The significance of AI is to liberate human labor and improve the minimum working capacity of most people. However, LLM currently still has great limitations. It requires back-and-forth dialogue to give suggestions, and users must personally execute the suggestions. There is still a gap before we can truly use AI to help us work.
Now, if you can use your computer to reply to emails, write reports, and even automate cryptocurrency trading by talking to AI, isn’t that getting closer to the vision of liberating productivity? This technology is the hot keyword in the field of AI - MCP
What is MCP?
MCP (Model Context Protocol) is a set of "standardized protocols" released by Anthropic in November 2024 to solve the problem that AI models in the past could only "speak" but not "do".
First, let's break down the MCP naming
- Model: Model refers to various large AI language models (such as GPT, Claude, Gemini, etc.)
- Context: context, representing additional information or external tools for the model
- Protocol: a universal, standardized "specification" or "interface"
To sum up, through unified standards, AI can not only "speak", but also directly control external tools to complete various tasks.
Generally, the most commonly used LLMs, such as ChatGPT, Grok, etc., can only perform "text input and text output" according to the content of the conversation. If we want AI to help us perform actual operations, such as reading files from computer folders, sending emails, searching databases, etc., we usually give instructions to LLM first, and the user will then operate by himself according to the response of LLM, and finally report the results to AI, and AI will give us text suggestions, and we continue to operate, and so on.
The emergence of MCP allows AI to not only read local files on the computer, connect to a remote database, but also directly operate specific network services. In other words, AI is no longer just outputting text, but can complete many repetitive or procedural tasks for you.
How it works
- MCP Host (Administrator): Responsible for managing and coordinating the operation of the entire MCP. For example, Claude Desktop is a type of Host that can help AI access your local data or tools.
- MCP Client (user end): receives user needs and communicates with LLM (AI model). Common examples include various chat interfaces or IDEs that integrate MCP (such as Goose, Cursor, Claude Chatbot).
- MCP Server (server): can be regarded as a set of "organized and annotated" APIs that provide functions that AI can use, such as reading databases, sending emails, managing files, calling external services, etc.
With MCP, AI can not only understand human language, but also directly translate specific text into action commands to complete automated operations. For example, it can help you organize sales reports, send emails to customers, and even perform 3D modeling directly on Blender through commands.
Reference: https://www.youtube.com/watch?v=FDRb03XPiRo&t=4s
Why is MCP important?
1. Build a bridge between AI and external tools
The limitation of LLM is that the data in it is pre-trained and not updated in real time, which means that the data of LLM is limited to the information it sees during training. Therefore, the model is unaware of new information generated after training.
Assuming that the LLM training was conducted in February this year, there is absolutely no information after February this year.
The current mainstream method is to use RAG (Retrieval-Augmented Generation), which is a way to combine "retrieval system" with "generation model". This architecture can retrieve the latest data before LLM reasoning, and provide the retrieval results as context to the model. Specifically:
- Retrieval: Before answering questions in LLM, use a search tool (such as Internet search, internal database query, etc.) to find the latest information that matches the current question.
- Generation: The retrieved information will be passed to LLM as auxiliary information (Context) to help it generate more accurate and timely answers.
For example, before answering a question, AI first searches for the latest information through Bing or Google, and then integrates the search results into the response. This is the way of using RAG.
The biggest difference between MCP and RAG is:
- RAG uses relatively static data to assist LLM in answering questions, while MCP allows AI to actually “do it”, such as searching a database, calling an API, or even modifying file contents.
2. Standardization & universality: Just like USB-C: Different manufacturers can develop their own functions that comply with the MCP specification, just like all devices can use the same USB-C transmission line. Without MCP, each developer has to define how to make AI call a specific API. This means that the same work will be repeatedly developed by different people. After the MCP is unified, everyone can integrate immediately by implementing the same set of specifications, avoiding the phenomenon of reinventing the wheel.
3. From passive response to active execution: Traditional AI tools can only answer questions but cannot really take action. With MCP, AI can decide what instructions to execute based on the current situation and take the next step by reading the feedback results. This ability to make continuous corrections based on the situation greatly enhances the practicality of AI.
4. Security and control: MCP does not force all data to be transmitted to the AI model. It can control data access through permissions, API key management, etc. to ensure that confidential information will not be leaked.
What is the difference between MCP and AI Agent?
What is AI Agent?
In Q3 last year, GOAT led the AI Agent trend. Most crypto users understand AI Agent from the perspective of Web 3. AI Agent usually refers to an AI system that can "automatically" handle specific tasks. It not only talks to people, but also takes actions based on the context, calls tools or APIs, and completes a series of steps. For example, the most common ability to post on Twitter autonomously also falls into the category of AI Agent.
