Tencent's Lobster team answers questions: The popularity of Lobster does exceed its capabilities, but it is the first step for ordinary people to embrace the AI wave.

Text by | Wang Yuchan

Edited by | Su Jianxun

In the past week, Tencent, known for its steady - style, has acted out of character. It not only boosted the public's enthusiasm for "lobsters" through an offline event of "helping to install lobsters for free", but also launched a large number of "lobster" products in a short period.

Currently available products include WorkBuddy, an office tool known as Tencent's version of "deployment - free crayfish"; QClaw, which supports one - click startup and integrates AI agent capabilities into both WeChat and QQ; Tencent Cloud Lighthouse, a lightweight application server for developers, and supporting ecological and security tools like Skillhub. CEO Ma Huateng revealed in his WeChat Moments that Tencent will also launch a series of products such as self - developed lobsters, local lobsters, cloud - based lobsters, enterprise lobsters, cloud desktop lobsters, secure isolation lobster rooms, cloud security guards, and knowledge bases.

OpenClaw, also known as "lobster", is the hottest open - source Agent framework at present. This Agent, which can be directly deployed on a local computer, can operate the system, access files, and reply to messages like a human being. At the same time, it can break through physiological limits and work 24/7.

After the Spring Festival, "lobsters" have become a "battlefield" for major tech companies. ByteDance's Volcengine launched ArkClaw, and Alibaba Cloud's Tongyi Lab introduced CoPaw. The threshold for ordinary people to install lobsters has been quickly lowered. However, the public still has many questions about what lobsters can actually do and how to avoid risks such as information leakage and accidental file deletion.

On the evening of March 12th, Tencent's "Lobster Special Task Force" was interviewed by several media outlets, including 36Kr. Ding Ning, the person - in - charge of Tencent Cloud's developer AI products, Zhong Yucheng, the director of Tencent's lightweight cloud, Su Jiandong, the general manager of Tencent Cloud Security, and Xie Yizhi, the person - in - charge of the AI Agent Security Center, revealed the ins and outs of the "stall - helping - to - install - lobsters" event and answered many questions about security, commercialization, and the product matrix.

The following is the transcript of the dialogue. 36Kr has made non - meaning - altering deletions:

About the "stall - helping - to - install - lobsters" event: An unintended national AI experiment

Q: How was the "stall - helping - to - install - lobsters" event under Tencent's headquarters building on March 6th realized? How was it prepared at the beginning?

Zhong Yucheng: The whole process was actually very fast. On March 1st, we distributed hundreds of free experience vouchers for the lightweight cloud OpenClaw within the company. We originally thought it would be enough, but unexpectedly, they were snapped up in less than 20 seconds. Then many colleagues asked about how to install and run it, so we decided to hold an offline computer - installation event at the north square of Tencent Binhai Building.

Our initial idea was quite simple. It was only for internal colleagues, so we scheduled it for Friday, March 6th. But later, when we talked to our colleagues about this idea, they thought we could expand the scope to allow users outside the company who were interested to participate. So we sent a preview through the video account two days before the event. Unexpectedly, the video went viral, exceeding our expectations.

In terms of personnel arrangement, we initially recruited only 10 volunteers. The core product and R & D colleagues of the team gave them a one - or two - hour quick training, teaching them how to help users deploy and demonstrate basic scenarios. After seeing the popularity on the video account, we expanded the number of volunteers to 20.

On the day of the event, since users started queuing up at 9 a.m., we temporarily reinforced with nearly 10 volunteers from the cloud server computing team.

The situation on - site was really touching. The users who came were from all walks of life: a retired engineer in his 60s, a mother pushing a baby stroller, primary school students, etc. About 80% of them had non - technical backgrounds and were simply interested in AI. We originally planned to finish the installation for each person in five minutes, but in fact, it took an average of twenty to thirty minutes for one - on - one introduction and debugging. Some people even stayed on - site for a whole day. That day, we helped more than 500 people complete the installation offline. The fermentation of the whole event completely exceeded our expectations.

Q: So was this just an "unintended" event? Before starting, did you have the marketing motivation to promote other "lobster" products of Tencent through this event? How was it evaluated internally after the event?

Zhong Yucheng: To be honest, we really didn't think that much at the beginning. At that time, WorkBuddy and QClaw were still mostly in the internal stage and not officially launched to the public. Our original intention was just that cloud - based deployment of OpenClaw was more complicated than local deployment, and we needed to explain it clearly to users. After the event's influence expanded, various products started to collaborate quickly and jointly formed the current "Tencent Lobster" product matrix.

Q: Many developers reported that "raising lobsters" consumes a lot of computing power and memory. In the past month, has Tencent Cloud's backend seen a significant increase in Token consumption? If the popularity continues, will it bring computing power pressure?

Zhong Yucheng: Due to issues related to user privacy and permissions, we haven't made an accurate statistics on the specific Token consumption. However, it is foreseeable that the consumption of computing power of the whole model has indeed increased a lot compared with before, but there is no specific quantitative data yet.

About security and architecture: The "moat" of closed - source products

Q: Someone on X once induced OpenClaw to hand over the API Key through instructions. If I use WorkBuddy, when it scans a similar instruction, will it also execute the command and send my private API Key to the comment area?

Ding Ning: It should be clear here that OpenClaw is an open - source solution, while WorkBuddy is a closed - source product. From an architectural perspective, WorkBuddy integrates our self - developed and long - put - into - production capabilities, such as the Gateway Server and Agent OS in CodeBuddy and Cloud Studio. We had a relatively mature Sandbox mechanism for developers a few years ago, and we just modified it to become the Agent OS.

