Clawdbot ist furchtbar beliebt. Ich sehe den himmelhohen Weg der Tencent-Yuanbao.

There is a project on GitHub that has risen by 9,000 stars within a day, from 7,900 in the morning to over 17,000 in the evening. By the time of publication, this project has already reached over 40,000 stars.

Such a growth rate is extremely rare in the world of open - source projects. Since this project can run on the Apple M4 chip, it has even triggered a buying spree for Mac minis.

And this project is called Clawdbot.

On weekends, everyone in the Silicon Valley AI scene was fascinated by Clawdbot.

Andrej Karpathy, the former head of Tesla AI, publicly praised it. Logan Kilpatrick, the head of Google AI products, followed suit and bought a Mac mini. Federico Viticci of MacStories consumed 180 million Anthropic API tokens within a week just to try out Clawdbot.

When I saw Clawdbot's fame, I suddenly thought: Isn't this exactly the chance that Yuanbao can replicate in the WeChat ecosystem?

But before we discuss Yuanbao's possibilities, let's first learn what Clawdbot is and why it has convinced so many AI experts in just a few days.

01

What is Clawdbot?

In short, Clawdbot is an open - source AI assistant that can be self - hosted.

The core of its innovation is that you don't have to operate it through a special website or app. You can simply send it messages via daily communication software (WhatsApp, Telegram, iMessage, Discord, Slack, etc.) and then let it handle various tasks.

It's as if you have an employee who works around the clock and sits at your computer. If you want to watch TV after work in the evening, you can simply tell Clawdbot what you want, and it will open the corresponding page. Then when you come home and sit at the computer, you can start watching TV directly.

Clawdbot's design concept is that it receives commands via a messaging application and forwards them to a large language model running in the background. This model converts the request into a local shell script that is executed on your computer.

In other words: It does the task for you instead of just telling you how to do it.

Its workflow is based on the architecture "Messaging application ⇄ Gateway ⇄ AI model + Tools". The gateway, which runs on your own device, acts as a coordination center and is responsible for message forwarding, state management, time control, and service integration.

Here comes the problem: To fully leverage the AI potential, Clawdbot requires high access permissions (reading and writing files, executing terminal commands). However, for security reasons, these must be limited.

Clawdbot's solution is the concept of "session isolation".

It distinguishes two types of usage scenarios: the "main session" when you chat with it privately, and the "non - main session", for example, in group chats or public channels.

In the main session, Clawdbot has full system permissions. It can read and write files, execute Bash commands, control the browser, and manage the calendar. Since only you communicate with it, the risk is manageable.

In the non - main session, Clawdbot automatically switches to the "sandbox mode". All operations are executed in a Docker container.

It's like a "training ground". In this isolated environment, it can still execute commands but has no access to the real data on your computer. Even if someone in a group chat tries to attack your system via a malicious command, Clawdbot will only execute the command in the sandbox without affecting the real environment.

Another technical problem is the coordination between the gateway and the tools.

The gateway is the control center on your device. It receives messages from different chat apps, manages session states, controls the AI model, and schedules time tasks. The tools are the various capabilities that Clawdbot can use.

The cooperation between these two components is not easy. When you send a message to Clawdbot in a chat app, the gateway must perform the following steps:

1. Identify which session the message comes from (main session or group chat)

2. Decide at what access level the message should be processed (full permissions or sandbox mode)

3. Forward the message to the AI model for processing

4. The AI model returns the required tools and parameters

5. The gateway executes these tools

6. The result is sent back to you.

During the entire process, the gateway must manage multiple sessions simultaneously. Each session can have different access permissions. In addition, it must ensure that the messages are processed in the correct order and that no states are lost.

It's like a traffic policeman who coordinates traffic from different directions to avoid collisions and ensure efficiency.

To solve this problem, Clawdbot has developed a "session model". Each session has its own state, history, and access permissions.

Although AI inference still relies on the APIs of Anthropic or OpenAI, the routing logic and automation are completely local. This ensures privacy and controllability. This architecture allows Clawdbot to leverage the strengths of cloud models while maintaining the principle of local priority.

The model's memory is limited. However, since Clawdbot is designed for continuous work, it stores all interaction records in Markdown format in a local folder. This enables permanent memory, similar to Obsidian's knowledge base.

It remembers your preferences, habits, and ongoing projects. In each conversation, it can refer to the previous context instead of starting from scratch every time, as is the case with traditional AI systems.

In addition, Clawdbot has a certain "self - improvement ability". For example, if you tell it that it should be able to generate images, it will research on its own how to connect to the image - generation API and improve its own code. Then it will inform you that it can now generate images.