Limitations of AI Agents
- Lack of standardization: Everyone can create an agent, but if there is no unified specification, there will be problems such as "this agent only uses the model of manufacturer A" and "that agent only calls the API of system B".
- Easy to operate independently: Although AI Agents can run errands, developers often need to customize a large number of API formats and rules. There is a lack of a shared ecosystem between different Agents, which makes integration difficult.
The relationship between MCP and AI Agent: MCP is a protocol, AI Agent is a concept or execution method
- AI Agent emphasizes AI's ability to take initiative and execute tools
- MCP focuses on how to enable different AI models to communicate with external tools, playing the role of a universal standard.
MCP helps AI Agents operate more efficiently
- Without MCP, AI Agent may have to write a set of API rules for different tools and platforms, which makes development and maintenance very troublesome.
- With MCP, AI Agent only needs to follow the MCP specification, obtain available tools from the "Server List", and then dynamically decide which tool to use to complete the task. Access to external resources is also safer and more convenient.
Different functional scope
- AI Agent: focuses on decision-making and logic, and determines what to do and what steps to perform based on requirements.
- MCP: Specializes in solving tool interface and standard formats, and how to provide external services, databases, and archive systems to AI in a unified way
The combination of the two: AI Agent + MCP = allows AI to understand both how to act and where to act.
What are the MCP concept projects in the current cryptocurrency circle?
1. Basic MCP
The framework officially developed by Base was launched on March 14, allowing AI applications to interact with the Base blockchain. Users can deploy contracts to the blockchain and use Morpho for lending and other functions through natural language conversations without the need for development capabilities.
BORK is the first token to use Base MCP deployment. It was issued on March 14 and its market value peaked at $4.6 million. However, it has now fallen back to $110,000 and its 24-hour trading volume is only $90,000. It can be judged that the life of the coin has ended.
Flock is a decentralized AI training platform. He pointed out that the current MCP still runs on external AI models and is processed by centralized LLM. Flock provides a Web3 proxy model, and AI-driven blockchain tasks can run locally, providing users with more control.
2. Lyra
LYRAOS, whose full name is LYRA MCP-OS, is also a multi-AI Aigent operating system that allows AI Agents to interact directly with the Solana blockchain and perform operations such as buying and selling cryptocurrencies.
They are currently exploring how to use MCP-OS to establish thousands of "AI16ZDAOs", AI-driven decentralized autonomous organizations for cryptocurrency investment. LYRAIOS plans to release a DEMO between March 21 and 22, 2025, and launch the official product next week.
The current token market value is 923,000, the highest is 2.64 million, the 24-hour trading volume is 3 million, and the number of currency holding addresses is 2,922
Conclusion: AI narrative dances again, but it will take time to observe
Although MCP provides a standardized rule that allows AI to interact with external tools more easily and safely, and seems to have great potential in the field of Web 3, the success cases are relatively limited. The reasons behind this may include the following:
Technology integration is not yet mature: In the Web 3 ecosystem, the contract logic and data structure of each chain and each DApp are different. It still requires a lot of development resources to encapsulate them into an MCP Server that can be called by AI.
Security and regulatory risks: Allowing AI to directly manipulate contracts and process financial transactions requires the design of a comprehensive private key management and permission control mechanism, which is difficult and costly.
User habits and experience: Most people are still skeptical about letting AI manage wallets or make investment decisions, and the operational threshold of blockchain itself is also high. If the experience is too complicated or lacks clear application scenarios, it will be difficult for novices to use or invest in it for a long time.
Aesthetic fatigue and market indifference: Previously, AI Agent set off a trend in the cryptocurrency circle. It was common for many unimplemented projects to have valuations of over 100 million yuan at their peak. However, it is facing the stage of bursting the AI bubble recently. Most projects have fallen by more than 90%, which is regarded as a fear of AI.
Back to the MCP narrative, it can be understood as a super-enhanced version of AI Agent. The market has experienced the crypto AI craze before, and has gradually understood what is concept hype and actual application. If there is no application with real innovation and practical value, investors and users will not easily pay for it. Pre-existing MCP projects like BORK did not attract much attention in the end because there was no obvious differentiation or application landing. This is also the most important key factor that the author believes that the current MCP concept has not yet become popular.
The combination of MCP and blockchain has potential, but it also faces the dual challenges of technical barriers and market pressure. In the future, if more mature security mechanisms can be integrated, more intuitive user experience can be created, and innovative applications that truly bring value can be discovered, "Web 3 + MCP" may be able to escape the fate of being a "hype topic" and become a new round of main narrative.
