QWE AI Academy Here’s the trap almost nobody leads with: search “DeepSeek use” and you can land on two unrelated products that both ship a dsh command. One is DeepSeek AI’s new agent runtime. The other is a community Python client for V4 API quirks. Mix them up and you’ll wonder why pip install never opens a Web UI.
The DeepSeek use developer preview is public under MIT – open source, no waitlist. This is not another architecture essay. It’s the boot path, the first-hour failures, and the cases where you should leave it alone.
What just shipped (and what it isn’t)
Compatibility will break. The GitHub README states that up front: developer preview, fast iteration, breaking changes expected. As of the mid-August 2026 drop, npm’s @deepseek-ai/dsh line sits around 0.1.0-rc.6 (community checks against that rc). Treat pins as mandatory, not polite advice.
On the official DeepSeek use page, capabilities show up as plugins – models, tools, skills, sessions, sandboxes, storage, loops, scheduling, UI – on the Cordis kernel. Design notes live in A Programming model for Spatiotemporal Composability. You get a local Web UI plus a headless runner. You do not get a frozen API surface.
Think of it less as a closed coding agent with a DeepSeek badge and more as a chassis you own. Closed tools hide the wiring. This one lets you swap it – which is the point, and also why setup has sharp edges.
Hands-on: run the developer preview in about 10 minutes
Engines first: Node.js ^22.19.0 or >=24. A generic “Node 22” install below 22.19 can fail the check. You also need a model API key. DeepSeek keys work without extra provider setup; catalog providers and custom OpenAI-compatible endpoints are documented in the official providers guide.
1. Boot the Web UI
npx @deepseek-ai/dsh webDefault listen address is http://127.0.0.1:3080. Start the command from the project directory you care about – that path becomes the process filesystem root – but chat still won’t work yet.
2. Models, then workspace (order matters)
- Open Settings → Models.
- Paste a DeepSeek API key and save. The route is usable on the next request; no server restart.
- Click Choose workspace, add your project folder, select it.
Until a workspace is selected, the session composer stays unavailable. That’s the onboarding rule in the official Web UI guide, not a glitch. Keys land in $DSH_HOME/.credentials.yaml; the UI is write-only for secrets after save.
Pro tip: One-shot without the browser –
dsh --profile headless "summarize this repo's package layout". Same runtime idea, no Web UI.
3. First task that proves the loop
Start a session and send something grounded:
Summarize this repository and identify its main packages.
List risky write operations you would need approval for before doing them.The agent can read and edit workspace files, run commands, and plan. Under the active permission policy, the UI asks before operations that need approval. Trajectory is where the product gets interesting: system prompts, reasoning, tool calls, results, and context injections go into an append-only session log you can resume, fork, search, or replay.
4. Optional: run from source
git clone https://github.com/deepseek-ai/deepseek-harness.git
cd deepseek-harness
pnpm install
pnpm run build
pnpm dsh webSource installs need the build step (pnpm via Corepack). Skipping build is a common self-own after a clean clone.
Common pitfalls to avoid
- Wrong package. Official path:
npx @deepseek-ai/dsh(TypeScript runtime).pip install deepseek-harnesspulls an unrelated V4 protocol adapter (e.g. HenryZ838978’s project). Same nickname, different job. - Composer looks dead. UI up, key saved – still pick a workspace before you type.
- Node too old within 22.x. Target 22.19+ or 24+.
- Images on hand-entered models. Custom models stay text-only until you set
input: [text, image]in$DSH_HOME/settings.yaml. DeepSeek’s own chat-completions route is text-only; config won’t flip that. Early errors you may hit:MISSING_CREDENTIAL,UNKNOWN_MODEL. - LAN bind fantasies. Community notes: the CLI rejects
--host 0.0.0.0. Local-first by design, not a multi-user service you casually expose. - Unaudited plugins + shell. Plugins showed up fast after launch. DeepSeek’s safe-use policy still applies: prefer a VM/container, don’t feed secrets, require human confirm for high-impact ops, only install code you’ve read.
One more sticky edge: after the first request, a session keeps the model recorded in its log. Delete that provider later and the composer can sit on “Select model” until you choose another route.
Modes, traceability, and what “good” looks like
Actually, mode choice changes what those paper numbers mean.
| Mode | What you actually get | When I’d pick it |
|---|---|---|
| Minimal | Basically shell + file editor | Fairer model benchmarks / less scaffolding noise |
| Code | Tools orchestrated via model-written multi-step programs | Long tool chains you want compacted |
| Standard | Full toolset (edit, shell, search, skills, planning, subagents) | Day-to-day coding agent work |
| Creator | Inspect runtime, trial plugins in memory, assemble presets | Building your own profile |
Scores that mention “use minimal mode” are not a promise of Standard-mode productivity. Minimal strips scaffolding on purpose. If your Standard run feels smarter than a paper’s Minimal number – or dumber – you’re measuring a different stack.
Full trajectory visibility is the piece early HN discussion latched onto (item coverage around the mid-Aug 2026 launch window). Closed agents often bury the chain. Here the event stream is meant to stay inspectable. Build the habit: fork a bad run, compare trajectories, tweak tools – don’t only retry the same chat.
When NOT to use this
Need API stability for a production agent product this quarter? Walk away. Core plugins and surfaces are still moving.
Want a hosted multi-tenant coding agent with SLAs? Wrong tool. This runs on your machine; you’re the SRE.
If you only need a thin terminal coding loop and don’t care about plugin composition, lighter runtimes will feel faster to learn. And if your threat model can’t tolerate a process that edits files and runs shell under approval policies, don’t “just try it” on a laptop holding prod credentials. Isolated environment first – the safe-use doc is blunt for a reason.
FAQ
Is DeepSeek use free?
MIT-licensed software. You still pay the model provider for tokens. Local compute is yours.
Do I have to use DeepSeek models?
No. Settings take DeepSeek keys, catalog providers such as Anthropic and OpenAI, and custom OpenAI-compatible base URLs. Handy when you want one model for bulk steps and another for harder judgment calls – just don’t expect a live session to silently reattach after you rip out its original provider.
Is this ready to replace Claude Code at work tomorrow?
As a learning and plugin lab: already useful. As a pinned production runtime: not yet. Upstream is explicit about breaking changes. Pin versions on purpose, keep profiles in git, and treat community plugins like untrusted code until you’ve read them.
If your team needs a frozen vendor agent next Monday, stay on whatever already ships tickets and run dsh beside it. Next action: throwaway git repo → npx @deepseek-ai/dsh web → key → choose that folder as workspace → ask it to map the tree → open Trajectory and read what the model actually saw.