WorkBuddy vs QClaw vs OpenClaw: What Tencent’s Two Crayfish Actually Are

Let me start with a fact that a lot of coverage is getting wrong.

Search for Tencent WorkBuddy and you’ll find headlines calling it “Tencent’s commercial product built on OpenClaw” or summarizing it as “OpenClaw’s core plus Tencent Cloud’s security layer.” That framing is incorrect.

WorkBuddy’s product lead Wang Shengjie told Chinese outlet Guanchazhe in March: “We are one hundred percent self-developed. We have never used a single line of OpenClaw source code.” He said the same thing to APPSO in a separate interview the same week.

Those articles aren’t inventing things. They’re confusing WorkBuddy with QClaw. Tencent shipped two crayfish-class products around the same time, and one of them genuinely is built on OpenClaw — but it’s QClaw, not WorkBuddy.

This piece separates the three by technical foundation, product philosophy, and who each one actually fits.

Positioning at a glance

Dimension	WorkBuddy	QClaw	OpenClaw
Maker	Tencent Cloud CodeBuddy team	Tencent PC Manager team	Open-source community
Technical base	In-house Agent framework	Wrapper on OpenClaw	Native open-source framework
Core positioning	Desktop workbench, plans and executes	Control your PC from WeChat	Fully customizable Agent framework
Interaction model	Desktop-initiated, multi-step task chains	IM command, reactive	Whatever you configure
Setup barrier	Install and go	Install and go	Environment, dependencies, API keys
Models	Multi-model, bundled	Bundled domestic models, BYO key optional	All self-configured
Cost structure	Credits	Credits, daily free allowance	Pay-as-you-go tokens, hard to predict
Permission depth	Higher, can touch system-level files	Lower, C: drive restricted	Whatever you grant
Best fit	Knowledge workers, teams	Individuals wanting remote control	Developers, tinkerers

The table covers the surface. What actually matters is that these three have fundamentally different philosophies about how much autonomy an agent should have.

WorkBuddy’s core: deliverables, not conversation

WorkBuddy is officially “an all-scenario workplace AI agent desktop workbench.” Translated into plain terms: its output is files, not text replies.

Tell it to sort the contract PDFs on your desktop, extract filenames and dates, and produce an Excel inventory, and it decomposes the task, reads the files, runs the analysis, and hands you an xlsx you can open. Traditional AI chat gives you “here’s how you could do that” and leaves the doing to you.

Four pillars support this:

Parallel agents with a unified board. Multiple tasks run simultaneously, with task decomposition, execution progress, and final deliverables all visible on a board rather than buried in chat scrollback. The design intent is clear: this is built for “I need to hand something to my manager,” not “I’d like to chat with an AI.”

Cloud and local coordination. Kick off a task and walk away. It keeps running locally or in the cloud. You don’t babysit a “thinking” spinner.

Claw remote control. Send a message from WeChat, WeCom, Feishu, DingTalk, or QQ to dispatch work to your computer.

Model matrix with auto-routing. Hunyuan, DeepSeek V4 (pro and flash, 1M token context), GLM, Kimi (2M token context), and MiniMax are all available. The /model auto mode runs a lightweight routing call before each request to assess task complexity, then picks the model and reasoning level automatically. For users who don’t track the differences between frontier models, that removes a real cognitive tax.

MCP protocol support rounds it out, letting it connect to third-party applications through a standard interface rather than only manipulating local files.

The timeline is what actually refutes the “wrapper” claim

An official denial isn’t very persuasive on its own. The chronology is harder to argue with:

• Second half of 2025: the CodeBuddy team is already building an open platform, SDK, and AI-driven framework
• Early January 2026: Claude ships Cowork; Wang builds a WorkBuddy prototype on the existing internal platform
• January 17: version 0.01 goes live for internal use — before OpenClaw broke out in China
• February: OpenClaw goes viral after Chinese New Year
• March 9: public beta launches; traffic far exceeds projections and Tencent scrambles a tenfold capacity increase
• June: monthly visits pass 20 million, reportedly more than the second and third place products combined

January 17 is the load-bearing date. The prototype’s direct inspiration was Claude Cowork, not OpenClaw, and the product is architecturally much closer to Cowork’s semi-automated shape.

One nuance worth stating precisely: it is compatible with OpenClaw’s Skills ecosystem at the architecture level, but it does not use the code. Those two claims get collapsed constantly. Supporting a skill format is not the same as borrowing an implementation.

QClaw is the one built on OpenClaw

QClaw comes from Tencent’s PC Manager team. Its positioning fits in one sentence: control your computer from WeChat.

Send a message from the couch and the machine at home starts sorting files, drafting email, processing spreadsheets. Data stays local, and a built-in sandbox monitors clipboard, camera, and other permission calls.

It is built on the OpenClaw framework, though with substantial optimization rather than a thin shell. The V2 release in April 2026 supports up to three agents working in parallel. Kimi-2.5, Minimax-M2.5, GLM-5, and DeepSeek-V3.2 ship built in, and you can supply your own API key.