Coupled with the solutions provided by Tencent Security Lab, it can well prevent behaviors such as software supply - chain poisoning. The most important thing is that WorkBuddy only takes effect on the content of the work folder specified by the user. For example, when you do local file organization or format conversion, it works within the scope you specify, and its permissions are not that large. So the kind of over - authorization behavior you are worried about will not occur.

Q: Then, if someone pulls code from GitHub or uses WorkBuddy to poison Skill Hub, can you detect it?

Xie Yizhi: We can detect it. There are multiple laboratories within Tencent, and we have innovatively proposed an automated review mechanism of "Agent against Agent". We have multiple lines of defense internally, from traditional rule - based and feature - based methods to AI - based detection mechanisms for malicious code, to prevent the inflow of malicious plugins from the source. This ability will be integrated in the creation and upload processes, and once a risk is identified, it will be actively intercepted.

Q: The general public is worried that there might be bundled pop - up ads and other functions in WorkBuddy.

Ding Ning: The WorkBuddy team is the CodeBuddy team that does AI Coding. We started from DevOps, code repositories, and artifact repositories, so we don't have the advertising gene you mentioned. The Skills built - in WorkBuddy have all been reviewed by ourselves and are fine. However, if users find plugins from third - parties themselves, it's the same as when users insist on installing un - reviewed applications on their mobile phones.

Paradigm revolution: Why did "lobsters" become popular?

Q: The popularity of OpenClaw reminds people of the ChatGPT moment. What are the differences between these two waves? Some people say that large models only represent the intelligence of lobsters. What do you think of the relationship between large models and lobsters?

Ding Ning: This is a paradigm shift from "dialogue" to "execution". Large models were just a dialogue box at first, while the current OpenClaw has accelerated the promotion of this model. When AI Coding develops to the Autonomous Development stage, Backend Agent and Skills can already support more general productivity scenarios. Even if you don't understand code, you can enjoy the benefits of "instant execution upon speaking". OpenClaw gives people hope. Although the product form has existed for a long time, its popularity has indeed made this model more precisely known to people.

Zhong Yucheng: I think lobsters are not a revolutionary technological breakthrough like ChatGPT. They mainly allow the general public to feel AI Agent up close. The things lobsters can do can actually be done in some tools in 2025, but the threshold for the public is too high. Although it is still relatively complicated to configure at present, it is still the best option on the market for most people to quickly feel the changes that AI brings to life.

Q: Some people question that the popularity of lobsters exceeds their capabilities and mock them as a product of the "aluminum - pot - on - the - head" era, or even say it is to solve the anxiety of slow Token consumption. What do you think of these voices?

Zhong Yucheng: There may be some anxiety factors. Currently, it really can't reach the level of a "one - person company". However, from the offline computer - installation event we held at Tencent Building last Friday, the situation is quite different. There was a retired engineer in his 60s and a mother pushing a baby stroller on - site. Although some people mocked them for not understanding lobsters and wasting time, I think the fact that they took the step to really get in touch with AI means they have made more progress than most people. For these people, lobsters are a specific way to embrace the AI wave.

Commercialization and the future: From "individual soldiers" to "teams"

Q: WorkBuddy has been used internally since January. Why did you choose to officially launch it this week? Is it to catch this wave of popularity?

Ding Ning: It is indeed related. The popularity has accelerated our pace. We originally planned to release it in mid - March. When DeepSeek was popular last year, everyone was working during the holiday. To prevent this from happening this year, we made preparations for an early release at the beginning of the year. Since the right timing came after the Spring Festival, we just launched it.

The biggest iteration of the current version compared to January is the connection of the long - link of IM tools, which allows for remote operation through mobile QQ, Feishu, and WeChat. Since we want to build AI Teams, we must cover these scenarios in advance.

Q: Tencent has launched several "lobsters" now. In addition to WorkBuddy, there is also QClaw. Will these products be integrated into a unified super digital avatar in the future?

Ding Ning: When high - value business scenarios require integration, we will naturally do so. Currently, it is still in the early stage. We need to build products based on successful closed - loop requirements and won't imagine a product form out of thin air.

Zhong Yucheng: Currently, WorkBuddy is more of a "local lobster", while the lightweight cloud (Lighthouse) provides a "cloud - based lobster". We provide a more user - friendly OpenClaw product form in the cloud. In the future, we don't rule out the possibility of trying to link or integrate the cloud and local.

Q: Pony (Ma Huateng) previously said that he would strongly support the internal incubation of AI applications. How has the company tilted in terms of computing power and manpower after the products were launched?

Zhong Yucheng: For the lightweight cloud, the biggest support is, of course, at the computing power level. In addition, teams such as Tencent Docs, Tencent Maps, and Emma Knowledge Base have provided a complete Skill ecosystem support, enabling us to offer users more diverse scenarios. We also created SkillHub, a skills plaza that is more user - friendly and secure for Chinese users, which solves the problems of overseas access failure and compliance filtering. Although the situations of different teams vary, currently everyone is moving forward rapidly in their respective areas of expertise.

Su Jiandong: I'd like to add something about the security of WeChat connection. Currently, the direct connection through the WeChat service account is equivalent to a robot dialogue channel. The instructions are sent to the backend for execution, not on the mobile phone. So it doesn't have the ability to retrieve your private data in WeChat.