DeepSeek Harness Official Release: Hands‑On Test Proves Its Power

Introduction

DeepSeek Harness ( dsh) is an open‑source agent harness released by DeepSeek on August 13. It follows the principle "Everything is a Plugin" and is driven by the Cordis framework, allowing model adapters, tool registries, session logs, and the agent loop itself to be swapped as plugins. Within a week the repository attracted over 150,000 stars, making it one of the fastest‑gaining projects on GitHub.

Installation

Prerequisites are a Node.js environment and an LLM API key (DeepSeek or any compatible provider). The harness starts with a single command: npx @deepseek-ai/dsh web The command prompts for installation, downloads dependencies, and launches a web UI at http://127.0.0.1:3080.

Model Configuration

In the UI’s Settings → Model page, built‑in models deepseek‑v4‑flash and deepseek‑v4‑pro are available. Users can also add custom providers. After selecting a model and saving the API key, the harness is ready to run tasks.

Usage Example: Building a Markdown Editor

The author demonstrates the harness by creating a Markdown editor with Vue 3 + Vite. The workflow includes:

Adding a workspace that points to the project directory.

Choosing the Standard mode (full‑featured coding agent with file editing, shell access, web search, skills, planning, sub‑agents, and workflows).

Setting permission to Workspace Write so the agent can modify files within the workspace.

Selecting the model DeepSeek‑V4‑Flash with inference level High.

A detailed prompt describing layout, editing features, toolbar buttons, autosave, status bar, theme switching, and rendering requirements is submitted to the agent:

帮我在当前目录下开发一个Markdown编辑器，使用Vue实现（Vue 3 + Vite），要求如下：
1. 布局：左右分栏，左侧编辑区，右侧实时预览区，中间可拖拽调整比例
2. 编辑功能：
   - 支持标准Markdown语法渲染：标题、加粗、斜体、删除线、列表、引用、链接、图片
   - 支持代码块渲染，带语法高亮（可引入highlight.js）
   - 支持表格渲染
3. 工具栏：提供加粗、斜体、标题、列表、代码块、引用、插入链接、插入图片等快捷按钮
4. 实用功能：
   - 编辑内容自动保存到localStorage，刷新页面不丢失
   - 底部状态栏显示字数统计和光标位置
   - 支持导出为HTML文件
5. 外观：
   - 默认深色主题，右上角提供亮/暗主题切换
   - 界面现代简洁，预览区排版参考GitHub风格

完成后告诉我如何启动项目并打开预览。

Agent Execution

Upon receiving the prompt, the agent first emitted a brief reasoning step, then listed seven tasks: create project skeleton, install dependencies, implement Markdown rendering, write components, build layout, verify, and write README. It executed each task automatically:

Generated files and folder structure.

Ran npm install in the background.

Built the project and started a dev server.

The UI’s "Trajectory" page records every model message, tool call, timestamps, token consumption, and a cache hit rate of 98 % (LLM time ~6 min, tool time ~1 min).

Result Verification

Running npm run dev opens the editor in the browser. The final product features a left‑side editor, right‑side live preview, syntax‑highlighted code blocks, table rendering, a status bar with word count and cursor position, automatic saving, dark‑theme toggle, and all requested functionalities.

Iterative Refinement

The author requested larger toolbar icons and synchronized scrolling between editor and preview. The agent quickly updated the code, producing a more usable interface.

Conclusion

With a single installation command and one detailed prompt, DeepSeek Harness can generate a functional application with minimal human intervention, while providing a transparent execution trace for post‑mortem analysis. The "Everything is a Plugin" architecture proves effective for rapid AI‑driven development.