What's this about?
DeepSeek released DeepSeek Harness (DSH) version 0.1.0 with the full source code under an MIT license — the most permissive kind, meaning anyone can use, modify, and sell software built on it. The design principle is that everything is a plugin: the model, the tools, the skills, the sessions, the sandboxes, the storage, the main loop, the scheduler, and even the user interface can each be swapped out or recombined. It ships with a web mode and a headless mode, and DeepSeek's Tianyi described it as an early preview aimed at people who build harnesses, not at end users.
What exactly is a "harness"?
The model itself only turns text into text. Everything that makes it behave like a working agent lives outside it: the code that decides which tools it can call, that feeds it your files, that runs the loop of think-act-observe, that keeps the conversation, that stops it when it goes wrong. That surrounding machinery is the harness. Claude Code and Codex are harnesses. The useful analogy is that the model is an engine and the harness is the rest of the car — and DeepSeek just published their car's blueprints and said every part is designed to be unbolted and replaced.
What's the community saying?
- **"But what *is* it?"** A common and upvoted complaint: the README is nearly bare beyond installation steps, and the landing page explains more than the repository does. Several people were puzzled that something this thinly documented hit HN #1.
- The Cordis plugin framework is the real content. Someone who read the underlying paper summarized it as adding hot-reload and dynamic enable/dispose to a plugin system — similar to what Pi agents do, but pushed further, all the way into UI components.
- Does a first-party harness actually help? A recurring genuine question: when the harness comes from the same lab as the model, does that pairing produce a real advantage over a third-party tool like Claude Code or Codex?
- Why is every harness written in Node.js? A widely-shared puzzlement about the ecosystem converging on JavaScript for agent infrastructure.
- Chinese developers wrote the substantive review. Baoyu (dotey) posted eight points of ordinary-user feedback — the startup path is too geeky for non-programmers with no Node.js environment, it's impressively fast, the Thinking output flickers because the Flash model is too quick and needs throttling, and Codex's multi-tab side panel should be copied. OrangeAI was blunter: "a toy of technological idealism — with models moving this fast, a system this open can't possibly be stable."