QClaw’s real weakness is task length. Complex multi-step work tends to break partway. Architecturally it’s a lightweight task dispatcher, well suited to clear single-purpose jobs. Asking it to run “read data, analyze the table, generate charts, write the report, export a deck” will probably stall out.

Permissions are also narrower. Try to save to the C: drive and you get a permission error, and it won’t explain its execution logic. WorkBuddy is more permissive on both counts — system-level file access, and it tells you what it’s doing.

One install note worth keeping: on some systems QClaw won’t launch until you append --disable-gpu --no-sandbox to the shortcut target. It’s a Chromium compatibility issue underneath.

Why Tencent declined to chase full autonomy

This is the part I find most worth thinking about.

The industry is racing toward fully autonomous, fully managed agents, with OpenClaw at the far end of that spectrum. WorkBuddy deliberately sits in the middle. Wang’s phrasing: “We are absolutely not a purely managed automation product right now. We’re maybe halfway there.”

He went further and publicly cooled the hype: “A lot of people have been using OpenClaw, and it deserves a bit of cold water — it has run too far ahead of what this moment actually needs.”

Two concrete concerns drive that.

Loss of control. His example: “Without my authorization, the AI posts to Xiaohongshu on its own, summarizes content I published before on its own — that’s genuinely frightening. Evolving and learning by itself without the user’s consent.”

Unpredictable bills. He attributes OpenClaw’s cost to the fact that “it’s quietly maintaining an autonomous evolution record in the background, and the context keeps getting longer.” For non-technical users, that token bill is nearly impossible to forecast. WorkBuddy bundles model compute and sells credit packs, which eliminates the variable.

There’s a regulatory layer too. Around March 2026, China’s MIIT issued repeated warnings about security risks in crayfish-class products, and CNCERT specifically flagged prompt injection, plugin poisoning, and destructive mis-operations. Several universities banned OpenClaw on work machines, and multiple companies abroad blocked it outright.

WorkBuddy avoided “pass-through” by design — the pattern where a local working environment connects directly to a remote endpoint, which is exactly what regulators singled out.

The cost of that choice is explicit: if you want maximum autonomy, WorkBuddy won’t give it to you. A human is deliberately left in the loop.

On version numbers, I came up short

Being straight about the gap: this information isn’t publicly available.

What’s confirmable is only the milestones — 0.01 internally on January 17, public beta on March 9, official documentation last updated May 20. The current shipping version number isn’t published on official sources. Getting an exact figure means downloading the client and checking the about screen. I’d rather say that than invent a plausible-looking number.

QClaw has a few scattered data points: timed tasks arrived in 0.2.14, and three-agent parallelism came with V2 in April.

Credit allowances also conflict across sources — 5,000 credits during public beta, a 500-credit starter pack plus a 3,200-credit promotional pack, and 3,700 for new users all appear. Most likely these are different promotional windows. Check the official page for current terms.

A second point of confusion: WorkBuddy is not CodeBuddy

Some articles describe WorkBuddy as “a multi-model AI programming platform with IDE, CLI, and plugin form factors, $10 a month, competing with Cursor and Claude Code.”

That contradicts the official documentation, which describes document generation, data analysis, deck and report production, deep research, and batch file processing. Workplace tasks, not programming.

What those articles are describing is CodeBuddy — WorkBuddy’s sibling product and the source of its technical foundation. They share the same AI-driven architecture, but the scope differs: CodeBuddy writes code, while WorkBuddy extends the boundary to “anything you can do sitting at a computer.”

When you see WorkBuddy benchmarked against Cursor, skip it.

How to choose

Pick WorkBuddy if you need deliverables. Multi-step task chains, scheduled report generation, non-technical teammates who need zero-setup onboarding. Of the three, it’s the only one genuinely designed around “I have to hand something in.”

Pick QClaw if remote access is the actual requirement. You’re out, and you want to fire off a WeChat message to handle something urgent on the machine at home. Simple, well-defined jobs — don’t expect long chains.

Pick OpenClaw if you’re a developer willing to pay for control. Configure your own environment, cover your own tokens, own your own security boundary. In exchange you get full customizability and no vendor constraints.

Running two is also reasonable. QClaw for lightweight remote control while you’re out, WorkBuddy for batch processing at your desk. The features overlap but the situations don’t conflict.

One closing judgment

Tencent shipping three crayfish products at once looks like internal horse-racing from the outside. Internally they call it diversity. Wang offered a framing worth remembering: in the future each individual product may be just a touchpoint, and a user’s materials, preferences, and memory — the context — may flow between multiple crayfish products.

If that direction holds, agonizing over which one to pick matters less over time, and the real competition shifts to who holds and moves your context. It also explains why Tencent is comfortable shipping several at once: the bet isn’t that one product wins, it’s on owning the entry layer.

For now the decision criterion is simpler than it looks. Do you want an assistant that talks, or a colleague that hands you finished work? Answer that and the choice isn’t hard.