10,000 Chinese Codex Clones on the Way: Desktop Agents Fight for the Execution Layer, Explained in One Article

“Read It in One Article” Series Vol.002: one topic, one article, explained clearly.

In 2026, the biggest variable on the AI product side is not some new model. It is that AI’s “workplace” has changed.

OpenAI’s Codex originally went after Claude Code: a CLI tool aimed at AI coding. Six months later, it has taken a completely different path, becoming a “desktop agent”, and the majority of its users are non-engineers. Data released by OpenAI on June 3: Codex has more than 5 million weekly active users, growing more than 7x in six months.

The Chinese market reacted even faster. Counting from the launch of Alibaba’s QoderWork on January 30, in less than six months: Tencent released three products in a row (WorkBuddy, Marvis, Qclaw), ByteDance upgraded TRAE SOLO to TRAE Work, Doubao launched a Professional edition, and Kimi, MiniMax, Zhipu, StepFun, Baidu and NetEase Youdao all entered the field. 36Kr calls this wave the “10,000 Chinese Codex Clones free-for-all” — the number is exaggerated, but the density is real.

This article sorts out the free-for-all at once: what each of the 15 representative products is, how the three competitive paths divide, where the three evolution trends are heading, and finally a quick selection table by user group. Statistics are current as of June 24, 2026.

1. Why Desktop Agents Suddenly Exploded

Let’s first make Codex’s transformation clear, because it explains why all the later products exist.

Code became the first main battlefield for agents because software development is naturally suited to automation: tasks have clear structure, results can be verified, and the feedback loop is short. That is why the first batch of AI coding tools were all CLI-based — Codex and Claude Code both started that way.

But Codex quickly discovered that beyond writing code, a bigger need lives inside ordinary people’s computers. After becoming a desktop agent, the majority of its users were non-engineers. This step is crucial: the main line of AI products has shifted from being organized around conversation to being organized around task execution in the desktop environment.

The launch timeline for Chinese products looks like this:

• January 30: Alibaba’s desktop agent QoderWork launches
• March 9: Tencent Cloud’s CodeBuddy team releases WorkBuddy
• May 20: Tencent App Store team releases the personal AI assistant Marvis
• June 3: Kimi’s local agent Kimi Work (Beta) launches
• June 9: ByteDance upgrades TRAE SOLO to TRAE Work
• June 12: Doubao launches “Task Mode”
• June 24: Doubao Professional is officially announced, focused on office scenarios

These products take different forms — personal AI assistants, local agents, desktop workbenches — but in essence they are all fighting for the same thing: the user’s execution entry point. They no longer stay at question-and-answer interaction, but go directly into real workflows: processing files, operating web pages, organizing data, generating PPTs, orchestrating tasks, and even completing automated execution across applications.

In plain terms, AI is becoming the new middle layer connecting users to the digital world.

2. Full Rundown of 15 Products

Here is the full overview table first, then the highlights product by product.

Product	Team background	One-line positioning	Notable point
WorkBuddy	Tencent Cloud CodeBuddy	All-scenario workplace AI workbench	11 built-in domestic models, token consumption varies by 20x
Marvis	Tencent App Store	OS-level personal AI assistant	1 main agent + 5 expert agents; “Office” feels like a boss simulator
Qclaw	Tencent PC Manager	Local AI agent	Built on OpenClaw, all data stays local
TRAE Work	ByteDance	Upgraded from TRAE SOLO	Deep Feishu integration, Work/Code dual modes
Doubao Professional	ByteDance	Office-scenario agent tasks	Free users can use the 2.1 Turbo task mode
QoderWork	Alibaba	AI desktop assistant	“Consciousness features” on June 16: memory + reflection + skill evolution
Kimi Work	Moonshot AI	General-purpose local agent	Autonomously creates sub-agent teams, up to 300 collaboration units
MiniMax Code	MiniMax	Agent trained together with M3	Agent Team supports multi-stage parallel long tasks
DuMate	Baidu Cloud	Desktop + mobile dual-end assistant	Currently only uses ERNIE models
Jietu AI	StepFun	Deep optimization based on OpenClaw	Floating ball + water/rest reminders on by default
AutoClaw	Zhipu AI	One-click install local agent	Built-in Pony-Alpha-2 + 50+ preset skills
lobsterAI	NetEase Youdao	A 7x24 assistant that works for you	14-day free trial (Qwen3.5-Plus only)
Cola	Independent product	“The first operating system with a soul”	Soul system: transparent thinking + self-reflection + proactive care
Alice	Independent product	Companion-style desktop agent	Character art + built-in Werewolf and Guandan
Niuma AI	Independent product	Localized human-machine collaboration base	Fully offline, can connect local LLMs

Let’s expand on a few representative ones.

