Kimi Claw actual test: Amid the OpenClaw craze, automated AI is still a semi-finished product

In 2026, a crayfish disrupted the entire AI circle, and the residual heat of OpenClaw continues to exert its influence after the year.

Recently, a number of domestic model manufacturers have launched products that compete with OpenClaw, such as MaxClaw launched by Mini Max and Kimi Claw launched by Kimi. Obviously, the AI ​​execution capabilities demonstrated by OpenClaw and the tolerance shown by developers for AI execution results have allowed the market to see value space.

Among the benchmark products, Kimi Claw has a relatively clear positioning. It is not a self-developed Claw product from scratch, but a managed cloud service based on OpenClaw. The data is hosted on the Moonshot cloud and is directly configured with 5,000+ ClawHub community skills.

Its advantages are that it is relatively stable to use, easy to deploy, easy to get started, and relies on the cloud to achieve 24/7 online execution. Open the Kimi official website, you only need to click to create, and Kimi will directly deploy Kimi Claw.

Kimi Claw one-click deployment｜Image source: Geek Park

In other words, Kimi Claw is not an independent new product. It is essentially a virtual machine opened remotely for users, allowing users to directly access the OpenClaw environment running in the cloud through Kimi.

It does not have any function deletion or additional packaging, and is almost the same as the local deployment of OpenClaw. It only completes the deployment, configuration, and environment setup steps for the user, but does not do any processing for the tuning process after OpenClaw deployment.If you have not learned to give instructions correctly and arrange tasks reasonably, it will still be difficult to get started.

For users who have never been exposed to OpenClaw products, this will also lead to a misunderstanding of expectations. Users think that they can do automated AI execution by connecting to OpenClaw, but in fact it is just an additional portable interface, and there are still many settings that need to be explored by themselves. Therefore, providing some popular preset Skills for OpenClaw products will become the next focus of many AI model manufacturers.

Currently, Kimi Claw is still in the beta testing stage and is only open to members above Kimi Allegretto.

01

Build an automated office workflow in 30 minutes:

The ideal is very full, and there are thresholds for implementation

We found that many users, like us, are still unclear about the boundaries of AI's execution capabilities after accessing OpenClaw. They are curious about what it can and cannot do, but they are also full of unknowns and don't know where to start after accessing it.

In fact, at present, whether it is deploying automated AI such as OpenClaw locally or directly connecting to an external interface such as Kimi Claw, the overall usage idea can actually be divided intoBuild an application from scratch和Optimizing apps starting from 0.5There are two paths. We have actual experience in these two methods. First, we choose to develop an application from scratch and optimize the workflow.

Before experiencing Kimi Claw, I first examined what tasks I have that can be built into a fixed workflow, or what tasks in my workflow can be made better with the help of AI. Until then, all I had to think about was which type of AI tool I was interacting with to get better results.

I chose the work diary link, combined with the daily work flow, work records, work summary, work reflection and other links to finally output a work report for the day. Finding a report used to be time-consuming for individuals to fill in. Now I hope that AI can automatically capture it and combine it with conversational interaction to automatically form a form.

I first gave the general idea to the AI ​​optimization instructions, and finally gave a very long and complex instruction from multiple levels such as role definition, skill configuration, data access, core workflow, multimedia table structure, memory focus, permissions and boundaries, and handed it to Kimi Claw.

Kimi Claw quickly analyzed the instructions and confirmed the execution details with me. For example, basic information, Feishu permissions, data storage and triggering methods. Then we started following the instructions to build the Feishu app on the Feishu platform, and will send the App ID and App Secret to Kimi Claw.

When there was a link where I needed to build a table in Feishu, I asked Kimi Claw to directly give me the style of the table, and then handed it to Feishu's built-in AI system to let Feishu automatically build the table.

One of the application pages built by Kimi Claw | Image source: Geek Park

After experiencing a series of problems such as not being able to find collaborators, not being able to find application pages, not being able to find IDs, etc., about half an hour later, I successfully received the first message from Kimi Claw.

Building this bot went faster than I expected. When I encounter a problem, I will directly tell Kimi Claw which link I am stuck in, and then choose the appropriate idea to implement it among the solutions given. If the solution given is not suitable, I will continue to ask Kimi Claw for other solutions.

Kimi Claw deploys to Feishu with one click | Image source: Geek Park

When building workflows, the importance of cross-platform capabilities is also more prominent. After opening 12 Feishu permissions in succession, I finally built the AI ​​application and it was not in the ideal state. Among them, I hope that AI can sort out my work tasks by reading my chat records with others, but after several rounds of attempts, the group chat list obtained by the AI ​​application is still empty, and it says that the Feishu AI application requires that the AI ​​can only read the conversations in which I participate, and the application cannot read the group chat list.

