DeepSeek Harness v0.2 发布官方桌面应用

DeepSeek has released an official desktop app for DeepSeek Harness (dsh), its open-source agent harness. The app ships with the v0.2 preview. Installers cover macOS (Apple silicon) and Windows (64-bit).

Is it deployable? Yes, today, as a preview. Download it from deepseek.com/harness or run npx @deepseek-ai/dsh web. DeepSeek warns that compatibility-breaking changes will follow.

What Shipped in v0.2

A harness is the runtime that turns a model into an agent. It reads files, runs commands and keeps a plan. The v0.2 preview targets everyday work as well as coding.

Built-in capabilities: commonly used tools for office work and development come preloaded; Plugin management: a new page installs, configures, enables and disables plugins; File and diff review: a right-hand sidebar previews generated files and code changes; Document work: users can send documents, spreadsheets or PDFs and ask for charts or slides; Automation Task: a plugin schedules recurring prompts, with run history and editable frequency

The desktop build also bundles the dsh command. Users no longer need a separate Node.js or pnpm install. Users can sign in with a DeepSeek account or add an API key. Per the rc.1 notes, account models get web search without an extra key.

Architecture: Everything is a Plugin

DeepSeek first open-sourced dsh in August 2026, under MIT. It runs on Cordis as described in the research paper ‘A Programming Paradigm for Spatiotemporal Composability‘. The model adapter, tool registry and agent loop are all plugins. Any layer can be replaced.

The harness is not locked to DeepSeek models. The provider guide covers third-party providers and custom OpenAI-compatible endpoints. The repository has passed 240,000 GitHub stars and 29,000 forks.

dsh also doubles as DeepSeek’s evaluation rig. The API changelog says Code Agent benchmarks ran on ‘DeepSeek Harness minimal mode.’ DeepSeek-V4-Flash-0731 scored 82.7 on Terminal Bench 2.1 in that setup.

4 Session Modes

Sessions run in 1 of 4 modes, per 36Kr’s hands-on and the release notes:

Standard: general work, and now the built-in default; Creator: describe a plugin in chat; the agent builds and installs it; PTC: programmatic tool calling, with Node code run in a separate process; Minimal: the lean loop DeepSeek uses for benchmark runs

So, What Changed Now

The v0.2.1-alpha.1 build adds an experimental Claude Code Mods compatibility layer. DeepSeek frames it as a test of whether the Mods API is broadly a subset of dsh plugins. It does not promise full compatibility. The build also adds a ‘Let Agent create a plugin’ entry, an optional Developer Tools bundle and a --public-url flag for reverse proxies. It removes runtime invariant plugins, a breaking change for some extensions.

Key Takeaways

DeepSeek Harness v0.2 ships official macOS (Apple silicon) and Windows x64 apps; The MIT harness treats every layer, including the agent loop, as a plugin; Creator mode lets the agent write and install its own plugins; It works with non-DeepSeek models through OpenAI-compatible endpoints; It is a developer preview; expect breaking changes between builds

Michal Sutter

Michal Sutter is a data science professional with a Master of Science in Data Science from the University of Padova. With a solid foundation in statistical analysis, machine learning, and data engineering, Michal excels at transforming complex datasets into actionable insights.

Michal SutterPrime Intellect Launches Prime Inference: Serverless and Reserved Serving for Frontier Open Models; Michal SutterMicrosoft AI Releases MAI-Transcribe-2-Streaming: #1 Real-Time Speech-to-Text Model on Artificial Analysis; Michal SutterOpenAI Releases GPT-6.1 Sol: Near-Astra Coding and Computer Use at One-Fifth of Astra’s Token Price; Michal SutterOpenAI Launches dots: Always-On GPT-6 Astra Agents That Work From Their Own Cloud Computers