DeepSeek Harness (dsh) vs Claude Code

At a glance

Openness

This is the cleanest difference and the one least likely to change. dsh is MIT licensed, developed in the open at github.com/deepseek-ai/deepseek-harness, and you can clone it, read how the loop is implemented, patch a behaviour you dislike and run your fork. Claude Code is proprietary: you configure it through its supported surfaces rather than editing what it does internally.

Whether that matters depends entirely on what you are doing. If you want an agent that works on your codebase, source access is a curiosity. If you are building a product on top of a harness, or you work somewhere that needs to audit what a tool does before it touches a repository, it is the whole decision.

How you extend them

dsh takes the plugin idea further than most tools do. Models, tools, skills, sessions, sandboxes, filesystems, the agent loop, orchestration and even the UI are all plugins, loaded into a shared runtime built on the Cordis meta-framework and composed through a cordis.yml loader config. There is no hard boundary between “the app” and “an extension” — replacing the sandbox is the same kind of act as adding a tool.

Claude Code extends through skills, hooks and MCP servers. Those are defined surfaces: you add capabilities and intercept events at supported points, and the core loop stays the vendor's. That is a genuine trade rather than a deficiency — defined extension points are far easier to keep working across upgrades, which is the exact problem dsh users are living through right now.

The practical consequence for dsh is that the ecosystem is where the capability lives. We index 632 repos carrying dsh topics, of which 348 show real plugin wiring; you can browse them all or scan them by category. That gap between topic-tagged and actually-wired is itself a useful signal about how young this ecosystem is.

Maturity, which decides most of it today

Claude Code is a product people already use daily. dsh was released yesterday, in the most literal sense of the word. Its own README leads with the warning that compatibility-breaking changes are coming, and that is not boilerplate: the loader format, the plugin APIs and the CLI options are all still being shaped, so a plugin that loads today may not load next week.

Nothing about that is a criticism of dsh — it is what a developer preview is for. It does mean the fair comparison is not “which is better” but “which risk do you want”. Choosing Claude Code means accepting a closed runtime. Choosing dsh right now means accepting churn, thin documentation and an ecosystem being assembled in public. Those are different kinds of cost, and only you know which one your project can absorb.

Models and where things run

dsh targets DeepSeek's V4 models; Claude Code targets Anthropic's Claude models. In both cases the harness is a local program that talks to a model over an API, so “self-hosting” mostly describes the harness rather than the model. There, the difference is real: you can run a modified dsh build on your own machines because the source is MIT licensed, which you cannot do with a proprietary client.

Because model providers are plugins in dsh, the architecture does not structurally bind it to one backend, and our directory already tracks a Models and Providers category. We are deliberately not claiming which providers work in which preview build — that is precisely the sort of detail that changes weekly and is best checked in the plugin's own repository.

Which one should you reach for

If you need an agent for real work this week, use the mature tool. That is the boring, correct answer, and it does not change because a new harness is interesting. If you are researching agent architecture, want to write plugins, need to read and modify the runtime, or are simply curious what an everything-is-a-plugin harness feels like, dsh is worth an afternoon — installing it takes one command. Plenty of people will sensibly run both.

If the underlying category is unfamiliar, our explainer on agent harnesses covers what these tools have in common before it covers where they differ.

Frequently asked questions

Is dsh a drop-in replacement for Claude Code?

No. DeepSeek Harness was released on 13 August 2026 as a developer preview whose README warns in capitals that there will be compatibility-breaking changes. It does the same category of work, but swapping an established tool for it means accepting a moving target in the middle of your workflow.

Which one is better?

That question does not have an honest answer this early. For work that has to ship, the established product is the safer choice. For building on the harness itself — replacing the loop, the sandbox or the model layer — dsh gives you MIT-licensed source that Claude Code does not offer. We have not run a head-to-head evaluation, and we would not trust one taken during launch week.

Can dsh run Claude models, or Claude Code run DeepSeek models?

dsh targets DeepSeek V4 models, and Claude Code targets Anthropic Claude models. Because model providers in dsh are themselves plugins, the architecture does not forbid other backends, and our directory tracks a Models and Providers category. We cannot verify which specific providers work in any given preview build, so check the plugin repo before relying on it.

Can I self-host either of them?

You can host and modify the dsh harness itself, because it is MIT licensed and builds from source. Claude Code is proprietary, so you configure it rather than fork it. Neither point removes the model API call at the centre of the loop, which still goes wherever you point it.

Do plugins written for one work in the other?

No. dsh composes plugins through a cordis.yml loader config, while Claude Code is extended through skills, hooks and MCP servers. MCP is a shared protocol, so tool servers built for it are not inherently tied to a single client, but that is a fact about MCP rather than a compatibility guarantee between these two products.