After the overall experience, I think Kimi Claw is familiar with some common workflow platforms such as Feishu, DingTalk and other developer tools. Basically, the instructions given can directly find the corresponding execution method, and users with zero basic knowledge can also read and execute them. However, this type of enterprise application will pay more attention to its own information permissions, and the opening configuration conditions are also relatively strict. Perhaps if you want AI to truly integrate into the workflow, you not only need to look at the tools of openers such as Kimi Claw, but also need to wait for the emergence of applications that are more suitable for integrating with AI.

Moreover, there will be many bugs during the operation. For example, during this process, the user's interaction tasks with Kimi Claw and the running Agent tasks will be mistakenly included in the personal work schedule. Learning to modify bugs has also become a key part of training AI.

If you choose to proactively customize the applications or functions you want from scratch, users need to think of a clear operation path and have basic product thinking. It is necessary to clarify the openness and connectivity of the interfaces at both ends of the information input and output, and at the same time control the cost of each call and operation.

To build this workflow, the entire token consumption is about 15k-25k, which is about 1 yuan according to Kimi’s pricing method. But it costs about 0.53 yuan a day, and about 15.9 yuan a month.

02

Automated AI news assistant construction test:

"Pre-made" applications are quick to use but difficult to modify

In addition to letting the AI ​​customize an app I envisioned, I also tried out some “pre-made” apps, such as letting Kimi Claw automatically crawl news.

When we did the first round of automated news crawling tasks, we tried to let Kimi Claw crawl the official website of a technology news media. When we give the instruction:

Please monitor xxxx's industry website and summarize the past week and the next 3 days. Whenever a new article containing the "AI" keyword is published, please automatically capture the title, abstract, and publication time, and summarize these contents into an online table. At the same time, please conduct analysis of popular articles in the report according to the style I set.

Kimi Claw will ask us for specific configuration information, but during the first round of news crawling tasks, we found that many official websites actually have anti-crawler settings, making it difficult to monitor the information of high-quality websites. Kimi Claw is also difficult to give an accurate range grab, so there will be idling, and each idling means a large number of tokens are consumed.

The monitoring task was run about 8 times from 4 a.m. to 11 a.m. today, consuming about 180K tokens and costing about 3.68 yuan. If you run it every hour according to the original settings, it will cost about 11 yuan per day and nearly 330 yuan per month.

Later, after consulting relevant people, we began to give up writing instructions ourselves, and instead downloaded a compressed package of related instructions from relevant websites such as ClawHub. Based on this basic instruction, we continued to customize related news.

Deploy Clawhub files to Kimi Claw｜Image source: Geek Park

Subsequently, we made more detailed settings for Chinese media, news filtering conditions, and the number and time of sending information. Finally, we can get a good version of AI news crawling results.

Kimi Claw automatic crawling results｜Image source: Geek Park

Obviously,If you only passively use pre-made applications, the key is to learn to select high-quality skill packages (skills) and adapt and optimize ready-made functions according to your own scenarios.。

However, if you want to make customized modifications to these pre-made AI applications, you will often go back to the difficulties encountered when building applications from scratch. Development and optimization are not easy, and the final effect may not be ideal.

In this process, users actually need to spend a lot of time to experience the convenience and adaptability of different Skills in the same type of product, and then decide which type of Skills to base on for secondary development, modification and expansion. These actually also consider users’ product thinking.

03

Kimi Claw usage experience:

AI execution is enhanced, instructions are productivity

The core value of Kimi Claw at this stage is just to lower the deployment threshold of OpenClaw so that domestic users can quickly access it. However, the product itself does not come with its own scenarios or skills. It is more like an "interface" rather than a "finished product".

During the experience, we also found that although the underlying model of Kimi Claw is also the Kimi K2.5 model, it is a combination of "naked model + native OpenClaw" and does not inherit the multi-round search, content enhancement, automatic error correction and other capabilities of the official version of Kimi that have been deeply optimized by the search team.

In other words, the official Kimi website is easy to use because there is a dedicated team behind it that has made a lot of optimizations and automatic completion capabilities for the model in high-frequency user scenarios; while the "naked" model accessed in the OpenClaw environment is closer to directly calling the API without special optimization, so the same instructions will appear, and the effect of passing it to Kimi Claw is not as good as directly passing it to the KimiK2.5 model.

After in-depth experience, I can clearly perceive the core differences between Kimi Claw and traditional AI and ordinary Agent products.It is concentrated in the two dimensions of AI execution and the importance of instructions., which is also the key logic for using this type of product.

