessing the full-stack self-developed Tongyi Qianwen model, and is also the No. 1 cloud service provider in China, making it a necessary condition for successful Agent+MCP implementation. Strong model capabilities ensure support for deep reasoning and the scheduling of complex tasks and tools, while ample cloud computing resources ensure that MCP services are stable, available, and efficient.
Specifically, the Alibaba Cloud Bailian platform integrates Alibaba Cloud Function Compute, over 200 industry-leading large language models, and nearly 100 mainstream MCP services, comprehensively addressing the computing resources, large language model resources, and application toolchains required for intelligent agent development. This eliminates the need for users to manage resources, develop deployments, and perform engineering operations, significantly lowering the barrier to entry for Agent development. For example, a user built an intelligent agent on the Bailian platform using the Bocha MCP service and the Tongyi Qianwen large language model, which can efficiently query massive amounts of data and quickly generate visual charts. The entire process was very convenient, taking only a few minutes to complete the development.
Bocha, as mentioned by Xu Zhiyuan, is an AI-based search engine that supports the online search function of large language models such as DeepSeek. Alibaba Cloud Bailian has currently deployed the Bocha MCP service, and online search is a fundamental tool that many intelligent agents must call during task execution. This tool will avoid a large amount of repetitive coding work.
Furthermore, the full support of MCP by internet giants, with their extensive business lines and application ecosystem systems, provides intelligent agents with a wealth of callable tools. For example, Alipay launched China’s first ‘Payment MCP Server’ on April 15th, providing native payment capability support for AI intelligent agents. Industry analysts say that with Alipay’s MCP service, developers can greatly shorten the process of developing payment links for various service applications. Within the intelligent agent, they can easily use Alipay to complete a series of closed-loop operations such as querying, transacting, and refunding, thereby opening up the ‘last mile’ of the commercial closed loop.
The aforementioned senior AI professional stated that with the addition of MCP, the number of lines of code required to build an intelligent agent with the same functionality has been reduced from more than 3,000 to less than 500, bringing a qualitative leap in the development efficiency of intelligent agents. Xu Zhiyuan revealed that within a week of launching their MCP service, the number of activated users has exceeded 10,000, and they are building intelligent agents for different scenarios based on the MCP service. Many Alibaba Cloud customers and partners have also joined the MCP ecosystem. Recently, the Alibaba Cloud Bailian platform has launched dozens of cloud-based MCP services, including Baiwang Finance and Taxation, Feichangzhun, Bocha Search, and Yingmi Fund, with more service providers gradually joining the platform. In the future, it will further enrich the ecosystem supply and accelerate the application of AI.
Still in a Rapidly Evolving Period
There is a general consensus in the industry that MCP provides a standardized method for AI models to communicate with different data sources and tools, and is the ‘key’ to accelerating the application of large language models. With the entry and layout of major internet companies, the ecological boundary of MCP will also be further expanded. However, the development of intelligent agents is still in its early stages, and correspondingly, MCP has not yet been fixed and finalized, but is in a process of rapid evolution.
Prior to MCP, OpenAI proposed Function Calling in June 2023 to help developers integrate large language models with external functions or tools. Function Calling is a very good design that has been regarded as the standard by the industry since its birth. However, the only problem is that the amount of work required to write external functions is too large. With the development of technology, the complexity of intelligent agents is increasing, and the difficulty of development is increasing exponentially. The advantage of MCP is that it unifies the original differentiated Function Calling standards of various large language models, forming a common protocol.
Following MCP, Google Cloud announced the open source of the first standard intelligent agent interaction protocol, Agent2Agent Protocol (A2A), in early April, aiming to break down the barriers between current intelligent agents and achieve mutual communication and collaboration between intelligent agents built by different manufacturers and different frameworks. For a time, the saying ‘MCP is outdated’ appeared in the developer community, with some believing that MCP may only be a transitional technology and will be a fleeting phenomenon for a period of time.
In response, a research report by Guosheng Securities believes that the battle for Agent communication protocols has not ended. Although A2A and MCP have different purposes, the former is for communication between Agents, while the latter is for the interconnection between Agents and external tools and data. However, in the complex situation where ‘tools may also be packaged as Agents,’ the functions of the two must have some overlap, but this competition helps to reduce the cost of large language models calling external tools and communication.
Whether it is MCP or A2A, the protocol itself does not have absolute uniqueness. It provides a more standard connection method for the model, activates supply, and reduces the difficulty of connecting the model with various real services. Ultimately, it is to release AI productivity and accelerate the explosion of applications. MCP is a natural product of the entire large language model development stage. Even if there is no MCP today, there are other protocols to achieve this step.
Xu Zhiyuan further pointed out that MCP currently has many problems, such as unified authentication, security protection, stable long connections, and multi-tenant management. From a personal point of view, the current problems are not terrible, but reflect the real needs that exist in developers and actual business implementation. Recently, we have also seen that the protocol evolution of MCP is continuing. As an open source protocol, it will continue to iterate and improve with the development of technology and the ecosystem, and will gradually reach a relatively stable state in the future.