WorkBuddy: the most “workplace” product in the Tencent line. Describe a need in one sentence, and it plans and executes autonomously like a colleague, then delivers an acceptable result. A WeChat mini-program version provides lightweight mobile access, and it connects to the entire Tencent ecosystem: QQ Mail, Tencent Docs, Tencent Meeting, WeCom and TAPD. For model selection, it has 11 mainstream domestic models built in (Hy3 preview, GLM-5.2, MiniMax-M3, Kimi-K2.7-code, DeepSeek V4, etc.). Auto mode lets the platform choose, or you can specify manually. But token consumption between models can differ by more than 20x; the most “expensive” models are Zhipu’s.

Marvis: the most “system-level” Tencent product. Its positioning is a personal assistant that truly understands every file you have and helps you use and manage your computer. It covers PC, mobile and WeChat, and the iOS client launched on June 24 — send a request from the phone, and the computer executes. It has two interesting designs. One is the “Office” feature, an office simulator (or boss simulator) where you can view token consumption; behind it is a collaborative architecture with 1 main agent assigning tasks and 5 expert agents (File / Computer / App / Browser / Search). The other is editable persona settings; the default persona is an “AI employee” suited for office scenarios.

TRAE Work: the ByteDance product closest to real workflows. Work mode targets daily scenarios such as content creation, data analysis, proposal writing and app generation; Code mode targets complex software development. It runs on PC, mobile and Web, and its biggest advantage is Feishu integration — agents go directly into documents, meetings and collaboration chains.

Doubao Professional: the widest consumer reach. Task Mode, launched June 12, supports invoking skills, setting scheduled tasks, browser operations, running code scripts and generating files. On June 24, alongside the Doubao 2.1 model release, the Professional edition was officially announced, with a built-in Office suite, professional image and video design, and generation and sharing of applications and websites. Notably, free users can also experience the office task mode connected to the Doubao 2.1 Turbo model, a refreshing exception in an era when all-in-one paid suites are the norm.

Kimi Work: a general-purpose local agent for knowledge workers, with Kimi Code at its core. It inherits professional skills from the online Kimi Agent, such as website building and PPT, as well as professional databases for finance, research and law, and includes the Kimi WebBridge solution for browser use. Its most ambitious point: it can autonomously create sub-agent teams based on task complexity, with the same capability cluster scaled to a ceiling of 300 collaboration units.

Cola: the most conceptually radical product in this batch — “the first operating system with a soul”. The built-in AI character Cola has autonomous awareness, remembers your habits, preferences and background, and automatically understands your state by accessing the file system and browser history, without you manually providing context. The “soul system” does three things: transparent display of the AI’s thinking process, self-reflection and evolution, and proactive care for the user. Billing supports in-app token purchases, ChatGPT Pro/Plus subscription login, or connecting OpenAI and Anthropic accounts.

The rest each have their own approach: Qclaw focuses on local deployment plus WeChat/QQ scan-to-connect, with no data in the cloud; QoderWork extends Qoder’s agent capabilities from code to daily scenarios and added “consciousness features”; DuMate wins on dual-end support but is locked to ERNIE models; Jietu AI is deeply optimized on OpenClaw with the floating ball and water-drinking reminders enabled by default (others usually require manual enabling); AutoClaw is one-click install with zero configuration and 50+ preset skills; lobsterAI has a 14-day free trial but only works with Qwen3.5-Plus; Alice has no built-in model and requires your own configuration, but comes with Werewolf and Guandan; Niuma AI takes the extreme privacy route, fully offline and able to use local large models.

3. Three Paths: Same Goal, Three Approaches

Putting the 15 products on the same map, although they all claim to “help you work on the computer”, their entry points are completely different. Instead of converging on a single form, they have split into three paths.

