Author | Sun Rui, Email | sunrui@pingwest.com
The biggest trend on the AI product side in 2026 is that both internet giants and model providers are developing their own versions of Codex.
Initially, Codex was OpenAI's product to compete with Claude Code. It started with only a CLI, targeting AI coding, but it quickly diverged from Claude Code's path, evolving into a "desktop Agent" product. Non-engineers now make up the majority of its users, and it is no longer limited to programming scenarios.
This move proves to be highly critical today. According to data disclosed by OpenAI on June 3, Codex's weekly active users have exceeded 5 million, representing a growth of over 7 times in half a year.
Behind this is the shift in the main trajectory of AI products, gradually transitioning from conversation-centric to task execution-centric within the desktop environment.
As Codex gains momentum, a batch of Chinese Codex alternatives has also been launched consecutively.
On January 30, Alibaba's desktop Agent QoderWork was launched;
On March 9, Tencent Cloud's CodeBuddy team introduced WorkBuddy;
On May 20, Tencent MyApp team released the personal AI assistant Marvis;
On June 3, Kimi's local Agent Kimi Work (Beta) went live;
On June 9, ByteDance's TRAE upgraded TRAE SOLO to TRAE Work;
On June 12, Doubao launched the "Task Mode," incorporating more comprehensive execution capabilities;
On June 24, Doubao Professional Edition was launched, focusing on office scenarios and capable of executing Agent tasks...
Whether they are personal AI assistants or local Agents, although their product forms differ, they are essentially competing for the user's execution entry point.
These products no longer stay at the Q&A interaction level; instead, they are beginning to enter real workflows. They can handle files, operate web pages, organize data, generate PPTs, orchestrate tasks, and even complete automated execution across applications.
This means that AI is becoming a new intermediate layer connecting users and the digital world.
1 The Chinese Codex Alternatives
To observe this desktop Agent boom more intuitively, we have compiled 15 representative products and attempted to analyze their differences in product positioning, capability building, and ecosystem layout.
A comprehensive comparison including pricing models and model usage designs is placed at the end of the article.
Workbuddy
Workbuddy is an all-scenario workplace AI Agent desktop workstation launched by the Tencent Cloud CodeBuddy team, designed for various functional roles. By simply describing a requirement in one sentence, WorkBuddy can autonomously plan and execute tasks like a colleague and deliver acceptable results.
Workbuddy has a WeChat Mini Program version, allowing lightweight access via mobile devices. It can also achieve ecosystem data integration with Tencent's products (QQ Mail, Tencent Docs, Tencent Meeting, WeCom, Tencent Lexiang, IMA, TAPD).
Currently, Workbuddy has 11 mainstream domestic models built-in, including Tencent's own Hy3 preview, as well as mainstream models from model providers like GLM-5.2, MiniMax-M3, Kimi-K2.7-code, and DeepSeek V4. Users can choose the Auto mode to let Workbuddy select the model, or choose the model themselves. However, the credit consumption for different models can vary by more than 20 times, with Zhipu's model currently being the most "expensive."
Marvis
Launched by the Tencent MyApp team, this OS-level personal AI assistant is based on the latest models such as DeepSeek V4 and Hunyuan 3 / Hy3. It is positioned as a private AI assistant that truly understands every file of the user and helps them use and manage their computer more conveniently. It supports AI search for local documents and images, allows APK and EXE applications to be invoked with a single sentence, and covers multiple platforms including PC, mobile, and WeChat. On June 24, Marvis was also launched on iOS mobile devices, allowing users to send requests from their phones for execution on their computers.
It can be seen that Marvis places more emphasis on the capability of "managing the computer."
During testing, there are two interesting points. One is the "Office" in Marvis, which is similar to an office simulator, or rather a boss simulator, where token consumption can be viewed. This actually reflects Marvis's multi-Agent collaboration feature, with a main Agent responsible for task dispatching and five expert Agents: File Agent for files, Computer Agent for the system, App Agent for applications, Browser Agent for web pages, and Search Agent for search.
Another point is that Marvis supports modifying the persona, with the initial persona being an "AI employee" suitable for office scenarios.
Qclaw
Qclaw is a local AI Agent product launched by the Tencent PC Manager team. Based on the open-source OpenClaw framework, it is positioned as a personal PC AI assistant, focusing on minimalist local deployment. Its core capabilities include WeChat/QQ interconnection, QR code binding, allowing remote control of the computer, file transfer, and setting scheduled tasks via mobile phone. It enables all-scenario automation, including file management, web browsing, office creation, and autonomous execution of multi-step complex workflows. Qclaw's data will not be uploaded to the cloud; all task execution, file processing, and data storage are completed on the local computer.
