DSH is the new open-source challenger; Claude Code is the polished incumbent. An honest comparison of philosophy, model lock-in, extensibility, maturity, and cost.
Claude Code and DeepSeek Harness both put an agentic coding loop on your machine — and that is roughly where the similarity ends. One is a polished, closed product from Anthropic; the other is a four-day-old MIT-licensed runtime whose core premise is that nothing deserves to be a privileged core. Here is how they actually compare.
Claude Code ships as a finished product. It is excellent at what it does, and its extension story (skills, configuration, hooks) is exactly that — an extension story around a core you cannot inspect or replace.
DSH ships as a runtime built on the Cordis plugin kernel, where the model adapter, tools, sessions, sandbox, scheduling, and even the UI are individually replaceable plugins. The closest historical analogy is Eclipse vs. a commercial IDE: a minimal core plus an ecosystem where everything, including the parts you’d assume are load-bearing, can be swapped.
Claude Code runs on Claude models, full stop. In DSH, the model adapter is a plugin — DeepSeek’s models are the default path, but the architecture is model-agnostic by design. If your priority is negotiating prices and capabilities across vendors rather than maximizing one vendor’s polish, this is the structural difference that matters.
Claude Code’s extension units are skills and configuration layered onto the fixed agent. DSH’s units are plugins (fifty-plus, each owning one capability) and presets — small YAML files composing plugins, tools, and persona into a named agent. A preset is inspectable before you run it and shareable with a curl command. The practical consequence: in the dsh world, “an agent that reads code but physically cannot execute it” is a two-file fork, not a feature request.
You can browse what that ecosystem looks like in the preset catalog.
DSH’s sessions are append-only logs: every prompt, tool call, and context injection is recorded, and sessions can be resumed, forked, and replayed. Combined with the MIT source, you can answer “what did the agent actually do, and why” from artifacts you control. With any closed tool, you ultimately answer that question with the vendor’s word.
Claude Code is production-tested by a very large user base. DSH is a developer
preview (0.1.0-rc.7 at the time of writing) whose own README warns in caps that
compatibility-breaking changes are coming. Plugin ids have already shifted between
releases. If “it must work, every day, with support” is the requirement, the
incumbent wins on every axis except price and openness — today.
The counter-argument is timing: the ecosystem conventions forming now (plugin ids,
preset formats, the dsh-plugin GitHub topic) are the ones the community will
inherit, and early participants shape them.
Claude Code is billed through Anthropic subscriptions and API usage. DSH is free software; you bring model access (DeepSeek’s API is aggressively priced, and adapters for other providers exist). Whether that nets out cheaper depends entirely on your volume — but the option to change your mind later is priced in.
| You should run… | If… |
|---|---|
| Claude Code | you want maximum polish today, one-vendor simplicity, and support |
| DSH | you want model freedom, inspectable agents, replayable sessions, and to shape the ecosystem |
The options are not exclusive — plenty of teams will run the incumbent for daily work and a dsh install beside it for automation, experimentation, and cost-sensitive loads.
Either way, if you’re kicking DSH’s tires: start with the practical introduction, then create your first preset.
