DeepSeek Harness

DeepSeek Harness (dsh) is DeepSeek's open-source, plugin-based agent harness. Connect its Web UI to Kourier to use deepseek-v4.1-flash on your Pro or Max plan.

You need Node.js, pnpm, and a Kourier API key. See Get an API key; no invite is needed. Your Kourier key is not a DeepSeek-platform key: configure a Custom model API, not the built-in DeepSeek card.

DeepSeek Harness is in developer preview and may introduce breaking changes. Review its safety notice before running it: the agent can edit files and run commands in your workspace.

1. Start the harness

From the project directory you want the agent to work in, run:

pnpm dlx @deepseek-ai/dsh web

The command starts the Web UI at http://127.0.0.1:3080 by default and opens your browser for a local launch. Keep the process running while you use the harness. If the browser does not open, visit the URL printed in the terminal.

2. Add Kourier as a custom provider

Open Settings → Models → Add model provider, then switch to Custom model API. Enter:

The provider ID is permanent, so use kourier from the start. Do not use https://api.deepseek.com or the Anthropic endpoint with this configuration.

Under Model catalog, choose Fetch available models, select deepseek-v4.1-flash, and choose Add selected. If discovery is unavailable, add the model manually with the exact ID below.

Check the model settings before saving:

The context window includes prompt plus output combined; see Model limits. Save or create the provider to persist the settings. Fetching the model list alone does not save it.

The harness stores the key in $DSH_HOME/.credentials.yaml and keeps a credential reference in its settings. The Models page shows a redacted descriptor after saving. Keep the credentials file private and out of version control.

3. Select a workspace and model

Click Choose workspace, add your project directory, and select it. A fresh Web UI does not select a workspace automatically, even if you launched dsh from that directory.

Select DeepSeek V4.1 Flash (Kourier) from the model picker and start a new session. Model selection also sets the default for new sessions; a session that has already sent a request keeps the model recorded in its own log.

For a first check, send:

Summarize this repository and identify its main packages. Do not modify files or run commands.

Confirm that the session uses the Kourier model and returns a response before asking it to make changes. Review permission prompts before approving file edits or commands.

Concurrency and subagents

Pro allows 3 simultaneous requests and Max allows 7. Each active model request, including a subagent's streaming request, uses a slot. If you receive 429 with Retry-After, wait for active requests to finish or reduce parallel work.

Troubleshooting

For upstream UI changes and advanced configuration, see the official model configuration guide.