First of all, in terms of execution, Kimi Claw can also perform tasks when you are not using the computer, instead of the traditional mode where users give instructions and then wait for the task to be completed. I can even tell Kimi Claw when to execute this command, and I can directly see the results of each timed output when I turn it on. But it also reminded me to remember to set a stop endpoint for some experiential applications to reduce unnecessary resource consumption.

Secondly, in terms of instructions, in the past, the instructions given by me and AI were relatively concise and pointed directly to the problem. When the solution direction given by AI was wrong, I would continue to adjust. However, every time Kimi Claw runs a complex command, it will call a large number of Agents to assist, and the tokens consumed will also increase exponentially. Therefore, when giving instructions, it is necessary to clarify the operation method, permission scope, execution path, security and cost control.

For example, in the past, my command when querying news was "Give me 10 news clues about OpenClaw and tell me its news attention value." Now the command I give is:

As an Information Retrieval Specialist, you have access to web search tools (restricted to web_search and web_open_url, access to paid news libraries requiring login is prohibited), subject to the following constraints:

1) First perform a keyword search for 'OpenClaw Latest News' and obtain only the first 5 high-weighted results (prioritizing technical media and official blogs, excluding forum posts);

2) When analyzing the news value of each article, it is strictly limited to the three dimensions of 'technical breakthrough', 'commercial impact', and 'security risks'. Each dimension is summarized in one sentence, and it is prohibited to expand the discussion without background;

3) Disable automatic browser clicks and deep crawler skills throughout the entire process to avoid triggering the anti-crawling mechanism and additional token consumption;

4) The output format is a table: news title | source | attention value label | brief basis (≤30 words/item);

5) If the search results are less than 10, the supplementary search will be stopped immediately and the actual number will be output directly. It is prohibited to initiate a second broad search just to make up the number. It is expected that the token budget will be controlled within 8K, and if the path deviation is found, it will be terminated immediately and reported instead of corrected by itself.

In most cases, I'll even let the AI ​​optimize my command expression before handing it off to Kimi Claw. Only by giving specific and accurate instructions can the best results be achieved within a reasonable token consumption range. Even on many public forums, Skills libraries specially prepared for OpenClaw can help users better get started with some popular application methods.

Precise and concrete instructions are the prerequisite for obtaining high-quality results within reasonable token consumption.The essence of the process of using Kimi Claw is the process of users making trade-offs between model capabilities, output results, and usage costs.。

Kimi Claw ｜Photo source: Geek Park

Finally, train the AI.

Even after you quickly build an AI application, you will find that the AI ​​bot is not easy to use from the beginning. Its division of many instructions and the merging of tasks are actually quite different from human understanding. You still need round after round of instruction training to explore the boundaries of the product. In particular, the interfaces to many information sources are not fully open to the public. Among them, it is not easy to truly access and transfer information rights.

After all, the current application effect of Kimi Claw is by no means a simple AI application such as Chatbot, with many AI functions for users to use directly, but a developer tool that requires users to understand the development process and be able to make choices after many comprehensive trade-offs. It's just that this developer tool can support some simple automated deployment.

04

Automated AI still has room for development

Although OpenClaw will completely ignite people's imagination of automated AI starting in 2026, judging from the recent frequent security incidents and new product testing experience, OpenClaw is still only a key and an opportunity, not the final answer.

Whether it is a real-world scenario that can be implemented or a commercialization path that can be scaled up, the AI ​​industry has yet to take a clear and mature route. In contrast, the market has continued to raise expectations for Claw products in rounds of hype, even attracting a large number of ordinary users to try high-risk operations beyond their own capabilities.

What is certain is that automated AI has been valued by the industry from the first day of AI's birth, but there is still huge room for verification whether OpenClaw and Kimi Claw can develop truly successful and scalable products. Especially now that such AI tools will directly gain permission to modify your terminal and files.

In the early days, the boundaries of AI capabilities were not clear to everyone. Many novices just released their permissions, and it was difficult to think of security restrictions and secondary permission confirmations. Giving such high operational power to AI is essentially a direct opening of system risks. This is why, if such products want to be truly scaled up and commercialized, security and authority management will be more difficult to overcome than "whether they are capable or not."

From talking directly to large models, to interacting with a single Agent, to collaborating with a cluster of Agents, to the way OpenClaw is used today, the industry has derived a large number of attempts with similar functions and different paths based on the same AI capabilities. This just shows that the entire industry is still in the exploratory period of AI functions. In addition to mature and stable interaction paradigms such as ChatGPT, people are still collectively exploring the usage logic, boundaries and value of new forms such as Agent and Claw.

Perhaps, we will have to wait until 2026 to see a batch of stable, usable, and real-value automated AI applications come to fruition.