Qclaw supports creating different Agents to complete tasks in various scenarios, continuing Tencent's consistent companion-style feature.
TRAE Work
TRAE Work is upgraded from the original TRAE SOLO. Currently, the Work mode is aimed at daily work scenarios such as content creation, data analysis, proposal writing, application generation, task advancement, and communication collaboration; the Code mode is aimed at more complex software development and code editing scenarios.
TRAE Work is available on PC, mobile, and Web. Compared with other "xx work" products, the advantage of TRAE Work lies in its integration with Feishu, allowing better access to real workflows.
Doubao Professional Edition
Doubao launched the "Task Mode" on June 12. The Task Mode supports invoking skills, setting scheduled tasks, completing browser operations, running code scripts, and generating files, among other tasks.
On June 24, at the Volcengine Momentum Conference, with the release of Doubao Large Model 2.1, Doubao Professional Edition was officially announced and launched. Doubao Professional Edition supports operating the local computer, using the browser, invoking Skills and scheduled tasks. It has a built-in Office suite and supports professional image and video design, as well as generating and sharing application websites.
Free users can also experience the office task mode integrated with the Doubao 2.1 Turbo model.
QoderWork
QoderWork is an AI desktop assistant under Alibaba. It extends Qoder's Agent capabilities from the coding field to daily work scenarios, describing requirements, executing automatically, and delivering results directly.
On June 16, QoderWork launched the consciousness feature, enabling QoderWork to have a complete system with self-reflection and continuous evolution capabilities, forming a closed loop composed of memory, reflection, and skill evolution. DuMate
DuMate is a product under Baidu Cloud. It has both a desktop client and a mobile app. It can see the screen, operate software, process files, and connect business systems. Its core scenarios include information processing, document generation, data analysis, and process automation. However, currently, DuMate cannot switch models and can only use Baidu's ERNIE large model.
Kimi Work
Kimi Work is a general-purpose local Agent for knowledge workers. Its core is Kimi Code, providing basic local Agent capabilities, including installing and using Skills, running scheduled tasks, etc. Kimi Work inherits professional Skills such as website building and PPT creation from the online version of Kimi Agent, as well as professional databases for finance, scientific research, and law. It also has a built-in Kimi WebBridge solution that can use the browser.
Kimi Work can autonomously create sub-Agent teams based on task complexity, setting the upper limit for clustering the same capability at 300 collaborative units.
MiniMax Code
MiniMax Code is an Agent product designed specifically for and trained alongside MiniMax M3. MiniMax Code can fully leverage M3's capabilities in long context, Coding/Agentic, and native multimodality.
For long-term complex tasks, MiniMax Code's Agent Team can break down large tasks into multi-stage, concurrent, and dynamically adjustable workflows, which are advanced collaboratively by the Agent cluster.
Step AI
Step AI, launched by StepFun, is a desktop AI Agent deeply optimized based on OpenClaw. It can be installed without the need for servers or command lines, can be online 24/7, and can connect to the local operating system and built-in browser to help users execute complex tasks.
What is relatively different is that Step AI reminds users to turn on the floating ball settings during the initial setup. The floating ball can open a direct dialog box and also remind you to drink water and rest at appropriate times. Other products also have similar features, but they usually require users to enable them manually.
AutoClaw
The local AI Agent launched by Zhipu has the slogan "Own a local AI Agent with one click, supporting one-click installation." There is no need to configure the environment, apply for an API Key, or write code. After downloading the installation package, it can be used by double-clicking to install and log in. Its core capabilities include the built-in Pony-Alpha-2 model, AutoGLM browser automation, 50+ preset skills (office, creation, web scraping, code, investment research, etc.), IM integration, and a self-evolution mechanism.
lobsterAI
Launched by NetEase Youdao, this all-scenario personal assistant AI product is positioned as an intelligent assistant that "works for you 24/7." It supports dual-end interconnection between mobile phones and computers. With a single sentence, users can remotely operate the computer to complete multiple desktop file organization tasks, extract key items from calendars and emails, clean and analyze business data, and generate weekly reports and PPTs.
Users can try it for free for 14 days, but they can only use the Qwen3.5-Plus model; using other models requires payment.
Cola
Cola is positioned as the "first operating system with a soul." The built-in AI character "Cola" has autonomous consciousness, can remember user habits, preferences, and background, and grows together with the user through voice or text interaction. It can operate computer files, browse the internet, execute commands, generate text/images/videos, etc., and supports complex task decomposition and parallel processing. Without the need for users to manually provide context, it automatically understands the user's state by accessing the user's computer file system, browser history, etc. Its soul system can realize the transparent display of the AI thinking process, self-reflection and evolution, and the function of proactively caring for the user.
