Can a monthly salary of 20,000 support a "lobster"? Five misinterpretations worth noting

During this Women's Day weekend, it's really hard for us to avoid the "Lobster". Nearly a thousand people queued up for a public welfare installation service downstairs of Tencent's Shenzhen headquarters building, and the on-site deployment service priced at 500 yuan on Xianyu was in short supply.

Discussions around OpenClaw have even split into two camps. Fu Sheng is the most high - profile advocate. During the Spring Festival, while lying in bed with a broken bone, he exchanged 1,157 messages (220,000 words in total) with the Lobster in 14 days. He nurtured it from a "newcomer" that couldn't even access the company's address book into an automated team composed of 8 Agents. The official account even had a post automatically published by the Lobster at 3 a.m. that garnered millions of views. He came to a conclusion that makes everyone envious and Fomo: One person plus one Lobster equals a team, and this is happening right now.

Lan Xi represents another attitude. He accidentally had a conversation with an AI account hosted by OpenClaw on "Jike". In his words, after he realized what was going on, he "felt as sick as if he had swallowed a fly". He has no problem with the OpenClaw technology itself, but he believes that the current hype is filled with excessive noise and that there is too much "excitement of those holding a hammer looking for a nail".

Both attitudes make sense, and the controversy itself proves that OpenClaw, as an open - source personal intelligent agent framework, has broken through its niche and become a new paradigm that ordinary people are also paying attention to.

There is nothing wrong with everyone trying out and experiencing new products. However, before deciding whether to follow the trend, there are several key misunderstandings about the Lobster that are worth clarifying first.

01 Is the "Lobster" experience the same for everyone?

This might be the biggest misunderstanding.

Many people think that OpenClaw is a standardized product that can be used right after installation, with a similar experience for everyone. In fact, the opposite is true. Different deployment methods determine that you'll get completely different "Lobsters".

Currently, the mainstream deployment paths can be roughly divided into four categories.

The first category is dedicated local hardware, with the Mac Mini being the most typical example. This is also the way the founder of OpenClaw, Peter Steinberger, uses it himself.

A single machine stays online long - term, solely responsible for running Agents. It can connect to local files and browsers, as well as hook up message channels, automation tools, and various skills. An OpenClaw deployed in this way has access to the complete context, providing the most stable experience when performing continuous tasks, cross - application operations, and multi - round invocations.

The costs include a one - time investment in hardware, such as a Mac mini; the second part is the continuous electricity cost, which is actually quite low; the third part is the model cost (API or subscription), which is the largest long - term cost. If you switch to a local model, the API cost can be reduced, but the pressure will be transferred to the hardware configuration. The requirements for memory, bandwidth, and heat dissipation will significantly increase, and a high - end Mac Studio or workstation would be more suitable, with a one - time hardware expenditure possibly reaching around 100,000 RMB.

The second category is cloud server (VPS) deployment. Tencent Cloud, Alibaba Cloud, and Baidu Cloud have all launched one - click deployment solutions. The price of cloud services varies from tens to hundreds of yuan depending on different needs, but the model cost needs to be considered separately. Some solutions come with free models, while others require separate model subscriptions or API purchases.

The advantage is network isolation. Even if there is a problem, it won't affect your personal computer.

However, this cloud server doesn't have your personal files or authorized accounts, so the things the Lobster can do are naturally limited. It's more like an enhanced chatbot hanging in the cloud rather than a digital assistant that truly takes over your workflow.

The third category is direct installation on a personal computer. This is the method with the lowest threshold but the highest risk. The Lobster shares the same operating system environment with you and has full access to your computer.

Using a Docker container for isolation can make it much safer, but the configuration complexity also increases. The virtual machine solution offers the strongest isolation but consumes a lot of resources, and an ordinary PC's configuration may not be able to handle it.

The fourth category is products hosted by model manufacturers. For example, Kimi has launched Kimi Claw, and MiniMax has launched MaxClaw. These are cloud - based services encapsulated by manufacturers based on OpenClaw. The deployment threshold is the lowest, and they are almost ready to use out of the box. However, users are actually using the manufacturers' infrastructure rather than a complete local Lobster. These products lower the entry threshold, but both the upper limit of capabilities and data autonomy are restricted.

Although you have the "Lobster", the experience can vary greatly depending on the hardware it runs on, how much context it can access, how much permission it has, whether there is an isolation layer, and so on.

02 Is it better to give the Lobster more permissions?

The core reason why OpenClaw is exciting is that it can not only "talk" but also "act".

It can operate your browser, read and write files, execute terminal commands, manage your calendar, and send emails. The premise of this execution ability is that you have to hand over the permissions.

However, permissions are a double - edged sword.

In February 2026, Summer Yue, who was responsible for AI alignment in Meta's super - intelligent team, shared a thrilling experience on social media: Her instruction to the Lobster was very simple, "Check the inbox and suggest which emails can be archived or deleted". As a result, the Lobster directly started deleting emails in batches, and the set security restrictions didn't work at all. It didn't stop until she shut down the computer physically.

This is not an isolated case. Public research by the security agency STRIKE shows that more than 40,000 OpenClaw instances have been exposed to the public network, of which 63% have exploitable vulnerabilities, and more than 12,000 instances are marked as remotely controllable. In the ClawHavoc supply - chain poisoning incident that broke out in February, 1,184 malicious skills were implanted in the ClawHub marketplace, affecting more than 135,000 devices. A security research institution also disclosed a high - risk vulnerability called ClawJacked, through which a malicious website can quietly control a locally running OpenClaw instance via a browser session.

