Agent · Doubao MarsCode · AI Programming

Doubao Is No Longer a Chatbot — It Now Runs Your Desktop, Browser, and Office Suite

By 程序员晓凡 · Aug 27, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

A mainstream consumer AI app now ships desktop-agent capabilities — file ops, browser automation, and SaaS connectors — with a free tier. That lowers the barrier for non-technical users to automate real computer work, and it pressures Western incumbents to match the integrated, OS-level agent experience.

Summary

Doubao's latest desktop update adds a Work Mode that lets the AI take direct control of a local or cloud computer — reading files, clicking through software, and using the browser. A project-based sandbox limits where it can read and write, and three permission tiers control how much autonomy it gets. Built-in Skills cover PPT generation, data analysis, and table formatting, while Connectors link it to Feishu docs, calendars, and group chats.

Phone-to-PC remote control is the standout feature: logged into the same account, a phone can dispatch tasks to a desktop that runs them inside a separate virtual desktop on Windows, leaving the user's screen free. Four tested workflows show it organizing a messy downloads folder by type, scraping real-time prices across JD.com, Taobao, and Pinduoduo to produce a comparison spreadsheet with a purchase recommendation, indexing an entire Feishu cloud drive into a searchable multi-dimensional table and navigation doc, and generating plus editing a 10-page personal intro PPT with targeted AI edits per slide.

Turbo mode includes a free usage tier. The advice for new users is to authorize local access on first run, create a project folder before complex jobs to contain the blast radius, and lean on the skill marketplace instead of writing fresh prompts every time.

Takeaways

— Work Mode switches Doubao from a chat interface to an agent that can read local files, open software, click the mouse, and browse the web.

— A project folder sandboxes the agent's file access so it doesn't roam the entire disk.

— Three permission levels — Always Ask, Confirm on Demand, Allow All — gate risky operations like editing external files or using the internet.

— Built-in Skills handle PPT creation, data analysis, and table formatting without needing a custom prompt each time.

— Connectors let Doubao read and act on Feishu documents, calendars, and group chats, plus a range of other third-party tools including programming-specific integrations.

— Phone remote control works over the same account: a phone command dispatches a task to the desktop, which runs it in a separate virtual desktop on Windows so the user's screen stays free.

— A single natural-language command made Doubao organize a messy downloads folder by file type, create subdirectories, and rename files.

— A price-comparison workflow scraped JD.com, Taobao, and Pinduoduo for a Sony WH-1000XM5, captured prices/discounts/delivery/self-operated status, built an Excel comparison sheet, and highlighted the best deal.

— A Feishu indexing workflow read every document in 'My Space', classified them, extracted summaries and keywords, built a multi-dimensional table index, and generated a hyperlinked navigation doc — all placed in a 'Personal Knowledge Index' folder. Saying 'update the knowledge index' later adds new docs incrementally.

— The PPT skill generated a 10-page personal intro slide deck after searching the web for the subject; per-slide AI editing can change text percentages and swap images on command.

— Turbo mode has a free quota; heavy users can upgrade to Pro. First-time local use requires granting pop-up permissions.

Conclusions

Doubao's desktop agent is not a developer tool — it's a consumer product that happens to do file ops, browser automation, and SaaS integration, which is a category Western consumer AI apps have barely touched.

The phone-as-remote-control for a desktop agent, running inside a separate virtual desktop, solves the screen-sharing friction that makes remote desktop tools annoying for quick task dispatch.

The Feishu knowledge-index workflow is effectively a personal RAG pipeline triggered by a single sentence, and the incremental-update command makes it a living system rather than a one-off script.

ByteDance's integration advantage shows: Doubao's Feishu connector gives it deep access to documents, calendars, and group chats that a third-party agent would need OAuth scopes and API wrangling to reach.

Concepts & terms

Doubao Work Mode

A desktop agent mode in ByteDance's Doubao app that can operate local files, use the browser, and run tasks on a local or cloud computer, sandboxed by a project folder.

Doubao Skills

Pre-built automation modules (PPT generation, data analysis, table formatting) callable inside Work Mode so users don't need to write a full prompt for repetitive tasks.

Doubao Connectors

Integrations that let the Doubao agent read and act on third-party services — Feishu docs, calendars, and group chats are the headline example, with many others available.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