Is DeepSeek Harness Really That Powerful? A Full Installation and Hands‑On Test

DeepSeek Harness (dsh) is an open‑source agent harness released by DeepSeek, promoted with the slogan "Everything is a Plugin". Within a week the repository gained over 150 k stars, highlighting its rapid community adoption.

Installation

The setup requires a Node.js environment and a large‑model API key. A single command npx @deepseek-ai/dsh web installs dependencies, starts the service, and prints the default address http://127.0.0.1:3080.

Model Configuration

In the Web UI, the user selects the model (e.g., DeepSeek-V4-Flash) and sets the inference level to High. The built‑in model list also includes deepseek-v4-pro, and custom providers can be added.

Developing a Markdown Editor

A new workspace is created and the "Standard" mode (full‑featured coding agent) is chosen. Permissions are set to "Workspace Write" so the agent can modify files within the workspace while keeping other directories safe.

The following prompt 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、代码块高亮、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，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。

The agent first outputs a brief reasoning, then lists a seven‑step task list (scaffold project, install dependencies, implement Markdown rendering, write components, layout, build verification, write README) and proceeds to execute each step automatically.

During execution the agent runs npm install in the background, writes source files, builds the project, and finally runs npm run dev. The "Trajectory" page records every model message, tool call, timestamps, token consumption, and shows a 98 % cache‑hit rate, making post‑mortem analysis straightforward.

Result and Iteration

The generated application displays a split view with a left‑hand editor and a right‑hand live preview, syntax‑highlighted code blocks, table rendering, a status bar with word count and cursor position, and automatic saving to localStorage. Minor issues such as small toolbar icons and unsynced scrolling are addressed by sending a follow‑up prompt; the agent quickly updates the UI, enlarging the buttons and adding scroll synchronization.

Conclusion

With a single command and a well‑crafted prompt, DeepSeek Harness can produce a functional Markdown editor end‑to‑end, offering a reproducible execution trace and requiring almost no manual intervention. The "Everything is a Plugin" architecture makes extending or customizing the harness straightforward, and the experience is considered strong for an open‑source agent framework.