Install DeepSeek Harness (DSH)
This guide walks you from a clean machine to a running DeepSeek Harness Web UI. DSH is also searched as DS Harness or deepseek-harness — the package you launch is @deepseek-ai/dsh. Commands below match the public developer preview; if upstream changes flags, prefer the official GitHub README.
Prerequisites
- Node.js ^22.19.0 or >=24.0.0 (check with node -v) — older 18/20 often fails engines
- Network access to fetch the npx package (official npm registry if mirrors lag)
- A DeepSeek API key from the official platform when you want model calls
- A local folder you are willing to let the agent read/write — not your entire home directory on first run
One-liners by operating system
Windows (PowerShell)
If Node is missing, install LTS (winget install OpenJS.NodeJS.LTS or the installer from nodejs.org), reopen the terminal, then run npx @deepseek-ai/dsh web. Prefer starting the command from your project folder so the default workspace is nearby.
macOS (Terminal)
Install Node via the official pkg or Homebrew (brew install node), then npx @deepseek-ai/dsh web. Grant Full Disk Access only if you truly need it — first tests should use a disposable repo.
Linux / WSL
Use nvm or your distro’s Node 22/24 package, then npx. Avoid running the agent as root. On WSL, keep the workspace on the Linux filesystem for fewer permission surprises.
Quick start with npx
1. Launch the Web UI
In a terminal, run:
npx @deepseek-ai/dsh web
The CLI downloads the developer preview package and starts a local server. Keep the terminal open while you work.
2. Open the local UI
Visit http://127.0.0.1:3080 in your browser. You should see the DeepSeek Harness interface for chats, settings, and workspace selection.
3. Configure the model
Open Settings → Model (wording may vary by build). Paste your DeepSeek API key and select the model you intend to use. Do not commit keys to git.
4. Choose a workspace
Add the project directory the agent may touch. Start with a disposable sample repo if you are evaluating permissions and shell tools for the first time.
5. Send the first task
Ask for a low-risk action first (list files, summarize README). Confirm tool calls look right, then graduate to edits and tests. Use Trajectory views later to audit what the model saw.
Verify the install
Success looks like: UI loads on :3080, the model setting saves, a workspace appears, and a simple prompt returns a response with optional tool traces. If the UI fails to bind the port, check whether another process already uses 3080.
Install from source
For contributors and plugin authors, clone https://github.com/deepseek-ai/deepseek-harness and follow the repository install scripts. Source installs track Cordis plugin development more closely than npx alone.
Troubleshooting
npx is slow or times out
Retry on a stable network, or mirror npm per your region’s guidance. Confirm Node.js is on PATH.
API errors after chat
Re-check the key, model id, and account quota on the official DeepSeek platform — this community site does not issue keys.
Agent cannot see files
Re-select the workspace root and confirm OS permissions. Avoid pointing at your entire home directory on first run.
Next steps
Read the tutorials path for architecture and modes, skim plugins for dsh-plugin discovery, and keep the comparison page handy if you are migrating habits from Claude Code.