DeepSeek Harness's 35K Plugins Reveal Agent OS Ambitions

In mid‑August 2024 DeepSeek Harness released v0.1 source code. By November 5 the project reported striking adoption metrics: 1.54 million daily active users calling the official API, consuming roughly 92 trillion tokens per day, and a community that had produced approximately 35,000 third‑party plugins. Six out of ten active users had installed or authored a plugin.

Plugin ecosystem extends far beyond coding tools

Unlike VS Code’s marketplace — dominated by language services, debuggers, and version‑control utilities — the four plugins highlighted in the official announcement demonstrate workflows outside software development:

dsh‑media‑desk (by creator woniuniuniu): a unified workspace for content creators that combines topic tracking, script writing, cover design, and storyboard planning in a single interface.

dsh‑guizang‑design (by op7418): packages web layout, rich‑text cards, and copy‑editing skills into a plugin with a persistent “My Works” panel; users can regenerate a liked style with one click. The author notes that the plugin system retains works, columns, and personal tone, yielding more stable output than stateless prompt‑only Skills.

agentero‑dsh (by research team Phil‑Fan): integrates the literature tool Agentero for PDF reading, highlighting, translation, and note‑taking, with an on‑call Agent to help untangle difficult passages.

dsh‑context (by Bilibili developer bowenliang123): visualizes the current context blocks, their evolution across turns, and inter‑Agent message passing — effectively a Chrome DevTools for model context.

All four plugins are installable directly from the plugin marketplace by package name.

Creator Mode: agents write plugins on the fly

The Cordis plugin architecture originally contained a dynamic plugin tool that let an Agent inspect available services and slots, generate code, and hot‑load it — but the feature was disabled by default and state was lost on restart. In early November the team shipped Creator mode (preview), surfacing this capability in the input bar. Users switch to Creator mode, describe a desired tool (e.g., a design workbench, a sandbox game, a 3D model visualizer), and the underlying DeepSeek‑V4.1‑Flash model queries the API, writes the plugin code, and installs it locally; the new UI appears in the workspace within seconds.

Strategic contrast with Coze and the Agent OS vision

On the same day (November 5), ByteDance’s Coze announced the shutdown of its enterprise tier and legacy plans. The article argues that early AI application platforms relied on rigid canvas‑based flowcharts because models lacked autonomous planning; as models gained task‑decomposition and self‑correction abilities, those flowcharts became liabilities. DeepSeek Harness inverts the model: it hands both UI and tooling to the model, letting agents create missing plugins at runtime and publish them to the community.

The official announcement states the goal explicitly: DeepSeek Harness aims to become the operating system of the Agent era, where “everything is a plugin” and even built‑in features can be replaced by third‑party plugins.

Real‑world constraints

The team acknowledges two immediate challenges: the 35,000 plugins vary widely in maturity, requiring work on API compatibility and security mechanisms; and the compute appetite is massive — 1.54M DAU burning 92 trillion tokens daily implies an average of ~60 million tokens per user per day. This economics explains why overseas incumbents restrict third‑party API access and herd users into closed clients, while DeepSeek chooses to subsidize plugin authors with official API tokens. The more indispensable tools users accumulate inside the window, the longer they stay, guaranteeing sustained model inference demand.