DeepSeek Harness's Officially Praised Plugins, Plus One That Automates Your Terminal
The plugin ecosystem turns DeepSeek Harness into a customizable agent OS, but the gaps these plugins fill reveal what agents still can't do out of the box: consistent design, token cost visibility, and direct terminal control. The terminal plugin in particular exposes a common pain point across agent IDEs that Western developers will recognize.
DeepSeek Harness's official API now draws 1.54 million daily active users, and third-party plugins have hit 35,000 within a few months of launch. The official team singled out four for praise: dsh-media-desk, a self-media workbench with five modules for inspiration, drafts, scheduling, review, and style; dsh-guizang-design, which bundles five Skills behind a unified design system so AI-generated web pages, PPTs, and social cards stop looking like blind-box outputs; dsh-context, a token forensics tool that breaks down where context windows are actually spent; and agentero-dsh, which wires the local-first research tool Agentero into the client for reading papers.
The reviews walk through installation and real tests. The media plugin needs a separate OpenAI-compatible image API key to generate visuals. The design plugin ships six magazine-style and four Swiss-style themes. The context plugin adds a /context command, a Fleet tab for sub-agent token distribution, and a 7-day billing estimate. The paper reader requires Node.js 24.9.0 or higher and keeps PDFs, highlights, and notes in a user-specified local folder.
The author then contributes a fifth plugin, paiwarp-dsh, to fill a gap: Codex, Claude Code, and DeepSeek Harness all lack direct terminal control. This plugin drives a real Bash terminal via natural language, reads on-screen output, checks exit codes, and records high-res PNG and MP4 artifacts. A "take over keyboard" button pauses the Agent mid-run so a human can intervene, then hands control back. Setup involves Python, Playwright, Chromium, and FFmpeg, and the bundled example runs a todo.py CLI through a full test scenario.
The token breakdown plugin exists because the default interface only shows a total and a percentage ring, hiding whether the cost comes from history, file reads, or tool definitions; that opacity is a real billing problem for any heavy agent user.
The design plugin's premise is that AI output lacks consistency, so a shared skill bundle imposes a fixed visual system; this suggests agent IDEs still can't maintain a coherent style across tasks without extra tooling.
The author's terminal plugin fills a genuine vacuum: none of the major agent IDEs support direct terminal use, so this workaround lets an agent act on live output and even hand control back to a human mid-run, which points to a missing primitive in the agent stack.