DSH vs Claude Code vs Codex

These three come up in the same conversations so often that it is easy to treat them as competitors in one category. They are not. Understanding why is more useful than any feature table.

The architectural difference that predicts everything else

Claude Code and Codex are vertical: a specific vendor's agent, tuned for software engineering, shipping features on a vendor's schedule, running on a vendor's models. You get a coherent product because someone made all the decisions for you.

DSH is horizontal: a runtime where the model, the interface, and every capability are replaceable parts. You get coherence only after you assemble it.

That single difference predicts almost every other comparison — setup cost, quality consistency, cost control, and the kind of person who is happy with each.

Where DSH genuinely wins

Model freedom

Point DSH at DeepSeek, at another vendor, at a local endpoint on your own hardware. Vendors of vertical agents do not offer this, for obvious commercial reasons. If your work involves sensitive data that cannot leave your network, this is not a preference — it is a requirement that only DSH satisfies.

Extensibility without permission

If you need a capability nobody has built, you write a plugin. No feature request, no roadmap negotiation, no waiting for a release cycle. In a vertical agent, an unbuilt capability is a wall. In DSH it is an afternoon.

Cost control as a first-class concern

Routing plugins send cheap tasks to cheap models. Context managers compress tool output before it bills. Cost meters give per-session visibility. A vertical agent gives you a monthly bill and few levers. For high-volume agent work this difference can be large.

Interface choice

Terminal, web, desktop, mobile bridge — all community builds over the same core. If your team works differently from the vendor's imagined user, you can change the interface rather than adapt your habits.

Non-coding agents

Research pipelines, document extraction, support triage, monitoring. Vertical coding agents have no opinion about these and you end up fighting the tool's assumptions. DSH starts neutral.

Where DSH genuinely loses

Setup cost

A vertical agent works ten minutes after install. DSH needs Node, a provider, a plugin set and some taste about which plugins are worth having. That is a real cost, paid in the least interesting kind of time.

Quality variance

Plugins are community code. Some are excellent. Many are abandoned. A few are actively unsafe. There is no review process and no warranty. You are trusting strangers, and you should read the source before installing anything that touches files or credentials.

No support contract

When something breaks in production, there is nobody to escalate to. For a hobby project that is fine. For a team depending on it, that is a risk to price in.

Churn

A fast-moving plugin ecosystem means breakage on upgrade. Plugins declare compatibility ranges and frequently fall out of them. Pin your versions. This is the hidden maintenance cost of the plugin model.

Documentation gap

Most of the ecosystem's documentation is in Chinese. For English-speaking developers this is a real obstacle — not to using the tool, which works fine, but to discovering what exists and understanding what a plugin does before installing it.

The practical decision

Choose Claude Code or Codex if your job is shipping code and you want the tool to disappear into the background. The vendor does the integration work. You accept their model, their interface, and their pace. That is a good trade for most professional developers.

Choose DSH if the agent itself is the thing you are building or tuning — non-coding agents, custom workflows, multi-provider cost optimisation, data that cannot leave your network, or a UI that does not exist yet.

Use both is not a cop-out. Many people run a vertical agent for routine coding and DSH for everything the vertical agent cannot do. The two occupy different niches and do not conflict.

A test that settles it in five minutes

Describe a task that requires a capability none of the three ships by default — reading a scanned invoice, querying an internal database, watching a directory and reacting to changes.

A vertical agent will need a workaround, usually a shell script glued to the side. DSH needs a plugin, and if that plugin does not exist, the architecture at least gives you a place to put it.

Whether that matters depends entirely on how often you hit the wall. If you hit it monthly, use the vertical agent. If you hit it weekly, the architecture is worth the setup cost.

What each one tells you about the market

One more observation worth making: the fact that DSH reached 259,000 stars while shipping almost no English documentation is evidence of how much demand exists for agent runtimes that are not controlled by a single vendor. The vertical agents are better products today. The horizontal approach is the one that produced an ecosystem.

That does not mean DSH wins. It means both patterns are now permanently established, and the interesting work is happening at the boundary — vertical agents borrowing plugin architecture, and DSH ecosystems shipping opinionated distributions on top of the runtime.

Next steps

If you have decided to try DSH, start with the install guide and then the plugin shortlist. If you are still deciding, the honest test is to try the vertical agent first — it is cheaper to evaluate, and knowing what a polished agent feels like makes the DSH trade-offs much easier to judge.