DeepSeek Harness Quickstart: Running in 60 Seconds

You need Node.js (for the npm package) and a DeepSeek API key — the same key you use for DeepSeek API calls^[1].

• Node.js installed
• DeepSeek API key (platform.deepseek.com)
• About one minute of patience

One command downloads and starts the harness with its Web UI^[1].

The npx command does double duty: it fetches the latest package and starts the web profile. If npx fails with a network error, install Node.js 20+ first or clone the repo and build from source (pnpm install && pnpm run build && pnpm dsh web)^[1].

npx @deepseek-ai/dsh web

# Then open http://127.0.0.1:3080 in your browser

In the UI: Settings → Models → add your DeepSeek API key. It is saved to $DSH_HOME/.credentials.yaml and becomes write-only after that^[3].

dsh defaults to DeepSeek models. If you want Claude, GPT, or a custom gateway later, the provider screen supports catalog providers and custom OpenAI-compatible endpoints — see the model configuration guide.

Key management note: once saved, keys are write-only — you cannot read them back from the UI or the credentials file. If you suspect a bad key, re-enter it in the Models screen rather than trying to edit the stored value.

Choose a workspace (your project folder), open a session, and ask for something real^[4].

The agent reads files, may run shell commands, and returns an answer. Open the Trajectory view afterwards to see exactly what it saw and did^[5].

If the agent cannot see your files, you chose the wrong workspace: the session uses the selected workspace as its file root, so pick the project directory in the workspace chooser before opening the session^[4].

Still nothing happening? Check the Models screen shows a saved key and that your DeepSeek account has credit — MISSING_CREDENTIAL and silent 401s are the two most common first-run blockers, both covered in the troubleshooting guide.

# In the session box:
"Summarize this repository's structure in 10 bullets"

# Or terminal one-shot:
dsh --profile headless "Summarize this repository in 10 bullets"

The three questions that come up in every first-run thread^[1][3][4].

Anything else: the error table is the fastest first stop, and the community Discord + GitHub Discussions are where the preview-era answers actually live^[1].

• Does dsh work without a DeepSeek account? — Yes: models are plugins, so you can configure any OpenAI-compatible provider. DeepSeek is just the default.
• Why can't the agent see my files? — You launched in the wrong directory or skipped the workspace chooser. Pick the project folder in the UI before opening a session.
• Is it safe to use with my API key? — Keys are stored write-only in $DSH_HOME/.credentials.yaml; only install plugins you trust, since plugins run with your credentials^[6].

The quickstart is the front door. From here the obvious next stops are^[1][6].

• Full tutorial — build a real agent setup step by step
• Plugins — add community extensions (616+ available)
• MCP setup — connect external tools
• Cursor integration — drive the harness from your IDE
• CLI reference — profiles, headless jobs, config inspection