Get it running in three steps, then stop reading

Step 1: Launch

Install Node.js, then run this in the project directory you'll use with DSH:

npx @deepseek-ai/dsh web

The terminal prints a full URL with a one-time token; open that. Hitting the bare 127.0.0.1:3080 without the token returns 401—that's not a dead service, and you don't need to check ports. After a restart, fetch a fresh URL; the old one stops working.

Step 2: Make it talk

Open Settings → Model, sign in to your DeepSeek account or paste your API Key; you can also add an OpenAI-compatible endpoint under Add Provider. After setup, send "please reply: connection successful"—if it returns, the model is live. Seeing the chat UI alone doesn't mean it's configured.

Step 3: Pick a workspace

Pick a local folder as the workspace next to the input box; without it the box is greyed out. One folder per task—don't select your whole desktop; everything it generates lands here.

Don't grant full permissions yet

Four common tiers: view-only, modify within workspace, full permissions, and the experimental auto-authorization review. Daily use only needs "modify within workspace": out-of-bounds actions are blocked and ask you first. Third-party plugins have their own data scope for code, network requests and cloud models; workspace permission is not a privacy switch.

Before upgrading

Back up your profile config and installed-plugin list first, then check the compatibility ledger. The community has reported plugin configs being overwritten after upgrades and the client not supporting rollback—don't test new versions on a live working environment.

See the compatibility ledger Look up by error, not by chapter

What we don't write here

We don't repeat architecture principles, full explanations of the four modes, or parameter tables.For systematic learning, read the official repo docs; for task-by-task copying, go to the recipes page.

• Official source repo
• Official docs
• DeepSeek Harness product page