Original video in Chinese.
Key Takeaway
- MCP (Model Control Protocol) is key to improving AI productivity, and this article recommends five practical MCPs.
- Tavily + Sequential Thinking: Tavily solves information retrieval, while Sequential Thinking enhances the model’s multi-step reasoning ability, improving answer quality.
- Filesystem + Markitdown: Filesystem lets the model access local files, and Markitdown converts PDFs to Markdown, allowing the model to handle PDF documents.
- Vector database MCPs (Milvus/Pinecone): used to connect personal note information, enabling local or cloud knowledge bases for storage and retrieval.
- All of these MCPs revolve around information, aiming to help the model obtain more information and improve its ability to process information, turning AI from a “brain in a vat” into a more complete tool.
MCP has been hot for so long—what are some that you’d actually use in daily life and that really improve productivity?
In this video, I’ll introduce five MCPs that I basically use every day. They may look plain and unremarkable, with no flashy presentation value, but they’re just useful.
I usually share these MCPs, along with more AI tools, inside the community. If you want to improve your AI productivity, remember to join.
Let’s start with the first two MCPs. They’re a pair.
Tavily is a search engine I’ve used for a long time, optimized specifically for LLMs. Its MCP includes two tools: search and extract. They’re especially practical. However, if you only use this one MCP, then although you can search online, the answers won’t be very good.
For example, I asked a question: Why has Real Madrid performed poorly this season? Please analyze in detail.
Claude used Tavily to find a bunch of web pages and seemed to answer my question. But if you look closely, its so-called detailed analysis—things like “unstable match results” and “poor Champions League performance”—those are symptoms, not causes. That doesn’t even count as a “detailed analysis.”
You see, as a model, Claude 3.7 Sonnet is strong enough, right? And yet the answer it gave was still so weak.
At this point, you need the MCP called Sequential Thinking. Its job is to make the model perform multi-step reasoning. Once you add it, the effect is very obvious.
Using the same question, I asked for a detailed analysis of the reasons behind Real Madrid’s performance this season. I specifically requested the Sequential Thinking and Tavily combo.
It starts the same way: gathering information. But after that, Claude called Sequential Thinking seven times.
If you expand it, you can see that with this MCP, the model’s logic became much stronger. Claude distilled Real Madrid’s problems, including injuries, fatigue, training methods, and physical management. And then look at the final answer—it’s clearly much more reliable than the previous one, right?
This pairing is simple: Tavily handles information, and Sequential Thinking handles logic. These are both very basic, very important capabilities, and you should definitely add them to the model.
OK, let’s look at the next two MCPs, which are also a pair.
Filesystem allows the model to connect to a specified folder, such as the desktop, read files on the desktop, and create new files.
Clients like Cursor, since they’re used for programming, naturally have strong file-reading and editing capabilities. But clients like ChatWise are a bit troublesome. That’s why you need Filesystem. As for security concerns, don’t worry: the folders are all specified by us, so there’s no need to worry about the model running wild.
The MCP that goes with Filesystem is Markitdown. This is a tool Microsoft released before that can convert PDFs into Markdown format. Now that an MCP version exists, the model can finally handle documents like PDFs.
Personally, I hate PDFs with a passion. From a technical perspective, this format is almost like an image, and it’s far too much trouble to work with. In the AI era, PDFs should be completely eliminated.
So how can Filesystem plus Markitdown be used?
For example, I have a PDF document on my desktop, the very famous Bitter Lesson. If I want to convert it to Markdown format, I can just tell the model directly. It will first use Filesystem to confirm whether it can access the desktop. Then it will use Markitdown to convert it. Finally, it will use Filesystem again to create a new file on the desktop and write the extracted content into it.
Or I can simply have the model summarize it for me.
You see, I tell the model to help me read and summarize the PDF document on the desktop. It immediately knows to use Filesystem and Markitdown to read and extract it, and then it can summarize it.
Claude’s context window isn’t very large, so handling a PDF that’s only two pages long is fine. If you run into one that’s a dozen pages or more, it’s better to switch to the Gemini family.
The two MCP pairs I just introduced are used to handle web information and document information, respectively. So how do you give the model your own note information through MCP?
If you use a local vector database like I do, you can choose Milvus. This database supports storage and retrieval. Once it’s deployed, keep OrbStack running in the background all the time. Then no matter which client you use, you can access the data anytime through MCP.
As for Milvus deployment, I released a community-exclusive video two days ago that goes through the key points one by one. If you’ve already joined the community, remember to watch it.
If you want to use a cloud vector database, you can choose Pinecone. It also has an MCP. If you’re even more hands-on, you can use AWS, which is even more powerful. I’ll also release a community-exclusive video introducing that in detail later.
So, those are the five MCPs I use every day. Have you noticed? They’re all centered on information—letting the model obtain more information and improving its ability to process information. With these tools, the model is no longer a brain in a vat.
OK, that’s all for this episode. If you want to learn about AI, want to become a super individual, and want to find like-minded people, come join our newtype community. See you in the next one!