In the macOS terminal, type npx @deepseek-ai/dsh web. A web UI pops open at 127.0.0.1:3080. You toss it a line: “write me a web-based Tetris game” — and DeepSeek Harness starts writing code by itself.
Your first instinct is probably to line it up against Anthropic’s Claude Code and OpenAI’s Codex, and think: just another “AI programmer” a big lab rushed out to chase the trend.
If that’s what you think, you’ve missed the most breathtaking gambit of this whole wave.
They’re not even in the same race
Comparing the three is like asking why a Tesla can’t outrun a nationwide intelligent traffic-control grid — the question itself is wrong.
- Claude Code / Codex are the “car”: ready-to-drive AI programmers. The vendor welds the model, prompts, and toolchain into a black box, sells you the whole car, and collects your subscription and API fees.
- DeepSeek Harness (dsh) is the “foundation”: an agent harness built on the Cordis engine. Everything is a plugin — swap the model, write your own plugins, rewire the workflow.
One fights for users at the application layer; the other sets the rules at the infrastructure layer. That’s a strike from a different dimension: tool vs. foundation.
The official definitions hide all the evidence
How does each of the three describe itself? The original words are the most honest.
DeepSeek (verbatim from the GitHub README):
DeepSeek Harness (
dsh) is an open-source agent harness developed by DeepSeek AI.It uses an architecture where everything is a plugin, and is powered by Cordis, whose design is described in A Programming Paradigm for Spatiotemporal Composability.
Claude Code (one line from the official docs):
Claude Code is an agentic coding tool.
Codex (the official page doesn’t say it outright; ChatGPT summarizes):
Codex is OpenAI’s AI software engineering agent.
Look at the keywords and it’s obvious: one is a harness, the other two are tools / agents. The vendors never claimed they were the same kind of thing.
| Dimension | DeepSeek Harness | Claude Code / Codex |
|---|---|---|
| Essence | framework / infrastructure | end product |
| Model | neutral, swappable | locked to Claude / GPT |
| Control | with the developer | with the vendor |
Architecturally, they’re two different species
- Traditional CLI tools are heavily hard-coded — swapping a component means restarting the whole process. The Cordis engine uses “reversible effects + reactive coeffects” to let components plug in and out dynamically, with precise state rollback.
- In DSH, even the agent loop itself is a plugin. It’s not a black box kidnapped by vendor logic — it’s an open kernel you can reassemble, even one that lets agents self-evolve.
- The sharpest point: Claude Code can only run Claude; Codex only runs OpenAI. DSH is model-neutral — DeepSeek today, whatever you want to mount tomorrow.
The real question: why doesn’t DeepSeek just build a Claude Code competitor?
Because a product wins a moment; a foundation wins the era.
Think back to the cloud wars: AWS welded customers into its cloud with a wall of proprietary protocols — want to migrate away? The cost was brutal. Until Google open-sourced Kubernetes and turned “private technology” into “public standards”, the giants’ walls started cracking.
Today, Claude Code and Codex are replaying AWS’s role — using “closed-loop end products + their own LLM” to lock developers inside walled gardens.
DeepSeek didn’t jump into that red ocean to ship yet another CLI. It open-sourced the underlying engine and handed the “rocket-factory foundation” to developers worldwide.
No matter whose model runs on top, once the harness becomes the de facto standard, the walled gardens fall — logically, irresistibly. This isn’t open source beating closed source. It’s an antitrust move over who gets to write the rules of AI’s future.