DeepSeek open-sourced a tool called Harness this week. Within seven days, community plugins under the dsh-plugin tag on GitHub have surpassed 900. What we care about more is the signal behind this number: the competitive focus of large model companies is quietly shifting from "whose model is smarter" to "who can wrap a more capable shell around the model."
What This Is
Harness (codename "Black Whale") is not a new model — it's an "execution layer" that wraps around large models. Think of it as a mobile OS: the AI brain is the chip, Harness is the system, and the various "plugins" are Apps installed on it. Its official slogan is "everything is a plugin": read files, write documents, send commands, operate other software — all assembled by plugins. Unlike most AI tools, Harness runs locally on your machine by default, with data never leaving the device — a plus for enterprise-sensitive data. The tradeoff: you need Node.js 22.19 or higher as a runtime, which is a non-trivial barrier.
Industry View
On the positive side, 900 plugins emerging in a week proves developers voted with their feet — this isn't a PowerPoint concept, it has a real ecosystem. Harness puts official and user plugins on completely equal footing, an approach aligned with Anthropic's MCP (Model Context Protocol, a standard for connecting models to external tools), but more open-source and more aggressive.
But we have to be honest about several risks. First, the cost of "everything is a plugin" is supply chain exposure — any third-party plugin can execute commands, and installing more means opening more backdoors on your machine; the article itself mentions security notes without elaborating. Second, the localization deployment threshold is high; most people who can actually use it are still programmers — the "everyone can write plugins" pitch remains at the marketing level. Third, without a strong model underneath, even the slickest shell is an empty husk; whether DeepSeek's model competitiveness can sustain this ecosystem remains to be seen.
Impact on Regular People
For enterprise IT: Internal automation scripts theoretically let non-technical roles "assemble" custom tools, but TypeScript knowledge is still required today. Real rollout to business units will likely take another year or two.
For individual careers: Mechanical work like daily status reports and weekly summaries will likely be partially taken over by AI within six months — the article's example of "reading yesterday's to-dos to generate today's draft report" is a textbook case.
For the consumer market: In the short term, ordinary users won't notice any change — this is still a developer's playground. But it foreshadows a direction: the AI in your future phone will be loaded with various third-party plugins, just like today's Apps.