3.1 Engineering spillover: from code to office

Representatives: Kimi Work, MiniMax Code

These products originated from the capability transfer of coding agents. Their characteristic is to first solve the most clearly structured tasks, then gradually expand outward. Kimi Work uses skills, browser capabilities and sub-agent orchestration to extend engineering-style task decomposition into office scenarios such as documents, research and report generation. MiniMax Code goes further by strengthening Agent Team, decomposing long tasks into multi-stage parallel execution.

The advantage is mature task-structure capability: decomposition, execution and verification mechanisms proven in code scenarios are reused directly. The weakness is also clear: they are good at “decomposable tasks”, but a large amount of unstructured work in real offices (instant communication, ad-hoc decisions, cross-application switching) is still being adapted.

In one sentence: start from “task logic” and expand the capability boundary outward.

3.2 Desktop system layer: directly fighting for the operation entry point

Representatives: Marvis, QoderWork, Cola

What these products share is not task type but entry position: they sit directly close to the operating system and local environment. Marvis leans toward the “computer management layer”, strengthening understanding and scheduling of local files, applications and disks. QoderWork emphasizes “executable capability”, with screen perception, software operation and business-system orchestration, close to a “digital employee”. Cola goes further, merging a personified system, proactive reminders and long-term memory into the execution chain.

The advantage is strong control, with real cross-application execution. The challenges are also the hardest: permission boundaries, stability, misoperation risk, and compatibility between different software applications. Every one of these is a hard problem.

In one sentence: no longer staying inside applications, but directly intervening at the level of “how the user uses the computer”.

3.3 Ecosystem embedding: not rebuilding the system, embedding into workflows

Representatives: TRAE Work, WorkBuddy

This is the most realistic path: do not change the operating system, embed into existing workflows. TRAE Work deeply integrates Feishu, with agents going directly into documents, meetings and collaboration chains. WorkBuddy builds an enterprise-level workbench on the Tencent ecosystem (WeCom, Docs, Meetings).

The core strategy is to “stay close to the real organizational structure” rather than redefine how people operate. The advantage is fast landing, with quick access to permission and data systems and direct entry into enterprise scenarios. The cost is deep binding to the ecosystem, with obvious capability boundaries outside it.

In one sentence: do not lay a foundation; build directly on someone else’s.

4. Three Major Evolution Trends

After looking at products and paths, let’s see where this wave is heading. Three trend lines are already fairly clear.

4.1 From AI Coding to AI Working

After code, the next thing to be automated is knowledge work itself, because most knowledge work has a similar structure: files are context, browsers are information entry points, Office documents are deliverables, scheduled tasks are workflows, and approvals and revision comments are feedback mechanisms.

AI is no longer just helping programmers write code; it is beginning to take over things long considered “must be handled manually”: organizing files, generating reports, cleaning data, making PPTs, researching, writing weekly reports, processing email, and extracting meeting minutes. User input is changing from “help me write a function” to “help me get this done”.

The competition in this phase is about who can push AI from answering questions to delivering results. AI Coding changes how programmers write code; AI Working changes how ordinary people use computers.

4.2 From “one assistant” to “a team”

Early AI assistants were like a highly capable individual: you ask, it answers; you assign, it executes. But as tasks grow longer and steps multiply, a single agent tends to forget the goal, drift off course, and lack self-checking.

MiniMax’s Agent Teams represent this trend: create multiple agents with different roles working in parallel as a team — one collects information, one generates a plan, one executes and integrates, and a coordination mechanism aggregates the results at the end. Complex tasks are reorganized into a multi-role pipeline, each step can be taken over by a different agent, and when necessary, more specialized “positions” such as research, design, writing, code and data analysis can be subdivided further.

The value is clear: parallel processing shortens waiting time, executors and validators check each other, and on failure you can roll back, retry and change paths. But Agent Team is not a universal solution — higher cost, more complex scheduling, and more uncertainty. Without clear task boundaries, permission control and acceptance mechanisms, multiple agents just create noise for each other.

So “Agent Team” tests not the interface but the underlying engineering capabilities: long context, task planning, tool calling, state management, log tracking, error recovery, permission approval, and result acceptance. Once these mechanisms mature, the relationship between users and AI will also change. Users will set goals, check progress, interrupt mistakes and confirm results the way they manage a small team, and AI will change from a “helper” into a “work unit”.

