QWE AI Academy Local agent. Real files. Spreadsheet preview. Work that keeps going after the window closes. No terminal babysitting. DeepSeek use Desktop landed with the v0.2 public preview in late September 2026 – same open-source agent that blew past ~240k GitHub stars since the August 13, 2026 open, now as a normal DMG/EXE.[[1]](https://rits.shanghai.nyu.edu/ai/deepseek-use-v0-2-desktop-app)
Before: npx @deepseek-ai/dsh web, a Node toolchain, a browser tab you could kill by accident. After: one installer. Also after: an agent that will rewrite disk paths you foolishly allow. The install is the easy half.
Why the old paths feel half-finished
Three routes still exist. Only one is “download and go” on supported hardware.
| Path | Setup friction | Best when | Watch-out |
|---|---|---|---|
| Official Desktop | DMG/EXE, bundled Node | Apple silicon Mac or Win10+ x64 | Developer preview; desktop analytics defaults |
npx @deepseek-ai/dsh web |
Need Node.js | Intel Mac, Linux, SSH, no-install experiments | Terminal + browser tab discipline |
| Community shells (Tauri/Electron wrappers) | Varies; often unsigned | Extra UI skins, Linux packages, version pinning | Not DeepSeek-affiliated; trust the publisher |
Apple silicon only for the official Mac build – macOS 13+ on the download page. Windows: 64-bit, 10 or later. Intel Mac and Linux stay on npx unless you gamble on a community wrapper.[[2]](https://www.deepseek.com/en/download/)
Cordis underneath: everything is a plugin, including the agent loop. Peking University and DeepSeek-AI wrote that up as spatiotemporal composability (arXiv:2608.25512). Desktop does not reinvent it. Electron shell, same web app, same sessions/tools/plugins.[[3]](https://www.deepseek.com/en/use/)
Install DeepSeek use Desktop on macOS and Windows
Downloads only from DeepSeek – the use product page or the download hub. “Latest” names on the public pages: dsh-latest-windows-x64.exe and dsh-latest-macos-arm64.dmg (download.deepseek.com). Community size checks in late September 2026 landed roughly 276-353 MB for the installers; installed footprint on macOS has been reported near ~1 GB in teardowns. Treat both as time-sensitive.[[4]](https://www.ai-indeed.com/encyclopedia/32237.html)
- macOS (Apple silicon): Open the .dmg, drag DeepSeek use into Applications, launch. First open may ask for folder access – grant only what the workspace needs.
- Windows (x64): Run the .exe through the NSIS wizard. Node and related tooling ship inside; you do not pre-install Node for the GUI path.
- Stay online on first launch. Host setup finishes, then the use UI opens in its own window.
- Auth: DeepSeek account (onboarding shows available credit) or paste a DeepSeek API key under Settings → Models. Other providers are key-based entries in that same models UI – not a consumer-chat SSO import.[[1]](https://rits.shanghai.nyu.edu/ai/deepseek-use-v0-2-desktop-app)
- Workspace: Empty test folder only (e.g.
dsh-sandbox). Not your home directory. Not a tree with secrets.
v0.2 adds a menu-bar action that can install the dsh CLI so plugin work does not need a separate Node/pnpm toolchain. Optional. Desktop, web, and CLI share one user-data tree – ~/.dsh on Mac/Linux, profile equivalent on Windows. Already used npx? Sessions can surface here after install. Backup that folder before a big upgrade; community notes flag one-time settings/session migrations (called out around the 0.1.7 line and still worth respect on later bumps).[[4]](https://www.ai-indeed.com/encyclopedia/32237.html)
A first task that proves the agent, not the chrome
Skip “hello world.” Drop a messy folder into the workspace: one PDF brief, one Excel export, one Markdown notes file. Then send something concrete:
You are working only inside the current workspace.
1) List every file with path and size. Do not modify anything yet.
2) Summarize the PDF in 8 bullets with page refs if possible.
3) From the spreadsheet, compute the top 5 rows by the amount column and write results to analysis.md.
4) Draft a one-page outline for a status deck in outline.md.
Stop and show the plan before any write. Ask before deletes.Three checks install pages never demo: read-only pass first, Office/PDF preview behavior (0.2 line improved this), write friction before anything hits disk.
Default model on the product page: DeepSeek-V4.1-Flash. App/use: free MIT. Tokens: you pay the provider. Off-peak Flash was listed around $0.15 input / $0.60 output per 1M tokens as of late September 2026 – cache hits much cheaper, peak hours roughly double. Live numbers move; check DeepSeek’s pricing page before you budget a long run.[[5]](https://api-docs.deepseek.com/quick_start/pricing)
Launch write-ups also waved a small time-limited login credit (around 6 yuan, short expiry into early October 2026). Treat it as possibly already gone. Do not design a workflow around promo balance.
Workspace risk and the settings people skip
Read SAFETY.md before the second prompt. Official line: experimental developer-preview software, not security-audited, not production-ready. Model output and third-party plugins can modify or delete files or spill credentials. Sandbox prompts help. They do not equal a security boundary.[[6]](https://github.com/deepseek-ai/deepseek-use/blob/master/SAFETY.md)
Pro tip: Dedicated sandbox folder. Enable only the plugins you need (Agent Teams, Voice, Scheduled tasks, and similar ship experimental). Review Creator-mode plugins before they stay installed. Disposable VM if the agent will touch anything you cannot lose.
Giving an agent your whole home directory feels efficient for about ten minutes. Then you remember it can delete.
The catch is the quiet defaults. Desktop builds have been reported to send product analytics on by default with no obvious UI toggle – community setup notes point at a product-analytics config flag; the plain npx web path does not share that telemetry story. Catalog churn in the 0.2 wave dropped some older third-party model IDs, so a saved model pick can fail after auto-update until you re-select. Close the window and background/scheduled work can keep running; a full quit can stop it, and the app warns when quit would kill in-flight jobs.[[1]](https://rits.shanghai.nyu.edu/ai/deepseek-use-v0-2-desktop-app)
Is a lighter desktop shell “enough” forever, or do power users drift back to CLI profiles for headless and SDK work? That split is still forming.
FAQ
Is DeepSeek use Desktop free?
Yes. MIT use, no seat fee. You only pay model tokens on whatever provider you connect.
Intel Mac or Linux – can I still run it?
Official installers stop at Apple silicon macOS and Windows x64. On an Intel MacBook you still use daily, install Node and run npx @deepseek-ai/dsh web – same UI, browser instead of Electron. Linux follows that path too. Community desktop wrappers exist for extra platforms; they are separate projects, so check signatures and release notes before you trust them with a workspace.
Account login vs API key – what actually changes?
Account login surfaces DeepSeek account models and available credit during onboarding. Per v0.2 notes, those account-backed sessions can use DeepSeek web search without a separate search key. An API key keeps billing on the platform key and still selects DeepSeek models under Settings → Models. Neither path pulls history from the consumer DeepSeek chat app – this is the agent use, not the casual chat product. People expect “sign in” to mean full product sync. It does not.
Next: download only from deepseek.com/en/use, create an empty dsh-sandbox folder, run the read-only list/summarize prompt above, and only then point the agent at a real project.