The Clawdbot team calls this process "Self - Expanding Skills".

It first searches for solutions in the official skills library and the locally installed skills. Only when no suitable resources are found will it conduct a full - scale API research and code development.

After the skill development is completed, Clawdbot conducts a syntax check, an access permission check, and a function test. If the test fails, it will automatically revert to the previous state and report the error to you without affecting the existing functions.

In addition, Clawdbot is an active AI assistant. Via cron time tasks, it can proactively send you messages. For example, it can send you a summary of your daily schedule every morning, create a summary of your work progress on Friday afternoons, or monitor certain conditions and warn you if necessary.

This activity breaks the traditional passive response pattern of AI systems. You no longer have to call it or inquire; instead, it will proactively contact you at appropriate times.

To run Clawdbot, you don't need expensive hardware. Although there are many stories of people buying a Mac mini to run Clawdbot, a $5 - per - month VPS is actually sufficient for basic functions such as chatting, summarizing, and API calls.

Only when you need to run a local large language model or perform intensive automation tasks should you consider using local hardware such as the Mac mini. The buying spree for the Mac mini is more of an expression of social proof and geek culture than a necessity.

Clawdbot can not only control your computer. Technically, it supports over 50 integrations, including Gmail, GitHub, Obsidian, Notion, Todoist, Spotify, Philips Hue smart lights, etc.

This means that you can control all the smart devices in your house with an SMS.

Clawdbot can connect almost all the digital tools in your life and work and thus become a real digital life center. Since it is open - source, the community constantly contributes new skills and plugins.

02

Why is Clawdbot so popular?

Clawdbot's popularity stems from the fact that it solves the central problems of existing AI tools: The "friction" is too high.

This problem sounds simple, but it is the biggest hurdle preventing AI from being truly integrated into our daily lives.

With traditional AI systems, you have to open a website, switch between different tabs, and copy and paste results. Each step interrupts your workflow.

You are writing an email and suddenly want to use AI to improve a paragraph. You then have to switch to the ChatGPT tab, paste the text, wait for the response, and then copy it again.

This process seems to take only a few seconds, but it interrupts your train of thought and makes the whole process disjointed. In contrast, Clawdbot is directly in your chat app. You can simply send a message and get a response without having to change the context.

Another reason is privacy.

In an era when AI services usually depend on the cloud, Clawdbot's principle of local priority, openness, and user control meets the core requirements of the technological community. Your conversation records, personal preferences, and automation scripts are only stored on your own device and can be viewed and edited at any time.

The "data sovereignty" is in your own hands. This is a great advantage for privacy - conscious users.

A deeper reason is that Clawdbot represents a new software paradigm. Traditional software is static: Developers write the functions, and users use them passively. In contrast, Clawdbot is dynamic and malleable.

You can teach it new skills at any time and have it adapt to your own needs.

This is actually a new model called "Software as a Chat" (SaaC). Each user can have a customized assistant that matches their own language style and needs.

From the community's reaction, it can be seen that Clawdbot's popularity also reflects a collective desire. People have been waiting for over a decade for Siri to become intelligent and for almost a decade for Alexa to become really useful. However, the products of these large companies still remain at the level of "setting alarms, querying the weather".

Clawdbot, on the other hand, has achieved in just a few weeks with open - source technology and the power of the community what these large companies have promised but not implemented for years.

The atmosphere of "Everyone is a product manager and everyone is a developer" has strongly resonated in the technological community.

Technically, Clawdbot is not particularly complex. Essentially, it is a "harness" that converts the strengths of large language models into practical assistant functions. Without the progress in these models, Clawdbot would not be possible.

But its product concept meets people's needs.

The focus of an AI assistant should not be on the AI but on the assistant. It should not be a separate website but integrated into your daily tools.

Even more important is Clawdbot's extensibility. Some people let Clawdbot summarize the health data of their Whoop bracelet every morning, others let it automatically process insurance claims or monitor the server status and automatically repair problems.

All these functions are not preset but developed by the users themselves to meet their own needs.

03

The chance for Yuanbao is here

Clawdbot's success shows that an AI assistant should be integrated into the existing communication tools of users instead of requiring users to install a new app.

In China, this communication tool is WeChat.

WeChat is not only the largest communication tool in China but also an integral part of the modern life process.

People chat, work, pay, and shop in WeChat. Almost everything related to digital life can happen in WeChat.

If Yuanbao can be integrated into WeChat in the form of a contact list, it has a natural advantage in reaching users. Users don't have to download a new app or learn a new user interface. They can simply send a message to Yuanbao as if to a friend. This user - friendly way of use is superior to any separate app.

However, the advantages of WeChat...