newtype.
ArticleOctober 11, 2025Free to read

MCP Is So Simple, Anyone Can Do It

There are two ways to configure MCP: AI auto-creation, like Cline’s Marketplace, and manual editing, like Cursor’s config file.

Originally published . English translation: . Read the Chinese original.

Original video in Chinese.

Key Takeaway

  • MCP (Model Control Protocol) is compared to AI’s USB-C, aiming to unify the interface between AI and various software so AI can call tools on demand.
  • MCP’s rapid development may have been influenced by the Agent concept, because it gives AI “hands and feet,” and is currently the best way to unlock Agent capabilities.
  • Compared with general-purpose Agents and complex workflow building, MCP is easier to configure, and models like Claude-3.7 Sonnet can independently choose and call tools.
  • There are two ways to configure MCP: AI auto-creation, like Cline’s Marketplace, and manual editing, like Cursor’s config file.
  • Through the cursorrules document, users can customize Cursor’s behavior so that when handling tasks it prioritizes retrieving local documents, searching online, and calling specific MCPs.
  • MCP server communication is divided into local stdio and cloud SSE; even a locally running MCP can access the internet by calling remote APIs.
  • Mastering MCP can significantly boost AI productivity, because when a model has more tools, its capabilities become stronger.

Full Content

My AI is stronger than yours not because it’s smarter, but because it has more tools in its hands.

For example, your Cursor can only code, while mine can use Blender to do 3D modeling.

Your Cursor can only code, while mine can rip an English webpage, translate it into Chinese, and save it to local documents.

All of this is possible because of MCP. I introduced this extremely, extremely hot protocol in the video before last. If you haven’t watched it yet, hurry up and watch it—it’s very important!

Simply put, MCP is AI’s USB-C. No matter what software you are, you all have to use this unified interface protocol. That way, AI can freely connect to various software and call tools as needed.

Just like USB-C, computers, phones, keyboards, mice, and so on all support it. One cable can both charge and transmit data, which is very convenient.

MCP has been around for a while. Over the past month, I suddenly felt it speeding up in development. Maybe it’s been influenced by Agent. Everyone realized that if you want to do Agent, you can’t do it without tools—without tools, AI has no hands or feet. After looking around, this seems to be the most reliable option. So more and more people are supporting it. Quite a few developers are already working on infrastructure-type projects. This field will definitely produce dark horses, so the VC folks should keep a close eye on it.

For users, MCP is the best way to unlock Agent at the current stage.

First, a general-purpose Agent like Manus isn’t yet at the stage where it can be deployed at scale. It still needs some time.

Second, the way platforms like Dify build vertical Agents and workflows actually isn’t suitable for ordinary users. It says “no code,” as if you can just drag things around on a canvas and be done. But if you really have no programming mindset or experience, you definitely won’t be able to handle it.

By comparison, the MCP route is much more reliable.

First, models like Claude-3.7 Sonnet are already very powerful. If you configure all kinds of MCPs, it knows on its own when to use what, so you don’t have to worry about it.

Second, the configuration method for MCP is simple enough. Right now there are only two methods: either AI creates it automatically, or you edit it manually.

The first method mainly applies to Cline. It has built a Marketplace internally and pulled in the mainstream MCPs. You just click Install, and it handles everything else. If a bug comes up during configuration, it will also figure out how to solve it on its own. That’s really thoughtful!

The second method is also simple. I’ll demonstrate with Cursor.

Before officially starting, you need to go into Beta and change Standard to Early Access. Then click Check for Updates in the upper-left corner. That will update Cursor to version 0.47.

Compared with version 0.46, version 0.47 adds config file support for MCP. We can make changes directly in it, which is very convenient.

For example, I want to add the File System MCP. Its purpose is to let the model operate on files under a specified path, such as reading and writing, creating and deleting, and so on. All of these functions are clearly listed under the tool list.

Remember, each MCP has several tools. The model will decide which MCP and which tool to use based on your request. So when you choose an MCP, make sure to check its tool list, and then you’ll know what’s what.

OK, let’s keep configuring. Scroll down to the NPX section. The lines in the middle are what we need to copy. Let me explain—it’s very easy to understand:

npx is a command-line tool that can execute commands in npm packages. The arguments below means parameters. It contains three parameters:

y means yes, which skips the confirmation prompt;

@modelcontextprotocal is the name of the npm package that needs to be executed;

The two lines below are both address parameters, telling the MCP where it can access files.

So what we need to do next is copy these lines. Then open the config file in Cursor and paste them in. Finally, change the address—for example, I specified that it can access files on the desktop.

After saving, go back to the MCP server page, and you’ll see that the filesystem MCP has lit up green, meaning it’s been configured successfully. Every tool it includes is also listed.

If you understood what I just demonstrated, then the other MCPs are basically the same operation. The biggest difference is that the parameters and environment settings are different.

For example, the Firecrawl crawler MCP requires us to fill in an API key. Everything else can be ignored—just copy and paste.

Once you understand Cursor’s settings, the same applies to other software. For example, Cline and Claude are operated the same way. You just open that config file, take one look, and everything becomes clear.

See, that’s the powerful thing about MCP. It doesn’t start by acting all high and mighty and trying to overturn everything. Instead, it tries its best to reduce costs for everyone involved. That’s why people are motivated to support it and complete interface unification.

If you’re a developer, you absolutely need to seize this MCP opportunity. If you’re a user, you should get hands-on and use it yourself. Remember what I said:

When the model is the same, whoever can call more tools has higher productivity.

OK, that’s it for this episode. If you want to talk about AI, come to our newtype community. See you next time!