Source checked October 11, 2026: the official DeepSeek Harness repository describes the project as an MIT-licensed developer preview and explicitly warns of compatibility-breaking changes. This is a documentation-based setup guide, not a hands-on benchmark.
Harness versus model versus client
| Component | Role |
|---|---|
| Harness | Coordinates the agent runtime and plugins. |
| Model provider | Performs inference under its own access, privacy and billing terms. |
| Web UI | Lets you interact with the local harness in a browser. |
| Plugins | Extend functionality; review their code, permissions and provenance. |
Launch the official local Web UI
Install a supported Node.js version and read the repository safety notice first. The command below downloads and executes the official npm package. Use a disposable folder without credentials or customer files; do not run it with administrator privileges.
npx @deepseek-ai/dsh webThe README documents a default local address of http://127.0.0.1:3080. A local launch opens a browser; an SSH launch prints the host URL. Add --no-open if you do not want automatic browser opening. Use the current official docs for provider configuration rather than copying an unrelated OpenCode config.
Before granting access
- Inspect the source and safety noticeFollow documentation from the official deepseek-ai repository, not a similarly named fork.
- Use a small test repositoryAsk for a read-only summary before permitting edits or shell execution.
- Check provider billingOpen-source software does not make the chosen model free. Set a provider budget and inspect one request.
- Review plugin changesInstall one trusted plugin at a time and verify the permissions and files it can access.
- Record the versionPreview APIs can break. Keep config backups and a known-good version before updating.
When to choose another interface
If you need an established terminal, desktop or browser workflow, compare the official OpenCode interfaces first. Wuwei is another separate desktop client; neither is a drop-in promise that DeepSeek Harness plugins will work unchanged. Compare required tools, permission controls, measured task cost and recovery behavior.
FAQ
Is DeepSeek Harness a new model?
No. It is an agent harness. Model availability, usage costs and data handling depend on the configured provider.
Is it production-stable?
The official repository labels it a developer preview with compatibility-breaking changes. Evaluate it in isolation before trusting important work.
Is the setup tested here?
The commands and local address were checked against the official README. We have not benchmarked a live authenticated task in this guide.