Five minutes is generous. The whole ceremony is one command, one key, one folder — the part worth understanding is what the agent can do to that folder, and how the pieces fit.
Step 0: the prerequisite
Node.js, that’s it. Any current LTS release. Verify with node --version — if that prints something reasonable, you’re done with prerequisites.
Step 1: launch the Web UI
cd your-project
npx @deepseek-ai/dsh webThis is the official path from the README: no global install, no clone, no build. The process starts a local server and prints its address — http://127.0.0.1:3080 by default. Open it.
Two things worth knowing about what you just started:
- The working directory is the default filesystem location.
dshtreats the directory you launched from as home base, which is why you launch it from your project. - Nothing is selected yet. The Web UI deliberately starts with no workspace bound — that’s the next step, not a bug.
If you ever want the from-source path instead (you usually don’t, unless you’re developing the harness itself): clone, pnpm install, pnpm run build, pnpm dsh web. Same UI at the end.
Step 2: the API key
Open Settings → Models, paste your DeepSeek API key, save. That’s the entire configuration ceremony — and notably, no restart is needed: model routing becomes available the moment you save.
A common surprise for people arriving from other harnesses: the default assumption is the official DeepSeek API. Other providers and OpenAI-compatible endpoints exist, but they’re configuration, not prerequisites — see the official providers guide when you need them.
Step 3: choose a workspace
Click Choose workspace, add the project directory (the one you launched from), select it. Until you do, the session input stays unavailable — the harness refuses to chat about nothing in particular.
This is gotcha #1 for new users: the composer is locked on purpose. A workspace is a scope of file access and command execution; picking one is the moment the agent gets a body.
Step 4: your first session
Start a session and send something real:
Summarize this repository and identify its main packages.
What the agent may do from here, per the official guide: read and edit workspace files, run commands, delegate work, and maintain a plan. When an operation needs approval under your current permission policy, the Web UI asks first — you are always in the loop for anything consequential.
Spend your first session watching when it asks. The approval rhythm is the single best introduction to how this harness thinks about safety: everything is permitted in principle, nothing consequential happens unasked.
Step 5: know what you’re standing on
The honest fine print, straight from the README: DeepSeek Harness is in developer preview and iterating rapidly — there will be compatibility-breaking changes. Pin your version, expect churn, and treat surprising behavior as an issue to file (GitHub Discussions is the official venue) rather than a personal failing.
Where this road goes
You now have the loop running. Everything beyond it — memory, themes, IM bridges, multi-agent teams, TUIs — arrives as a plugin, mounted beside the rest rather than forked into the core. Two doors from here:
- Plugin 101 — write your first plugin in the time it took you to read this guide
- The resource map — 50+ verified ecosystem entries, every link checked
The five minutes are up; the agent is waiting in the tab you left open.