01 · What WorkBuddy is

In one sentence

WorkBuddy is Tencent's desktop AI work agent: you describe what you want in a sentence, it breaks the job into steps, reads the local files you have authorised, calls tools, and hands back something you can use — a Word document, a spreadsheet, a deck, a web page, an image — written to the folder you chose.

The official line is "say what you need, it runs the task, it delivers the result." The valuable word in that sentence is runs.

How it differs from the AI chat you already use

Ordinary AI chat gives advice. WorkBuddy gives output. In daily work the gap shows up concretely:

	Ordinary AI chat	WorkBuddy
What it can do	Talk, suggest	Actually execute tasks
What it can reach	Whatever is in the conversation	Local folders you authorise, browsers, external services
Task size	Single-step, simple	Multi-step, can run for tens of minutes
What you get	A block of text	Files you can accept or reject
Cost of a mistake	You ignore the answer	It edited a file — you need a backup to undo it

That last row gets overlooked and it is the real dividing line. A helper that only talks costs you three minutes when it is wrong. A helper that acts can lose you files while tidying your desktop. This is why an entire chapter of this book is about permissions and the sandbox.

Four core capabilities

The official docs summarise the product as four capabilities. In plain terms:

It understands natural language. No command syntax to learn. "Sort the images in Downloads by date" is a valid instruction.

It plans and executes on its own. You give the goal, it sequences the steps. A "generate a sales analysis report" task walks itself through read → aggregate → chart → lay out → conclude, without you driving each step.

It handles multiple modalities. Documents, spreadsheets, decks, images, data analysis. Handing it a screenshot to turn into a table, or a recording to turn into minutes, is routine.

It operates on local files. This is the biggest difference from browser-based AI — it reads folders you have authorised and processes what is inside them in bulk.

Three work modes

When you create a task you pick one of three scene entry points. They set different defaults and output habits:

• Daily office (日常办公) — the default, for general workplace work: documents, data analysis, decks and reports, deep research, email and weekly reports, bulk file handling. This is where most people spend 90% of their time.
• Code development (代码开发) — generation, review, bug fixes, refactoring, understanding an unfamiliar codebase, requirements to shipped site.
• Design ideas (设计创意) — deeply integrated with Tencent's own Ardot design canvas: one sentence to a UI mockup, poster, deck or logo; jump into Ardot for pixel work; then generate front-end code from the finished design.

They are not three products. Switching changes the default capability mix and output preferences. See 06 Work modes and model tiers.

What it can actually do today

Pulling together the official practice cases and hands-on notes, the boundary looks roughly like this:

• Files and documents — reads local PDF, Word, Excel, PowerPoint, images and archives; produces Word, Excel, PowerPoint, Markdown, PDF. Bulk rename, sort by content, format conversion are all routine.
• Data analysis — reads Excel or CSV, aggregates along the dimensions you name, draws bar and line charts, writes a conclusion, lays the whole thing out as a report.
• Research and writing — searches the web, reads pages, produces a briefing or report with source links.
• Browser operation — with the Agent Browser Skill installed it scrolls, clicks, expands and screenshots, handling pages whose content only appears after interaction.
• Audio and video — the Local Whisper Skill transcribes recordings offline, then feeds straight into minutes; a downloader Skill pulls subtitles and audio.
• Code and small tools — generates a runnable local app from a description, and fixes it in the same conversation when it breaks.
• Tencent ecosystem — Tencent Docs, Tencent Meeting, Tencent Lexiang, QQ Mail and WeCom all have official connectors or Skills you can drive in plain language.
• Remote triggering — bind WeChat, WeCom, QQ, Feishu or DingTalk and a phone message starts a job on your desktop, with the result coming back to the chat window.

Where it sits in the field

Through 2026, desktop work agents became something every major vendor shipped at once: Tencent's WorkBuddy, ByteDance's Doubao Work, Alibaba's Qianwen Office, plus Anthropic's Claude Cowork and OpenAI's Codex. They belong to the same product category with near-identical interaction models — you state a goal, it plans and executes, files land on disk.

If you have used any of them, WorkBuddy needs almost no learning curve.

Three things make it distinctive:

Native depth in the Tencent office stack. Tencent Docs, Tencent Meeting, WeCom, Lexiang and Tencent Drive are first-class citizens — no wrangling APIs and auth yourself. For a team already running on those tools, that removes the most tedious part of the job.

An unusual number of mobile trigger paths. Six platforms: WeChat, WeChat customer-service account, WeCom, QQ, Feishu and DingTalk. For users in China, "message my desktop from WeChat" is far more reachable than the alternatives.

Concrete security boundaries. Default permissions, a program blocklist, a forced-approval blocklist, automatic backups, malicious-domain blocking — these are not marketing bullets, they are switches you can set individually in the Security Centre. Enterprise plans add VPC-isolated deployment.

Internal rollout numbers

Tencent says WorkBuddy began internal testing on 19 January and is now a routine AI work surface for more than 12,000 Tencent employees, across HR, admin, operations and sales.

The number matters not as a brag but as evidence: everyday usability has survived a real-scale trial, not just a good demo.

What it is not good for

Naming the boundary is more useful than listing features:

• Do not point it at your only copy of anything. Contracts, financial workpapers, sole-copy project material — duplicate first.
• Do not use it as a search engine for facts. It answers from model memory unless you pair it with the Web Search Skill and demand sources.
• Do not expect one shot at a big job. "Read this 50-page report and turn it into a deck for my boss" succeeds far more often split into three rounds than attempted in one.
• Do not let it make the big calls alone. Medical, legal, investment and compliance conclusions need a qualified human to check them — the official docs say so too.

Next

Now that you know what it is, go install it.