dsh-desktop

English | 中文

An Electron desktop shell for DeepSeek Harness: double-click an icon and get a standalone window running dsh web — no terminal, no manual commands, no browser tab juggling. Region screenshot to ask the AI, system tray, session manager, auto environment setup.

Pure shell — it does not modify dsh itself. The server reuses your globally installed @deepseek-ai/dsh.

✨ Features

• 📸 Region screenshot → ask AI: one-click capture → full-screen overlay with real-time selection (GPU-accelerated, 8-direction handles to fine-tune) → auto-pastes into the chat → the AI answers from the image; the chat window hides itself during capture
• 🗂️ Session manager: Alt → File → Session Manager… lists every session, delete with one click (permanently removes chat history)
• 🪟 Standalone window: native desktop window loading the Harness Web UI
• 🍱 System tray: closing the window minimizes to the tray; right-click to show or quit
• 🔄 Easy reload: F5 / Ctrl+R / floating button — reloads the Web UI; frontend plugins (skins, layout, enhancements) take effect on reload without restarting the app
• 🧭 Auto environment setup: on first launch it detects Node.js / dsh and guides a one-click install — friendly to non-technical users
• 🔧 Configurable port: npm start -- --port 3081 when 3080 is taken
• 🧠 Smart reuse: reuses an already-running dsh service instead of starting a duplicate
• 🧹 Clean shutdown: only kills the dsh process tree it started — never an external service
• 🔒 Single instance: prevents double-launch conflicts
• 🎨 Hardened: contextIsolation + sandbox + no Node integration; external links open in your system browser
• 📦 One-click installer: npm run dist produces an NSIS setup

How it works

1. On startup, probes whether a dsh service is already listening on 127.0.0.1:3080;
2. If not, spawns dsh web --port 3080 automatically (logs to %APPDATA%/DeepSeek Harness Desktop/dsh.log);
3. Once ready, opens an Electron window loading the Web UI;
4. When the window closes, kills only the dsh process tree it started (an externally started service is never touched).

📸 Region screenshot

1. Click the 📸 button (bottom-right) — the chat window auto-hides
2. Drag to select an area on the full-screen overlay (live preview; drag the 8 handles to fine-tune after drawing)
3. Click ✔ OK (or double-click / Enter) — the chat window restores
4. The screenshot lands in the chat box automatically → type a question → the AI answers from the image

Pair with a vision bridge like ModLens so text-only DeepSeek models can read images too. Screenshots are also saved to ~/.dsh/screenshots/ (open the folder from the tray menu).

Prerequisites

• Node.js ≥ 22 (v24 verified)
• Globally installed dsh: npm i -g @deepseek-ai/dsh (0.1.0-rc.6 verified)
• To override the dsh binary path, set the DSH_BIN environment variable (e.g. point it at a source build of apps/cli/src/bin.ts)

Install & run

cd dsh-desktop
npm install
npm start

If the Electron binary download is slow or fails (common on some networks), use a mirror:

set "ELECTRON_MIRROR=https://cdn.npmmirror.com/binaries/electron/"
npm install

Use a different port (when 3080 is taken):

npm start -- --port 3081

On first launch: go to Settings → Models, enter your DeepSeek API Key, then pick a workspace.

Package an installer

npm run dist

Output lands in release/: dsh-desktop-<version>-setup.exe (NSIS installer).

Custom icon: drop an assets/icon.ico (≥256×256) and add "icon": "assets/icon.ico" under build.win in package.json.

Troubleshooting

• Launch fails / error dialog: check %APPDATA%/DeepSeek Harness Desktop/dsh.log; make sure dsh --version works.
• Port already in use: run npm start -- --port <new-port> instead.
• After upgrading dsh: just re-run npm i -g @deepseek-ai/dsh — the shell needs no changes.
• Screenshot not in the chat box: click the input and press Ctrl+V (the image is already on your clipboard).
• Closing the window doesn't quit: that's the tray — right-click the tray icon and choose "Quit".

Contributing

Issues and PRs are welcome! The project is tiny — all the code lives in src/.

License

MIT