DeepSeek Harness Install: Node and Web UI
Check Node and npm, then run the official launcher pinned to 0.2.0-rc.2, the npm latest identity checked on 2026-10-04. Keep its terminal open, visit the printed local URL, select a workspace and configure a model. A reachable Web UI and a successful model request are separate checks.
Sharing includes the public link only, without local progress or private configuration.
Start with this command
This is the narrow command relevant to this page. Read the checks below before changing flags, credentials or profile patches.
npx @deepseek-ai/dsh@0.2.0-rc.2 web1. Check Node, npm and your project directory
Open a terminal in the repository you want the agent to inspect. Run the two checks below in that same terminal. Both commands should finish and print a version; a “command not found” or “not recognized” message means the runtime is missing from this shell’s PATH. Restart the terminal after installing or switching Node before testing again.
Do not assume that an npm launcher without an engines field supports every Node release. Check the official requirements for the exact dsh release you chose. Native PowerShell, WSL and a remote SSH shell can each resolve a different Node executable, even on the same computer.
node --version
npm --versionWhat to check: Each command prints its own version. Record the values; this checks the shell environment, not DSH startup or provider access.
2. Verify the pinned DeepSeek Harness launcher
Use an exact package version while diagnosing installation. An unpinned npx command may download a different release on a later attempt, making before-and-after comparisons unreliable. The launcher version check should complete before you add a model provider or optional plugin.
If npm cannot download the package, preserve the package-manager error and investigate the registry, network or proxy connection first. If download succeeds but the process exits during startup, preserve the first startup error. These are different failure stages; neither proves that your API key is wrong.
npx --yes @deepseek-ai/dsh@0.2.0-rc.2 --versionWhat to check: The launcher should report 0.2.0-rc.2. A package download error or a startup stack trace is a failed check; the exact text can vary by release and platform.
3. Start the Web UI and keep the server running
Run the command below from your intended project directory. The documented default local URL is http://127.0.0.1:3080; use the address actually printed by your terminal. Leave that terminal running while using the browser. Closing the browser tab does not stop the process, and closing the server process makes the local page unavailable.
A local launch can open the browser automatically. An SSH launch prints the host URL instead; forward the loopback service through your SSH client or editor and use its local forwarded address. If automatic browser opening fails but the server stays up, open the printed URL manually before changing the installation.
- For a deliberate manual browser launch, append --no-open.
- If the default port is already occupied, stop the unrelated process you own or retry with web --port 3081.
- For a local second-port check, open the URL printed for that invocation rather than the old browser bookmark.
npx --yes @deepseek-ai/dsh@0.2.0-rc.2 webWhat to check: A running server prints a local URL and the browser can load its Web UI. This is the expected documentation behavior; the site’s retained HTTP 200 test used rc.7, not the package pinned above.
4. Select a workspace and configure the model
In the Web UI, choose the workspace you intend to use. Check the selected directory before sending a task. Then open Settings → Models and configure the provider you actually want to call. Enter credentials only in your local DSH interface or its documented credential source; this website does not need them.
For a custom endpoint, prepare its base URL, supported protocol and exact model ID. Fetch available models is a convenience: a gateway can support inference while exposing no compatible model-list response. Save a manually entered model only when the gateway’s documentation identifies it. Select the saved model in a new session so an older session’s recorded model does not obscure the test.
5. Verify one small, inspectable task
Ask: “Find this repository’s test command and name the file that defines it. Do not change files.” Open the named file and compare the answer yourself. This exercises the selected model and workspace with a result you can check without starting a large coding task.
If the UI loads but the task fails, keep the exact error code and the selected provider/model identity. MISSING_CREDENTIAL points to credential resolution; UNKNOWN_MODEL points to the configured model route. A provider-side rejection has a different cause. Follow the corresponding diagnostic page rather than reinstalling Node after every failure.
- Record the dsh version, Node version, operating system and launch shell.
- Record whether the failure happened during download, startup, page loading or the first model request.
- Keep a sanitized error excerpt; exclude secret values and private workspace content.
Release identities and what the evidence establishes
| Channel or test | Version | Meaning |
|---|---|---|
| npm latest | 0.2.0-rc.2 | Default npm channel checked 2026-10-04; metadata does not establish a successful local run. |
| npm next | 0.2.0-rc.2 | A separately moving npm channel; pin its version explicitly when testing it. |
| Reviewed upstream documentation | dsh-v0.2.1-alpha.1 | The documentation contract used for these instructions, independent of the npm channel names. |
| Retained local runtime test | 0.1.0-rc.7 | Linux x64, Node 22.22.1: the Web UI returned HTTP 200 on the recorded test port. |
| Python SDK | 0.1.5rc1 | A separate Python package and installation path; its version does not identify your Node launcher. |
Frequently asked questions
Do I need a global dsh installation?
The official npm quickstart uses npx. The pinned commands on this page select a package version without requiring a global dsh command. If you already installed dsh globally, record which invocation you use so its version does not get mixed with the npx package.
Why does npx @deepseek-ai/dsh web keep timing out?
First identify the stage. A package download timeout requires a registry or network check. A startup error requires the first terminal diagnostic. A running server with an unreachable page requires the printed URL, port and SSH forwarding checks. A loaded page whose model request times out requires provider and endpoint checks. The same symptom does not justify the same fix.
Does an HTTP 200 response mean the model works?
It establishes that the local HTTP page responded. It does not establish provider authentication, model availability, workspace selection or a successful inference request. Complete the small first-task check separately.
Which version should I use to reproduce this site’s Linux test?
The retained test used 0.1.0-rc.7 with Node 22.22.1. Newer registry versions are shown separately. Use the exact historical version for that reproduction; use a current pinned version for a new test and record its own result.
Can I reuse the same instructions for Windows and WSL?
The Web UI workflow is similar, but the executable, path, environment and process live in the shell that launched it. Diagnose native PowerShell and WSL separately. Start with the matching platform guide and keep the workspace in the filesystem you intended to test.
First-party sources
Open the pinned or retrieved source before acting on a changed release candidate.
- Official DeepSeek Harness run guideOfficial documentation · retrieved 2026-10-04↗
- Official CLI behavior referenceOfficial documentation · retrieved 2026-10-04↗
- Official Web UI guideOfficial documentation · retrieved 2026-10-04↗
- Official model configuration guideOfficial documentation · retrieved 2026-10-04↗