Setup guide · updated 12 Nov 2026
DeepSeek Harness, set up step by step
DeepSeek Harness (dsh) is DeepSeek's open-source coding agent: it reads your files, runs commands and edits code, the same kind of job Claude Code does. Everything in it is a plugin, including the model, so you choose the brain. This guide gets it running in about 10 minutes.
Read this first
The software is free (MIT license). The model is your choice: the DeepSeek API or Claude cost money per use, while a local model through Ollama costs nothing but needs a decent computer and is weaker than the big cloud models. dsh is a developer preview: it changes fast and can run commands on your computer, so start in a test folder (step 05).
01What you need
- Node.js 24 LTS (22.19 or newer also works).
- A test folder with a small project in it, not your real work.
- One model: a DeepSeek API key, or Ollama for a free local model.
02Start dsh
Open a terminal inside your test folder (dsh uses the folder you start it from as its workspace) and run:
npx @deepseek-ai/dsh web
It starts the Web UI at http://127.0.0.1:3080 and opens it in your browser. Keep it on this local address.
03Connect a model
Option A: DeepSeek API (paid per use, strongest)
- Create a key at platform.deepseek.com and add a little credit.
- In the dsh Web UI, open Settings → Models, paste the key and save. No restart needed.
Option B: a local model with Ollama (free)
Install Ollama, then let it set up and launch dsh for you:
npm install -g @deepseek-ai/dsh@latest
ollama launch dsh --model qwen3.8
Ollama installs dsh by itself if it is missing, so the first line is optional. To pick or change the model without starting, run ollama launch dsh --config.
Careful: a model name ending in :cloud (for example deepseek-v4-flash:cloud) runs on Ollama's servers, not on your PC. For a truly local, no-bill setup, use a model without :cloud. Small models run on most laptops but make more mistakes; bigger ones need lots of RAM or a good GPU.
Other models
Claude, OpenAI and any OpenAI-compatible server (vLLM, llama.cpp) plug in through model adapters. See the official docs for the exact settings.
04Run your first task
- Click Choose workspace and select your test folder. The chat box stays disabled until you do.
- Start simple:
Explain what this project does.
- Then a real job:
Fix the failing test or Add a dark mode toggle.
- Read each command it wants to run before you approve it.
05Stay safe
From the project's own safety notice: dsh is not security-audited and can run model-written code, load plugins and touch files it can reach.
- Use a test folder, a virtual machine or a container, not your main projects.
- Keep backups of anything it can access.
- Don't give it passwords, client data or keys you can't afford to leak.
- Only install plugins you trust.
06If something goes wrong
- Node version error: update Node.js to 24 LTS and run the command again.
- Can't type in the chat: you haven't chosen a workspace yet (step 04).
- Port 3080 in use: close the other app, or with Ollama run
ollama launch dsh -- --port 3081.
- Local model is slow or makes mistakes: try a smaller model for speed, a bigger one for quality, or use the DeepSeek API for hard tasks.
- Something broke after an update: it's a developer preview with breaking changes; check the repo's latest README.