DeepSeek 开源的 Agent 运行容器：一切皆插件

DeepSeek Harness (dsh) is an open-source agent harness from DeepSeek AI, built as a plugin-centric runtime for AI agents. Instead of a monolithic framework, dsh treats every capability as a plugin and relies on the Cordis runtime for orchestration. It is aimed at developers who want to assemble, run, and observe agent workflows through a Web UI.

The core design statement is simple: everything is a plugin. dsh does not hard-code components; instead, capabilities are contributed and connected as plugins. The runtime relies on Cordis, whose design is described in the paper 'A Programming Paradigm for Spatiotemporal Composability.' This makes the harness highly composable and extensible, though the project is still iterating quickly.

DeepSeek Harness is explicitly in developer preview. The README warns that the project is iterating rapidly and compatibility-breaking changes will occur. Developers who plan to adopt it should be prepared to track updates closely and pin versions.

The project requires Node.js. When launched via npm, the command starts the Web UI at http://127.0.0.1:3080 by default and opens it in the default browser; pass --no-open to run without opening a browser. SSH launches only print the host URL. To run from a repository checkout, use:

git clone https://github.com/deepseek-ai/deepseek-harness.git
cd deepseek-harness
pnpm install
pnpm run build
pnpm dsh web

pnpm run build prepares the repository artifacts, and pnpm dsh web uses those artifacts without rebuilding.

Feedback and bug reports go through GitHub Discussions. Plugin authors can add the dsh-plugin topic to their repositories for discoverability. There is also a Discord community for real-time discussion. Contributions should follow CONTRIBUTING.md; new developers can start with the development guide, architecture docs, and AGENTS.md.

dsh fits developers building custom agent workflows who want a plugin-based, UI-observable harness. Because it is in developer preview, it is not yet a stable foundation for production systems. The README does not include benchmark numbers or a detailed feature list, so evaluation should be hands-on.