The fatal flaw of the MCP protocol for large AI models

In recent days, there have been a lot of articles about the MCP protocol in my circle of friends, so I won’t go into the details here.

Simply put, how can we summarize in one sentence the role and significance of the MCP protocol for AI agents and AI big models?

To give an example, it is like an AI that can only brag, but suddenly it has 1,000 "little brothers" who can actually do work, and the MCP protocol is a way to find these "little brothers" .

For example, if you ask a large AI model like DeepSeek to write an ancient poem, it will be fine. But if you ask it to order takeout, sorry, it can't call Meituan, and Meituan has no way to provide an interface for AI to order takeout.

The MCP protocol is like establishing a communication mechanism between all large AI models and Meituan, so that you can ask DeepSeek to order takeout for you.

I have been in contact with the MCP protocol for some time. At first, I thought it was pretty good and had a bright future. But when I used it myself, I found a fatal flaw -  it can only run on the local computer .

Currently, MCP only supports stdio and SSE transports.

This means that you either need to download the code locally or run the service locally and provide an SSE interface. It does not support the common REST API method at all, and the official hosting URL is not provided. Although I think the official design from the perspective of the open source community is a very long-term thinking and design philosophy, but in this way, for the majority of novice users, the MCP protocol and MCP service are no different from a pile of code that cannot be run.

For example, I want to use the MCP protocol, open the Cursor and prepare to add the MCP Server, but there are only two options.

(Cursor adds MCP service, there are two ways: SSE and Command. Do you know what it means?)

This is confusing. You must be wondering: Do I need to fill in a URL? But I can’t find a place to fill it in. What should I do? Do I need to run the MCP server code? What is Git? What is the command line? What is npm? What is node js? (10 more are omitted)

For novice developers, it is extremely difficult for ordinary people to execute a simple bash command, let alone get a Node JS service up and running within an hour.

Although starting a nodejs service is not too difficult, if they are asked to do production deployment, they need to configure a managed Kubernetes cluster and learn Docker container deployment, which is frighteningly costly and greatly increases the difficulty.

You know, in today's cloud era, what stumps most people may not be the software itself, but the software's operation and maintenance tools.

It’s like everyone knows that cooking is fun, and can even be considered an art, but most people are too lazy to wash and cut vegetables, thinking that these foreplay preparations are too troublesome.

The starting point of the MCP agreement is actually quite good.

You know, in the software era, the connection and docking between software can cause a lot of trouble for many programmers and product managers.

If the data is not connected, it will give rise to a lot of troublesome solutions, and the connection will be even more troublesome. This is a big problem.

In addition to the transport network solution, the MCP protocol has some other problems, such as user authentication, network performance, and security.

However, I also see that its ecology and related tools are developing very rapidly. This is the key. Only when the ecology develops well can it be really good.

I will continue to pay attention to alternatives to MCP. If you have any ideas, please leave a comment in the comment section.

Although this is quite challenging, where there are problems there are opportunities, and we look forward to more alternatives to MCP emerging.

After all, the big model is like the brain, and the intelligent agent is like the brain stem. To truly complete a task, it requires countless "hands and feet" and tools of all kinds.