Currently, Cola supports in-app token purchases, ChatGPT Pro/Plus account subscription login, or connecting to OpenAI and Anthropic accounts for billing. It does not support API keys from other model providers for the time being.
Alice
Alice is a companion desktop Agent. The most significant feature of Alice is that she is a "personalized AI assistant" with a specific character illustration and detailed character settings. Alice itself does not require payment, but since it does not have a built-in model, a model needs to be configured to use it.
In addition to performing routine tasks such as file management and scheduled tasks, Alice also has built-in casual games like Werewolf and Guandan, offering more entertainment features than other desktop Agents.
Niuma AI
Niuma AI is positioned as a localized human-machine collaboration base station, emphasizing personal data privacy. It supports a fully offline local operation architecture and can use the user's own local large model to run directly offline. Niuma AI uses the Claude model by default. If you have a Claude account, you can log in directly; if not, you need to configure it yourself.
2 The Entry Point Becomes the Real Battlefield
If this round of desktop-level Agents is placed in the same product map, it can be found that although they all seem to be "helping users work on computers," their actual entry points are completely different. Instead of converging into a single form, they have diverged into three paths along different scenario entry points.
From Code to Office: General Capabilities Spilling Over from "Engineering Tasks"
A category of products represented by Kimi Work and MiniMax Code originally stemmed from the capability migration of Coding Agents.
The characteristic of this path is to first solve the tasks with the clearest structure and then gradually expand outward.
Kimi Work places more emphasis on the "general knowledge work Agent." Through Skills, browser capabilities, and sub-Agent orchestration, it extends the originally engineering-oriented task decomposition capabilities to office scenarios such as document processing, research, and report generation. MiniMax Code goes a step further to strengthen the Agent Team, breaking down long tasks into multi-stage parallel execution, and handling more complex production tasks through role division and verification mechanisms.
The advantage of this type of product lies in its mature task structuring capabilities, but the shortcomings are also clear. They are better at "decomposable tasks," but are still adapting to the large number of unstructured operations in real office work (instant communication, ad-hoc decision-making, cross-application switching).
It can be understood that they start from the "task logic" and expand their capability boundaries outward.
Entering from Desktop and System: Directly Competing for the Operation Entry Point
Another more aggressive path is the "desktop system-level Agent" represented by products such as Marvis, Qoder Work, and Cola.
Their commonality is not the task type, but the entry point location—directly close to the operating system and local environment.
Marvis leans more towards the "computer management layer," emphasizing the systematic organizational capabilities of files, applications, and disks, essentially strengthening the understanding and scheduling of the local operating system. Qoder Work places more emphasis on "executable capabilities," including screen perception, software operation, and business system connection, approaching a "digital employee." Cola goes a step further, integrating the personalized system, proactive reminders, and long-term memory into the execution chain, making the Agent not only a tool but also a continuously existing interaction layer.
The key point of this type of product is that they no longer stay within the application but attempt to directly intervene in the layer of "how users use computers."
The advantage lies in stronger control, enabling true cross-application task execution. However, the challenges are equally obvious, including permission boundaries, stability, the risk of misoperations, and compatibility issues between different software.
Entering from the Office Ecosystem: Not Reconstructing the System, but Embedding into Processes
Different from the first two categories, TRAE Work and WorkBuddy take a more realistic path, not changing the operating system but embedding into existing workflows.
TRAE Work is deeply integrated into the Feishu system, allowing the Agent to directly enter the document, meeting, and collaboration chains; WorkBuddy relies on the Tencent ecosystem (WeCom, Docs, Meetings, etc.) to build enterprise-level workstation capabilities.
The core strategy of this type of product is to be "close to the real organizational structure" rather than redefining the operation methods. Their advantage lies in fast implementation, enabling rapid access to permission and data systems and entering enterprise-level scenarios.
3 Collectively Moving Towards AI Working
The evolution of this round of domestic products is essentially expanding around this "execution closed loop" to broader scenarios, organizational methods, and system entry points. It is precisely in this process that several gradually clear directional changes can be seen.
The first trend is from AI Coding to AI Working.
The reason code first became the main battlefield for Agents is that software development is naturally suitable for automation. However, after the Coding Agent matures, its capabilities will naturally migrate outward.
Most knowledge work essentially has a similar structure: files are the context, the browser is the information entry point, Office documents are the deliverables, scheduled tasks are the workflows, and approvals and revision comments are the feedback mechanisms. They are also multi-step tasks that can be decomposed, executed, and verified.
