DeepSeek did it again. This time it’s not about a model, but about a tool that could reshuffle the market of coding agents: DeepSeek Harness (dsh) – an open-source framework where everything is a plugin. The GitHub repository gathered 166,000 stars in three days since its release (August 13, 2026). One of the fastest launches in AI project history.
What is it, exactly?
DeepSeek Harness is not a model and not another API client. It is an open-source agent harness – an environment that connects language models to files, terminals, tools, sessions, subagents, a permission system, and a user interface. In other words: a complete runtime for building your own coding agents.
The defining idea is simple but ambitious: everything, including the agent loop itself, is a plugin. The engine is built on the Cordis framework, whose design is described in the paper A Programming Paradigm for Spatiotemporal Composability. DeepSeek released the code under the MIT license – free to use, modify, and redistribute.
How does it work?
Getting started is a single command:
npx @deepseek-ai/dsh webIt requires Node.js ^22.19 or >=24.0. After launch you get a local web UI at http://127.0.0.1:3080 (with a --no-open flag for SSH sessions). The first run asks for a DeepSeek API key – note: a local UI does not mean local inference. The model still calls the API and you are billed by your provider.
A sizable community has already grown around the harness. Just look at the dsh-plugin topic on GitHub: there are design plugins (open-design – almost 90,000 stars), vision bridges for text-only models, desktop builds. The ecosystem is growing faster than any previous tool of this kind.
How is it different from Claude Code, Codex, and Copilot CLI?
So far, coding agents were closed monoliths: one company decides which tools the agent has, how the loop works, and which models can be plugged in. Claude Code, Codex, and Copilot CLI follow the walled-garden approach – they work, but you can only extend them within limits set by the vendor.
DeepSeek Harness flips that model: the agent is a runtime, not an application. Every layer – tools, permissions, sessions, even the decision loop itself – can be swapped like a building block. Add open source (MIT) and DeepSeek pricing, which is an order of magnitude lower than the competition. This is not a “cheaper Claude Code” – it is a different category of product: a platform for agents, on which anyone can build their own tool.
A sober note
Since this sounds like a bombshell, let’s be honest: the project is in developer preview. DeepSeek itself warns that it is iterating rapidly and compatibility-breaking changes are expected. The tested package is version 0.1.0-rc.6 – pre-release. Serious adoption requires pinning the package version and following the documentation.
But even with that caveat: 166,000 stars in three days is a signal that the community was waiting for an open, modular harness from DeepSeek. If the pace keeps up, the coding agent landscape could look very different by the end of the year.
What’s next?
We will soon publish a technical guide: step-by-step installation, the Cordis architecture, building your first plugin, and a performance comparison with Claude Code and Codex on real tasks. Leave a comment if you want us to test something specific.
This content was generated using artificial intelligence and reviewed by a human editor. Read the Polish original.
Leave a Reply
You must be logged in to post a comment.