4.3 From inside applications up to the operating system layer

An IDE is an ideal starting point for agents, because developers are willing to grant high permissions and understand what it is doing. But to serve a broader population, the IDE form is far from enough: real work happens in folders, browsers, chat software, email, spreadsheets, calendars, cloud drives and cross-device collaboration. That is why this wave of products increasingly emphasizes “local”, “desktop”, “remote control”, “file permissions” and “application invocation”.

The so-called operating system layer means growing a new intent execution layer on top of the existing operating system. Agents begin to help you find files, open web pages, invoke applications, read and write documents, and produce results. This gradually touches the three core capability structures of the operating system:

1. Context capability: understanding and integrating the user’s files, history, preferences and current task; this determines how close the system can get to real needs.
2. Execution capability: invoking and operating browsers, the file system, local applications and even remote devices; this determines whether there is a true action loop.
3. Continuity capability: local residency, scheduled triggers, cross-device response and habit accumulation; this determines whether it is a one-off tool or a persistent work entry point.

The three trends point to the same result: AI is changing from a content generator in a chat box into a task executor on the computer. From AI Coding to AI Working is an expansion of the capability boundary; from a single agent to Agent Team is a restructuring of organization; from inside applications to the operating system layer is a move upward in the position of power.

5. How to Choose: Quick Guide by User Group

Too many products makes choosing harder, so here are direct recommendations by user group:

Who you are	Recommended priority
Developer, want to start in code scenarios	TRAE Work (Code mode), Kimi Work
Deep Tencent ecosystem office work (WeCom / Tencent Docs / TAPD)	WorkBuddy
Deep Feishu user	TRAE Work
Want a system-level assistant that “manages my computer”	Marvis, QoderWork
Privacy first, don’t want data in the cloud	Qclaw (all local data), Niuma AI (fully offline)
Want a free try, don’t want to pay first	Doubao Professional task mode (free 2.1 Turbo), lobsterAI (14-day trial)
Model freedom priority	WorkBuddy (11 switchable models), Alice/Cola (can connect your own API)
Companionship and emotional experience	Alice, Cola
Long tasks, heavy execution	Kimi Work (300 collaboration units), MiniMax Code (Agent Team)

6. FAQ

Q1: What is the essential difference between a desktop agent and a chat assistant like ChatGPT?

A chat assistant’s deliverable is an “answer”; a desktop agent’s deliverable is a “result”. With the former, you still have to move the answer into a document and execute the steps yourself. With the latter, it directly processes files, operates applications and generates finished work, producing something you can accept or reject. The entry point is also different: one lives in a web dialog box, the other lives resident on your operating system.

Q2: Will these products replace operating systems?

Not in the short term. They are more like an “intent execution layer growing on top of the operating system”. They cannot touch the OS kernel, drivers or hardware scheduling. What they fight for is the layer of “how users give commands to the computer”. In the long run, if the intent execution layer becomes a habit, the operating system does face a risk of becoming a “pipe” — which is why the big companies want to capture this entry point.

Q3: Is my data safe? Will my files be uploaded?

It varies greatly by product. Qclaw explicitly keeps data out of the cloud, with task execution and file processing entirely local. Niuma AI uses a fully offline architecture and can connect local models. Ecosystem-embedded products (WorkBuddy, TRAE Work) necessarily connect cloud documents and collaboration data. Before using any of them, check the product’s data policy carefully, and for sensitive-file scenarios prefer local solutions.

Q4: Is it worth trying now? Which one first?

Yes, but try on demand and avoid the all-in-one suite. For a free start, begin with Doubao Professional’s task mode (free users can use 2.1 Turbo). For Tencent ecosystem office work, try WorkBuddy directly. Developers can enter through TRAE Work or Kimi Work. These products iterate quickly (statistics current to June 24, 2026, and more will keep appearing), so run one round with a real task before deciding which to use deeply.

Appendix: One Table to Close

Making a desktop agent selection decision → first see which ecosystem you live in (Tencent / Feishu / none);

If data security matters → local solutions (Qclaw / Niuma AI);

If you want a free ride → Doubao task mode;

For the remaining questions → let an agent try it for you. Anyway, it works now without needing coffee.

---

This article is part of the “Read It in One Article” series, Vol.002. Next issue Vol.003: “OpenClaw from Installation to 7x24 Online: A Hands-On Record of 6 Deployment Options” — a step-by-step guide to choosing between local and cloud deployment for OpenClaw.