Let's talk briefly about MCP and Computer Use. Just understand them.

Due to the popularity of DeepSeek, everyone has been paying close attention to the development of AI this year. Recently, Manus has continued to attract attention, and the concept of Agent has become popular again.

As a result, some low-frequency nouns have re-entered the field of vision, and even been highly praised, such as computer use and MCP . Let’s talk about MCP first...

unsetunsetWhat is MCPunsetunset

Model Context Protocol (MCP) is an open standard protocol developed by Anthropic that enables AI models to securely access and interact with external data and tools.

Similar to the http protocol, it is an agreed upon standard that everyone can follow.

The reason why MCP appeared is very simple. To truly solve the problem, the large model must interact with various external interfaces, including browsers, databases, file systems...

Before MCP was proposed, how did we need external information? The answer is customization.

Write an intermediate program to directly call the big model (get request). After receiving the big model request, call the API to read and write the database, or make API calls based on some parameters returned by the big model. In short, what the user accesses is actually the intermediate program, which completes the bridging of the gap for the expansion of the big model's capabilities .

But one person was unhappy: he felt that such a weird thing as an intermediate program should not exist , so he first implemented a fixed format API call at the bottom layer of the model , so that subsequent users could directly access the large model, and the large model could automatically call the API to read and write the database .

Later, demands for file reading and writing, browser reading and writing, etc. arose. In order to improve efficiency, the previous standards were adopted. Finally, it was found to be very useful, so a protocol was formed.

With MCP, there is a unified standard for the interaction between AI and the outside world, enabling AI applications to seamlessly integrate external data sources and tools.

Before MCP, each developer needed to create their own way for AI to interact with the outside world, which led to a large number of incompatible systems and security vulnerabilities.

The above is a very rough description, you can understand it yourself, there is nothing much to say...

unsetunsetAdd a sentenceunsetunset

To deepen your understanding, I will just add some content here. Actually, there is nothing much to say about MCP. Just write it according to the requirements...

User → Intermediary → Large Model → Intermediary → API → Return Result

Actual technology evolution:

# Traditional adaptation layer pseudo code
user_input =  "Check Beijing weather"  
model_response = llm(user_input)   # The model may output "need to call weather API, city = Beijing"
if  "Weather API"  in  model_response:
    city ​​= extract_city(model_response)   # Developers write their own parsing logic
    result = call_weather_api(city)       # Manually call the API

MCP stage:

Rather than completely eliminating the middleman, the process is standardized into a protocol layer. The process becomes:

User → Model → MCP Client (Structured Request) → MCP Server (Protocol Conversion) → API → Return Result

- **Key upgrade**: The model directly outputs standardized instructions (such as JSON format), and the MCP protocol layer replaces customized code

---

### 2. Core differences: protocol layer vs model capabilities
| **Dimension** | **Traditional Mode** | **MCP Mode** |
|----------------|-----------------------------|---------------------------------|
| **Call initiator** | Developer code triggers API call | Model autonomously generates MCP instructions |
| **Interface Specification** | Each API needs to be connected independently | Follow the MCP protocol format |
| **Security Control** | Rely on developers to implement permission management | Built-in sandbox and permission strategy at the protocol layer |

**Typical case comparison**:  
- **No MCP**: The model outputs  "Please call the weather API to check Beijing" , and the developer needs to write a regular expression to extract the parameters  
- **With MCP**: The model directly outputs structured instructions:
```json
{
 "action" :  "query_weather" ,
 "params" : { "location" :  "Beijing" },
 "auth_scope" :  "user_weather"
}

unsetunsetWhat is computer useunsetunset

Computer Use was also launched by Anthropic (Claude's parent company) in October 2024.

The goal is to allow AI to operate computers like humans: look at the screen, move the cursor, click buttons, and type, and then expand functions on this basis, such as helping to book air tickets, fill out forms, and check the weather.

In fact, RPA has been doing this automation work before and has done a good job.

Computer Use can be understood as RPA’s expansion capabilities and the behavior of operating computers, but the underlying logic is completely different.

Computer Use is to list an action plan after accepting a task, and think, plan and operate the next step based on the real-time screen changes. He can actively explore and try and error unknown tasks;

RPA is more of a fixed process operation, which is a huge difference. In the long run, Computer Use may make RPA companies feel desperate...

The technology is implemented as follows: simulating human operation through screenshot + virtual keyboard and mouse, similar to "teaching interns to use computers".

It is still in the testing phase, with slow operation speed and high error rate.

The meaning is also very simple: without tools such as RPA, the big model can use the computer by itself ...

Just like MCP, there are too many gimmicks for the time being. Just keep paying attention. When it really matures, there will be very complete documentation. There is no need to worry at all.