I asked for it, and Deepseek listened

I have been saying for a while that you should own your harness, so you are not dependent on a certain model. DeepSeek just open sourced their agent harness under MIT, and that is exactly what it lets you do, similarly to Pi and Open Code.

The beauty is that you can customize it the same way you would with Pi, maybe even more, but it is web-based, it looks way more pleasing to the eye, and you can switch models, or integrate any kind of model with a custom plugin if that is required.

What DeepSeek actually shipped

DeepSeek's own framing is: Agent = Model + Harness. The model is still the intelligence. The harness is the Cordis plugin layer around it: model adapters, tools, skills, sandboxes, the loop, the session log, and the UI. Those pieces mount and unmount, and you recompose them in configuration instead of forking the runtime. Nothing in that design ties the harness to DeepSeek's own models either. The provider catalog already covers Anthropic, OpenAI, Bedrock, Azure, Gemini, DeepSeek, and OpenAI-compatible gateways, and if something is missing you write an adapter plugin.

Same ownership I already want

That is the same ownership I already want with PI-Agent. One foundation under my control, so the model can change and the way I work does not have to.

I have not run this in production, and they say there will be breaking changes, but the idea is right.

Source is here: github.com/deepseek-ai/deepseek-harness