DeepSeek has released dsh, an open-source agent harness for developers. It is a Node.js-based runtime, licensed under the MIT license, that can be launched locally with: npx @deepseek-ai/dsh web It starts a web interface on localhost. The project is currently labeled a developer preview, and DeepSeek warns that the API may change significantly, including compatibility-breaking changes. The interesting part is its plugin-oriented design. The project treats many parts of the agent stack as replaceable components, including: Model adapters Memory backends Session logging Routers Output parsers Sandboxing Storage Scheduling The user interface These components use a common plugin interface and run on DeepSeek’s Cordis meta-framework. That design could make it easier to experiment with different models, tools, memory systems, and agent workflows. At the same time, it raises practical questions about plugin security, supply-chain risks, API stability, sandbox isolation, and whether the architecture will remain manageable as the project grows. The repository has attracted substantial attention since its release, although GitHub stars are not necessarily a reliable measure of technical quality or long-term adoption. What do you think? Does a highly modular plugin architecture make agent development more flexible, or does it introduce another layer of complexity that will be difficult to maintain?