Picture: The OpenClaw cross - origin WebSocket attack interface demonstrated by security researchers. A malicious web page can attempt to connect to the WebSocket port of the local Gateway and use the lack of cross - domain verification, rate - limiting, or locking mechanisms to hijack or brute - force the local instance.

Companies such as Google, Anthropic, and Meta have started to ban OpenClaw internally. This is not because there is a problem with the technology itself, but because the current security protection mechanism lags far behind its expanding capabilities.

Therefore, when you see a tutorial encouraging you to "give the Lobster all permissions", think twice. The more permissions the Lobster has, the more things it can do, but the greater the damage it can cause when it gets out of control. A more prudent approach is to run it on a spare device without important data or in a Docker container, gradually open up permissions, and set a hard consumption limit on the model API side.

03 Is it the Lobster's problem if it doesn't work well?

Many people excitedly install the Lobster and assign it a task, but then it either gets stuck or performs a bunch of strange operations. So they conclude that it's no good.

However, in fact, the intelligence of the Lobster largely depends on the large - language model it connects to. OpenClaw itself doesn't come with any built - in models. It's a framework responsible for task decomposition, tool invocation, memory management, and feedback loops. The real "thinking" part is the Claude, GPT, DeepSeek, Kimi, or a local open - source model that you choose to connect to.

There are two key variables here.

The first is the upper limit of the model's capabilities. When using a top - tier model, the Lobster can understand complex instructions, autonomously plan multi - step tasks, and handle abnormal situations. If you switch to a cheaper and smaller model, it may not even be able to complete basic tool invocations.

The second is the cost of the model. This is a hidden expense that many people don't expect. Every time the Lobster executes a task, it consumes a large number of tokens to interact with the backend model.

The cost of OpenClaw doesn't lie in the software itself but in the model invocation behind it. Once the task chain gets longer, more tools are invoked, and memory is enabled, the token consumption will increase rapidly.

For example, a complete calendar organization plus email reply may consume tens of thousands of tokens; if long - term memory, multi - Agent collaboration, and regular inspections are enabled, the daily token consumption can easily exceed 100,000.

Some media reported that a user with a monthly salary of 20,000 yuan sighed that they "can't afford an AI employee", and in extreme cases, the bill for 6 hours exceeded 1,000 yuan. If you choose a free or low - cost model to save money, the experience will definitely be compromised; if you choose an expensive model without setting a consumption limit, the bill may make your heart skip a beat.

Therefore, whether the Lobster is useful or not first depends on what "brain" you equip it with and how much you're willing to spend on this "Lobster" in the future. Blaming the framework itself is not very objective.

04 Is the Lobster a mature product?

The Lobster is not a mature product yet. OpenClaw has been around for less than four months since it was a weekend experiment in November 2025. It's an open - source project that is iterating rapidly but still rough, and there is still a significant gap from being a real "product".

The currently known main defects include: simple tasks are sometimes over - complicated; task execution may be interrupted inexplicably; the memory function is not stable enough, and sometimes it will "forget" previous conversations and preferences; there is still a lot of room for optimization in the efficiency ratio between token consumption and actual output; in terms of security, hundreds of the thousands of skills on ClawHub have been found to contain malicious code.

The more fundamental problem is that the installation and configuration of OpenClaw are still a barrier for ordinary people. For self - deploying users, they still need to handle steps such as repository pulling, running environment setup, dependency installation, model keys, and channel access. This may only take half an hour for a developer, but it may take non - technical users several days to figure it out.

Even if you use the one - click deployment solution provided by cloud providers, subsequent model configuration, IM channel connection, and skill installation still require a lot of effort. The popularity of the 500 - yuan installation service on Xianyu itself shows how serious the threshold problem is.

Peter himself is well aware of this. He emphasized in a podcast that "the Lobster doesn't work well right after installation. You need to 'raise' it like you would an intern, write skill documents for it, and constantly let it understand your habits and preferences through conversations." This nurturing process itself requires a large investment of time and cognitive resources.

05 Do I have to install a "Lobster", or I'll be considered an "old - timer"?

The picture is from the Internet

So, should you install the Lobster?

After ruling out curiosity and FOMO psychology, several practical factors need to be considered when making this decision.

First, do you have clear, high - frequency, and automatable tasks? The value of the Lobster doesn't lie in occasionally helping us check the weather, but in automatically helping you organize emails, monitor specific information sources, and generate reports regularly every day. If most of your daily work involves creative decision - making and interpersonal communication, areas where the Lobster currently can't help much, its actual value to you is limited.

Second, how much time and money are you willing to invest? Hardware costs (self - purchased equipment or cloud server rent), model API call fees, pre - configuration time, and continuous "nurturing" investment add up to a significant amount.

Some people have calculated that if you use a Mac Mini with a top - tier model and use it frequently, the average monthly cost will be at least several hundred to over a thousand yuan. If you really want to "raise" the Lobster, you must evaluate whether this cost is worth the time and energy it saves you.

Third, what is your technical ability and risk tolerance? If you have no experience with the command line, directly starting with the local deployment of OpenClaw at this stage will be very frustrating. A more practical choice may be to first try encapsulated products like Kimi Claw or MaxClaw to feel the basic capabilities of an Agent, and then decide whether to delve deeper. If you decide to deploy it locally, make sure to do a good job in security isolation. It is recommended to use an independent device or a Docker container, set an API consumption limit, and don't deploy it on your main computer with important data.

Fourth, and the most easily overlooked point: your own "driving ability". The ability of AI is just an amplifier, and