DeepSeek Harness tutorial: first run

A DeepSeek Harness tutorial from npx to the first coding task. Modes, API key, and what not to expect from v0.1.

Last updated: 2026-08-14

This DeepSeek Harness tutorial is unofficial. Commands come from the official README.

1. Install and open the Web UI

npx @deepseek-ai/dsh web

Browser: http://127.0.0.1:3080. Details: install DeepSeek Harness.

2. Add a DeepSeek API key

Create a key on the DeepSeek platform. In the Web UI settings, point the model plugin at deepseek-v4-flash first (cheaper) or deepseek-v4-pro.

DeepSeek Harness does not replace the model. Agent = Model + Harness.

3. Pick a mode

  • Standard — daily coding agent.
  • Code — let the model write TypeScript to chain tools.
  • Minimal — bench only (bash + editor).
  • Creator — inspect plugins and compose presets.

Read DeepSeek Harness modes before you live in Creator. Dynamic plugins currently vanish on restart.

4. Give it a real repo

Open a local project folder. Ask for something checkable: “run the tests and fix the first failure.” Watch the Trajectory / session log. Official claim: every run is traceable.

5. Add one plugin, not ten

The topic dsh-plugin is already noisy. Start from the plugin list. If you need a terminal UI, see TUI.

6. What this tutorial will not do

It will not turn v0.1 into Claude Code. As a finished coding agent, DSH is early. As a plugin kernel, it is the thing DeepSeek actually open-sourced. Comparison: vs Claude Code, vs Codex.