DeepSeek Harness is DeepSeek’s official AI agent client — an open-source desktop app that turns DeepSeek models into a working agent on your own machine. Open-sourced in August 2026 and released as a v0.2 public preview on September 29, 2026, it runs natively on macOS (Apple silicon, Apple-notarized) and Windows, with Linux supported through npm. Instead of chatting in a browser tab, you point Harness at a local folder and it works inside that workspace: reading your files, running code, producing documents, and keeping long jobs alive in the background.
Harness is an Electron app built around a local workspace and a plugin architecture — everything is a plugin, and agents can even write new plugins of their own. It ships with built-in Python and Node.js runtimes, with pandas, python-docx, python-pptx and openpyxl preinstalled, plus LibreOffice for in-app previews, so it can generate and edit Word documents, spreadsheets and presentations without extra setup. A background daemon keeps tasks running after the window closes, scheduled tasks and agent-team presets cover recurring and multi-agent work, and the GUI and the dsh CLI share the same data.
Harness itself is free and open source. Model usage is billed through your DeepSeek account balance when you sign in — no separate API key is required — or through your own API keys if you connect another model provider.
Harness suits developers, analysts and knowledge workers who want an agent that works directly on local files: automating document and spreadsheet production, running code against real project folders, or delegating long background jobs — all while keeping data on their own machine and choosing the model provider.