DeepSeek Harness: Everything Is a Plugin

What It Is

DeepSeek Harness is an open-source agent framework released in August 2026. The core idea: everything is a plugin. Models, tools, memory, storage, UI, and loops are all separate plugins you wire together via YAML config. No code changes to swap out any piece.

The framework is built on Cordis, a plugin dependency injection system originally from the Koishi bot framework. MIT license. The repo hit 180,000 GitHub stars within days of release.

How to Run It

npx @deepseek-ai/dsh web

That starts a local web UI. You configure your agent by editing a YAML file and restarting the process. No Python environment to set up, no LangChain dependency graph to debug.

What Makes It Different

LangChain and CrewAI require code changes when you add a capability. With DeepSeek Harness, you add a plugin to the config and restart.

Want a different model? Swap the model plugin. Want persistent memory? Add a memory plugin. Want a new tool? Add a tool plugin. The agent loop does not change.

This is the same pattern Hermes uses internally. The difference is DeepSeek Harness is a standalone framework you can build a product on top of, not just a personal agent runtime.

Who It Is For

Developers who want to build custom AI assistants or agent products without forking a framework. The plugin model means you can ship updates to individual components without breaking the whole system.

It is not ready for production deployments yet. The August 2026 release is a developer preview. The plugin ecosystem is small and the documentation is incomplete. But the architecture is clean enough that it is worth watching.

Compared to Hermes

Both treat agents as composable systems. DeepSeek Harness is more opinionated about the plugin interface and targets external developers building products. Hermes is a personal agent platform with more surface area (cron, multi-profile, Telegram delivery). They solve adjacent problems.