AI is no longer just helping programmers write code; it is beginning to help knowledge workers complete tasks that have long been considered "must be handled manually": organizing files, generating reports, cleaning data, making PPTs, searching for information, writing weekly reports, processing emails, extracting meeting minutes, and tracking industry information. What users input is no longer "help me write a function," but "help me get this done."
Therefore, the competition in this stage is about who can better advance AI from answering questions to delivering results.
AI Coding changes the way programmers write code, while AI Working changes the way ordinary people use computers to complete work.
The second trend is that Agents evolve from "a single assistant" to "a team."
Early AI assistants were more like highly capable individuals. Users ask questions, and the AI is responsible for answering; users assign tasks, and the AI is responsible for executing. However, when tasks become longer, steps increase, and context becomes more complex, a single Agent is prone to hitting bottlenecks. It may forget the goal, deviate midway, or lack self-checking during execution.
Taking the Agent Teams launched by MiniMax as an example, it allows users to create multiple Agents with different role settings and combine them into a team to work in parallel.
Different Agents can process the same task simultaneously from different perspectives, such as one responsible for information collection, one for solution generation, and one for execution and integration. Finally, a coordination mechanism aggregates the results, thereby improving the processing efficiency and stability of complex tasks.
Complex tasks are being reorganized into a multi-role pipeline: from understanding requirements to planning, execution, and then verifying results, each step can be taken over by different Agents and further subdivided into more specialized "positions" such as research, design, writing, coding, and data analysis when necessary.
The value of the Agent Team is to provide AI with a more stable work structure. It can process multiple sub-tasks in parallel, shorten waiting times, and allow executors and verifiers to check and balance each other, reducing low-level errors. Moreover, it can roll back, retry, and change paths when a task fails, precipitating complex work into reusable processes.
However, the Agent Team is not a panacea. Multiple Agents will bring higher costs, more complex scheduling, and more uncertainty. Without clear task boundaries, permission controls, and acceptance mechanisms, multiple Agents may instead create noise for each other. Therefore, the real key is not simply stacking more Agents, but whether the product can organize them into controllable workflows.
This is also why the ultimate test for "Agent Teams" is not the interface, but the underlying engineering capabilities: long context, task planning, tool invocation, state management, log tracking, error recovery, permission approval, and finally, result acceptance.
As these mechanisms gradually mature, the relationship between users and AI will also change. Users can set goals, check progress, interrupt errors, and confirm results just like managing a small team. AI also changes from a "helper" to a "work unit," and the product becomes the scheduling system for these work units.
The third trend is that the position of Agents is rising from within applications to the operating system layer.
The IDE is an ideal starting point for Agents because developers can proactively grant it higher permissions and understand what it is doing. However, if Agents are to serve a broader population, existing only in the form of an IDE is not enough. Real work happens in folders, browsers, chat software, emails, spreadsheets, calendars, cloud drives, local applications, and cross-device collaboration. This is also why this round of domestic products increasingly emphasizes "local," "desktop," "remote control," "file permissions," and "application invocation."
The so-called operating system layer is growing a new intent execution layer on top of the existing operating system.
After the emergence of Agents, it begins to help you complete execution. Agents can help you find files, open web pages, invoke applications, read and write documents, and produce results...
This allows Agents to gradually touch the three most core capability structures of the operating system.
First is the capability at the context level, involving the understanding and integration of user files, history records, preferences, and current tasks. It determines to what extent the system can approach the real needs themselves.
Second is the capability at the execution level, reflected in the invocation and operation of browsers, file systems, local applications, and even remote devices. It determines whether the system has a true action closed loop, rather than just staying at information generation.
Third is the capability at the continuous level, manifested in local resident operation, scheduled triggering of tasks, cross-device response, and long-term precipitation of user habits. It determines whether the system can transform from a one-time tool into a continuously existing work entry point.
From AI Coding to AI Working is the outward expansion of capability boundaries; from a single Agent to an Agent Team is the reconstruction of organizational methods; and from within applications to the operating system layer is the upward shift of the power position. These three trends actually point to the same result: AI is transforming from a content generator in the chat box to a task executor in the computer.
Finally, a horizontal comparison table of key information for current major products is attached for a more intuitive feel. Products of this form will continue to emerge constantly.
Note:
The above statistics are as of June 24, 2026.
The starting monthly paid price for some products is calculated based on the starting price of the continuous monthly subscription package.
Most of the common domestic models mentioned here include the GLM series, DeepSeek series, MiniMax series, Kimi series, and other models.