DeepSeek Harness Desktop: installation, plugins, and limitations

⚡ The gist in 5 seconds - DeepSeek Harness has received a desktop version: installers for Windows 64-bit and Mac on Apple, both builds are preliminary. - Node.js, a terminal, and pnpm are no longer needed to run it or to install the first plugin — everything is done through the interface. - Limitations: updates are installed manually, product analytics is enabled by default with no off switch, and environment variables on Mac are only picked up after a restart. ### 🔍 What was found Until the end of September, DeepSeek Harness was launched with a single terminal command: npx @deepseek-ai/dsh web, with the interface opening in a browser at a local address. It required Node.js, a terminal, and pnpm. On September 30, a desktop version was announced: an installer and account login. There are two builds: Windows 64-bit and Mac on Apple, both preliminary. The first launch offers two paths: signing into a DeepSeek account via the system browser or using an API key from platform.deepseek.com. The working window starts with choosing a working folder; there are four agent modes: standard, "creator," PTC, and minimal. Plugins are ordinary npm packages with a dsh.bundle section in package.json; they can be installed from the official npm registry, a mirror, or a custom address. The awesome-dsh-plugin catalog has amassed 17.8 thousand stars: 764 plugins for interface improvements, 288 for sessions and messages, 235 for spend and balance, 213 for models and providers, and 153 themes. Third-party shells appeared back in August: DSH Desktop by anywhere-labs has 29.9 thousand stars, and the dataelement project has 11.9 thousand. ### 💡 Why it matters The barrier to entry for DeepSeek agent scenarios has dropped significantly: instead of setting up an environment and a command line, you just download the installer, sign in, and add the plugins you need. The app keeps running when the window is closed — tasks continue to execute, and the session and draft are preserved. Two independent control levers have appeared: the sandbox (where commands can write) and the confirmation policy (what the agent must ask about); in the interface they are combined into a single switch with ready-made presets. The dsh team can also be installed via the Manage dsh Command menu and uses the app's environment, so separate Node and pnpm are not needed. ### 🧩 Context Previously, DeepSeek Harness existed only as a CLI tool: launched via npx, with the interface in a browser, and mandatory Node.js and pnpm for installing plugins. The desktop version removes this dependency but keeps some nuances: the app does not read system proxy settings, only the HTTPS_PROXY and HTTP_PROXY environment variables, and on Mac, variables from ~/.zshrc and ~/.zprofile are invisible to graphical applications — now a login shell is launched at startup and collects the environment within ten seconds. Updates are checked at launch and roughly every ten minutes, but are installed manually. The danger-full-access preset removes all sandbox restrictions. Environment variables on Mac are only picked up after the app is restarted.

🔗 Read on habr.com