DeepSeek Harness vs Claude Code

This is not a scoreboard. They are not the same product.

What each one is

Claude Code is a mature terminal coding agent. You pay Anthropic (or point it at another API). The loop, tools, and UX are closed and polished.

DeepSeek Harness is an open-source plugin kernel (Cordis) plus an official coding-agent preset. Models, tools, sessions, sandboxes, the loop, and the UI are replaceable plugins. v0.1 is a developer preview with a Web UI.

Side by side

| | DeepSeek Harness | Claude Code | |---|---|---| | Job | Build or remix a harness | Sit down and ship code | | UI in v0.1 | Official Web UI + headless | Official TUI / IDE flows | | License | MIT | Closed product | | Default model | DeepSeek V4 family | Claude | | Stability | Will break | Production | | Extensibility | Everything is a plugin | Skills / MCP / hooks, not the kernel | | Cost shape | You bring a DeepSeek (or other) key | Subscription / API |

When to use which

• You want to write software today with a finished agent → Claude Code (or Codex, Pi, OpenCode).
• You want to swap the loop, the UI, or the sandbox without forking a 7k-commit agent → DeepSeek Harness.
• You want DeepSeek V4's price/perf inside a mature TUI → run V4 inside Claude Code / Codex. DeepSeek publishes those adapters too.

Early testers have said the same thing out loud: as a coding agent, DSH is behind Claude Code and Codex right now. As a framework, the plugin split is the point.

Sources

• Official DSH: github.com/deepseek-ai/deepseek-harness
• Official product page: deepseek.com/harness