Quick Answer
DeepSeek Harness's Web UI is a local console at http://127.0.0.1:3080. It is not a "chat window" — it is the front-end of an agent runtime: you give a goal, the agent loops over tools until done, and the UI is where you observe and intervene.
Launch:
npx @deepseek-ai/dsh web
Open http://127.0.0.1:3080. To make it do real work, point the model backend at TeamoRouter:
export DEEPSEEK_API_KEY="sk-teamo-your-key"
export DEEPSEEK_BASE_URL="https://api.teamorouter.com/v1"
export DEEPSEEK_DEFAULT_MODEL="deepseek-v4-pro-free"
Below, each part of the UI and how to use it.
1. Session Management: Parallel Tasks
The session list (left or top) lets you:
- Create sessions — each is an independent agent context.
- Switch / rename — name by task ("fix payment bug", "refactor reports"), easier to find than by date.
- Run several in parallel — dsh supports parallel sessions, useful for a long agent loop in the background plus quick questions in another.
A useful habit: separate exploratory and deterministic work into different sessions. Let the agent roam freely in one; lock down exact steps in the other — avoids diluting context with irrelevant content in a long session.
2. Model Selection: One Key, Many Models
The model picker switches model IDs per session. With TeamoRouter wired in, the common choices:
| Model ID | Position | Free quota |
|---|---|---|
deepseek-v4-pro-free |
Flagship Pro, hard work | 200/day |
deepseek-v4-flash-free |
Cheap Flash, light work | 200/day |
deepseek-v4-pro |
Paid Pro, no account limit | — |
deepseek-v4-flash |
Paid Flash, no account limit | — |
Light work on Flash, hard work on Pro is the standard cost-saving combo (see V4 Pro vs Flash).
3. Tool Panel: The Agent's Hands
The tool panel lists what the current agent can call — the essential difference between a harness and plain chat. Typical tools:
bash— run shell commands: how the agent reads files, runs tests, installs deps.- File read/write — read, write, and edit files in the working directory.
subagent— spawn sub-agents to work sub-tasks in parallel, then collect results.web_search/web_fetch— search the web and fetch pages.ask_user_question— pause at a key moment and ask the user.
Understand these tools and you understand why dsh edits code rather than just talks about code: every tool call is a real read, write, or command execution.
4. Goals: Breaking Down Long Tasks
dsh models long tasks as Goals, supporting:
- Create / edit — split "build a feature" into "write endpoint → write tests → pass".
- Pause / resume — pause a wandering agent, correct it, resume.
- Mark complete — explicit closure for review.
For small one-off tasks, a one-line instruction works. For multi-step, trackable work, Goals are steadier.
5. A Complete Walkthrough
- Set the env vars (above) and start dsh.
- Start a session, pick
deepseek-v4-pro-free. - Give a real task:
Find package.json in the current directory, bump the version to 2.0.0, and add a release note line to the README.
- Watch the tool panel as bash and file tools are called in sequence, then the agent reports back.
Those 30 seconds give you the feel of "the harness turns a model into a worker."
FAQ
Why is 3080 blank / not connecting?
Confirm the npx @deepseek-ai/dsh web process is still alive. dsh's SDK server exits when stdin closes, so if you launch it with nohup ... &, keep a never-ending stdin attached.
The tool panel is empty / the agent says it has no tools?
Tools come from the agent preset. The default preset includes bash + file read/write + glob/grep; if trimmed, check whether the launch config loads the matching tool packages.
Is the Web UI local-only?
It binds to 127.0.0.1 by default — local machine only. Don't expose it to the public internet; the agent behind it can run shell commands, so an open port is dangerous.
Can I change the port?
Default is 3080. Change it via launch flag or config when needed, and update the access URL accordingly.
How do UI actions relate to the underlying agent?
The UI is an observation and intervention layer. The state you see (goals, tool calls, sessions) maps to real events in the agent runtime; stopping the UI doesn't stop a running agent session.
With TeamoRouter's free tier wired in, this console gives you 200 free V4 Pro requests a day. Register TeamoRouter for a key and start.