Original video in Chinese.
Key Takeaway
- MCP (Model Control Protocol) is key to AI’s development. It evolves AI from a “feature phone” into a “smartphone,” solving the problems of inconsistent traditional API interfaces and a lack of semantic understanding.
- MCP gives AI a unified tool interface, allowing it to freely access and control massive numbers of tools and carry out more advanced tasks.
- This article demonstrates how an AI tool (Cline) uses MCP to call multiple tools such as Obsidian (notes) and Tavily (search), showing MCP’s real-world application.
- MCP passes the user’s query, tool descriptions, and parameters to the large language model in a structured way, and the large language model decides how to handle them.
- Claude Desktop and Cline are currently the best tools in terms of MCP support, and users can also develop MCP tools themselves.
If you care about AI and care about agents, then you absolutely need to pay attention to MCP, really understand it, and use it a lot.
There are two very clear paths in AI’s development, and I’ve said this in the community before: one is mastering more information, and the other is controlling more tools.
Two years ago, a large language model’s sources of information were only the data it was trained on, and the things we told it during inference.
Later, we added web search so it could access broader and more up-to-date information; we added RAG and knowledge bases so it could access those non-public pieces of information in vertical domains.
That still wasn’t enough, because all of that was just text information. So then we added multimodality, and now it can even understand images.
You can see that all of these developments follow the first path: mastering more information.
Information alone isn’t enough. To change the world, you also need tools in your hands. That is the second path. I believe MCP is the key to the second path.
With MCP, the AI in your hands evolves from a feature phone into a smartphone.
As always, I’ll demonstrate first, then explain.
On the left side of the screen is Obsidian, a note-taking app. For the demo, I created a new vault and put in three previously written video scripts as notes.
On the right side of the screen is Cline. It’s an AI plugin with very strong programming capabilities. If your main setup is VS Code, you absolutely need to use it.
What I want to show everyone is that through Obsidian MCP—you can simply understand it as an interface—AI tools like Cline can directly read and modify notes in Obsidian.
Let’s start with the simplest example. I ask Cline: what notes are in the vault?
After receiving the request, Cline looks around the environment and finds that Obsidian is connected to an MCP server. So it decides to use the MCP server and call the list notes tool to query what notes are available.
In the RESPONSE, the MCP server returns the queried information to Cline, and then Cline gives its final answer.
Using the same principle as querying what notes are in the vault, Cline can also read the specific contents of notes through the MCP server.
I ask it: what is the main content of this note?
At this point, Cline uses the read note tool to read the note. After extracting all of the more than 1,000 characters, it summarizes the content and then gives the final answer.
If you remember, you should recall that I previously made two videos, both about Cursor modifying Obsidian notes. The prerequisite for that approach was that Cursor had access to the locally stored Obsidian vault.
But not every AI tool has the conditions or ability to access an entire note vault. That’s when an MCP server is needed, serving as a universal interface connecting both sides.
What I just demonstrated only connected to one MCP server. In fact, to complete a task, AI can connect to multiple MCP servers and call multiple tools. Let me show another example.
This is an article on OpenAI’s official website about GPT-4.5. I gave the link to Cline and wanted it to scrape the content of the article.
At this point, Cline found that there was also a Tavily MCP it could connect to. Tavily is a search engine specifically optimized for large language models. Its MCP includes a tool for extracting webpage content.
Using this tool, Cline successfully scraped the webpage content. Then it reorganized and summarized all the information, ultimately giving a very clear answer.
Next, I gave a second instruction: take this summary and save it as a new note in Obsidian.
You can see that it called the create note tool in Obsidian MCP and began creating the note. Along the way, it ran into a formatting issue, mainly a parsing error caused by special characters. I guess it was a Chinese punctuation issue. Fortunately, Cline was smart enough to proactively avoid special characters on the second attempt, and in the end it completed the note creation.
From the demo just now, you can see that AI applications like Cline and Claude can call all kinds of tools and complete more advanced tasks. That is the value of MCP.
To put it simply: MCP is the USB-C port for AI.
Before MCP came along, how did AI call tools? Through APIs, right? But APIs have major limitations and are not suitable for AI.
First, the APIs of different tools each have their own specifications.
For example, if I want to call a weather forecast tool, I have to tell it the city name and date in its required format, such as the word “Beijing” and “0315” (that is, March 15). If I want to call a calculator tool, I have to tell it the numbers and symbols in its required format, such as “1”, “+”, “2”.
There are massive numbers of tools on the market. It’s unrealistic for AI to know and satisfy every single tool’s requirements. It’s like some devices use USB-A, some use micro USB, and some use mini USB. As a user, seeing all these interfaces would make your head spin!
So AI needs a unified standard to unify all tool interfaces. It’s just like USB-C today: phones, computers, everything uses this port. And now new monitors even support connecting to a MacBook via USB-C, handling both charging and signal transmission, so you don’t even need HDMI anymore. One cable to rule them all is the trend.
Second, APIs focus on transferring data, not on the meaning of the data.
Like in the example just now, if I ask: how’s the weather in Beijing next Monday? Or, what is one plus two? Data like this can’t be accepted by APIs, because they don’t understand, and don’t need to understand, what it means. Their job is just to take the data and pass it along.
But MCP can understand it.
Actually, it’s not that MCP itself understands it. Rather, it packages everything up and gives it to the large language model to understand. MCP passes the user’s query, the tool’s description, and the parameters to the large language model in a structured way, and the large language model decides how to handle them.
At present, the best tools on the market for MCP support are Claude Desktop—after all, MCP was proposed by them—and Cline.
In the two MCP examples I just showed, the Tavily one was ready-made and could be found in the Cline Marketplace. Then Cline would download the code repository and deploy it automatically, which is very convenient.
The Obsidian one, on the other hand, was generated automatically by Cline for me. During the process, I didn’t write a single line of code—I just kept clicking “Agree,” and it was done in a few minutes. If I need to add or modify a tool, I can also leave that to Cline.
MCP itself isn’t complicated. With the help of AI coding tools, you can also develop your own MCP tools and then deploy them locally or in the cloud, such as on Cloudflare. The most important index document can be fully understood with the help of AI.
Or you can just use ready-made ones. There are already several MCP navigation websites out there, such as Smithery.
On the commercialization side, I’ve also seen the first paid MCP product on the market, which can help design UI.
MCP is gaining recognition from more and more people. Everything is moving fast. Like I said earlier, if AI could only call limited, preconfigured tools, wouldn’t that just be the old feature phone? With MCP, AI can freely access massive numbers of tools. From this moment on, AI is finally moving from the feature phone era into the smartphone era.
OK, that’s it for this episode. If you want to understand AI, come to our newtype community. See you in the next one!