DeepSeek Harness is an open-source agent desktop app, and its plugin architecture is the real story

What DeepSeek Harness is

Harness is DeepSeek's agent harness as a product. It runs as a desktop app on macOS (Apple silicon) and Windows (64-bit), or as a web UI you launch from code with npx @deepseek-ai/dsh web. Out of the box it handles four kinds of work: everyday tasks like organizing files, analyzing data and drafting documents and slides; coding across real repos, from exploring to fixes, features and tests; research with source citation; and background tasks like scripts and batch processing with progress tracking.

The app works on the files people actually have: Word documents, Excel spreadsheets, PDFs, Markdown, HTML and TypeScript and Python code, reviewed and refined through conversation. It also ships with the operational pieces agent products usually skip: scheduled recurring tasks, voice input, and developer tools that expose execution traces and per-call timing for every tool invocation.

Everything is a plugin

The architecture comes from Cordis, DeepSeek's "everything is a plugin" design, laid out in a companion paper, Exploring Super-agent Design: Towards a More Autonomous Agent System. Tools, skills and interface elements are all plugins. Official plugins cover agent teams, scheduled tasks, auto-approval review and the terminal, while agent loop, subagents and voice input sit behind experimental flags. A Creator mode lets you build new plugins through chat: DeepSeek's own demo builds, installs and verifies a floating Pomodoro timer plugin in about five minutes without leaving the conversation.

For anyone building agent features, that is the interesting commitment. Capabilities ship as composable, inspectable units with a plugin ecosystem around them, rather than as one monolithic agent you take or leave.

The open-source part, with an honest caveat

Everything is MIT-licensed. Two repositories matter: deepseek-ai/deepseek-harness carries the app, and deepseek-ai/dsh-superagent-prototype carries the research prototype, TypeScript throughout. DeepSeek's own README on the prototype is unusually candid: it is research code, not an official product, adapted for publication so it differs from the internal version, and some designs in the paper are future directions rather than shipped behavior. Treat it as a reference architecture to learn from, not a dependency to bet a product on, which is also the fair reading of a preview with evolving core plugins and APIs.

What DeepSeek Harness means for builders

Three things stand out if you build SaaS with agent features. First, the plugin and skill model is becoming the default shape of agent products, and it is the same shape independent developers already use; my Agent Skills case study is the reusable-instructions version of the same idea. Second, the pricing pressure on agent workflows keeps coming from the open side: a capable, MIT-licensed harness resets what buyers expect to pay for. Third, the observability work, execution traces and inspectable tool calls, is a reminder that agent features need debugging surfaces from day one, a theme across the agentic models in my GPT-6.1 Terra note.

FAQ

Sources

1. 01Product site and download: DeepSeek Harness (opens in a new tab)
2. 02Research prototype: deepseek-ai/dsh-superagent-prototype on GitHub (opens in a new tab)
3. 03Star velocity and analysis: The Week in AI, October 26 to November 1 - siliconnoon (opens in a new tab)

Elsewhere on this site

The selected work catalogue lists the SaaS products and AI builds behind notes like this one.