Install DSH Desktop First, Then Learn DeepSeek Harness
Windows/macOS beginners should start with the desktop client: install DSH, configure a model and run the first task. Linux or command-line users can use the npm/npx tutorials instead.
On this page
If you simply want to start using DeepSeek Harness, you do not need to set up Node.js, pnpm or a terminal first. The packaged DSH Desktop installer includes its normal runtime dependencies. DSH Desktop is independently maintained by the community and is not an official DeepSeek product; Linux users should use the npm/npx path below.
Step 1 — download the DSH Desktop client
For Windows and macOS beginners, Desktop is the shortest first-run path. The installer packages a fixed DeepSeek Harness version with the runtime it needs, avoiding a separate first-time Node.js, pnpm and local Host setup.
The packaged desktop app includes the normal runtime dependencies, so ordinary Windows/macOS users do not need Node.js or pnpm first. Desktop is a community project; the official npm/npx path remains available for Linux, developers and people who prefer the command line.
Using Linux or prefer the command line? Choose an OS guide
If you use Linux, want explicit control of your Node environment or simply prefer the official npm quick-start path, choose the matching operating-system guide. The core launch command remains npx @deepseek-ai/dsh web, with the Web UI served at 127.0.0.1:3080 by default.
Windows 10 / 11
Windows Terminal or PowerShell. Commands are shown for PowerShell.
Open Windows 10 / 11 CLI guide →
After Desktop or CLI opens DSH, complete these five steps
The two installation paths meet here. If one step fails, fix it before moving to the next step.
Configure a model provider
Open Settings → Models, save a DeepSeek API key or another supported provider's credential, then select one of its models.
Choose a test workspace
Select a disposable test folder before pointing DSH at an important project.
Run a read-only repository tour
Ask DSH to summarize the file structure or explain the project entry point, then compare the answer with the files on disk.
Make one controlled change
Send a small reversible task, review any approval request and verify the resulting file yourself.
Choose the next session preset
Use the documented preset capabilities to decide what tools a later session needs.
What a harness adds
A model generates an answer. A harness connects that model to a workspace and the tools needed to carry out a task. In DSH, an agent preset is the plugin composition for one session: its tools, prompt and capabilities.
Model
Choose the configured provider and model that will reason about the task.
Workspace and permissions
The selected workspace sets the session's working folder. Available tools and the active permission policy determine file access and which operations need your review.
Preset and plugins
The official project describes DSH as “Everything is a plugin.” A preset assembles the tools and capabilities available to a session.
That is why the first task should be small and inspectable: you are checking both the model response and the environment around it.
What you need before starting
For the DeepSeek path, create a DeepSeek API key. DSH also supports other providers listed under Settings → Models → Add provider; use the custom-provider guide for a gateway or service outside that catalog.
After DSH opens successfully
Continue through the learning path: configure the model, read repository facts, make one controlled file change and then learn the capabilities available in later sessions.
Configure another model provider
Each provider has its own standalone setup article with the official key page, route values, protocol, model-selection path and failure checks.
Advanced guides — only open these when you actually need them
The beginner path above is enough for a normal local DSH setup. These guides cover the four areas that create the most extra configuration and failure modes.
DeepSeek Harness official repository · DSH Desktop community repository. Desktop is independently maintained; DeepSeek Harness is also still a developer preview, so labels and behavior can change between releases.