DSH vs Codex CLI

If you are choosing a terminal agent today, these two come up in the same conversation. They are not really competitors — they are answers to different questions — but the difference only becomes obvious after you have used both, which is an expensive way to find out.

The core difference, in one sentence

Codex CLI is a product; DSH is a runtime. Codex ships a coherent experience that one team designs and maintains. DSH ships a thin shell and expects you, or someone on GitHub, to supply everything else.

That single distinction predicts almost every other difference on this page.

Extensibility

DSH's entire thesis is that capabilities are plugins. The directory behind this site catalogues 715 plugins across 15 categories — interfaces, model providers, memory layers, desktop wrappers, browser control. That breadth is real and it is DSH's strongest argument.

The cost is consistency. Those 715 projects were built by hundreds of different people with no shared design review. Some are excellent. Many are weekend experiments: the median plugin in the largest category carries 2 stars, and about half of the catalogue sits below three. You are assembling a product from parts of unknown provenance.

Codex CLI's extensibility is narrower by design. You get fewer knobs and far fewer surprises. If you want to spend your attention on your work rather than on your tool configuration, that is a feature.

Model lock-in

This is the most consequential difference and the one least discussed.

Codex CLI assumes an OpenAI model. You can point it elsewhere, but the tool was designed around one family and behaves best there. DSH has no default assumption at all — the provider is a plugin, and 118 routing plugins exist to prove it.

What that buys you is optionality: when a cheaper or better model appears, you install a plugin. What it costs you is the thing Codex gets for free — a model that was tuned alongside the tool that drives it.

Ecosystem maturity

DSH's ecosystem is large and shallow. Large because contributions are cheap: publishing a plugin requires no approval. Shallow because most contributors move on. Of the 715 catalogued projects, 603 have been pushed since late August 2026 — a high number that mostly reflects how recently the ecosystem started, not how actively it is maintained.

Codex's ecosystem is small and deep: fewer integrations, but the ones that exist are maintained by people with a reason to maintain them.

Failure modes

Worth comparing, because you will hit both:

• DSH fails by silence. A misconfigured plugin usually just does not load. No error, no log, no clue. Our troubleshooting guide exists because this is the single most common complaint.
• Codex fails by refusal. When it will not do something — a sandbox boundary, a policy limit — it says so. Blunt, occasionally frustrating, but you always know where you stand.

Given a choice between a tool that fails loudly and a tool that fails silently, most people should take the loud one.

Who should pick which

Choose Codex CLI if you want to start working today, you are content with one model family, and you would rather not maintain a toolchain. The installed base is larger, the documentation is coherent, and nothing will surprise you in week three.

Choose DSH if you enjoy assembly, you need models other than the default, or you want a specific interface and cannot get it otherwise. The ecosystem is genuinely more capable at the edges — browser control, memory, desktop clients — and those edges are where interesting work happens.

If you are unsure: run Codex for a week, then spend a weekend with DSH. The switching cost in either direction is a few hours, and no article — including this one — will tell you as much as that.

Also worth reading

DSH vs Claude Code vs Codex covers the three-way comparison including Claude Code's permission model. For the ecosystem picture, see the category breakdown.