DeepSeek Harness

Developer preview

The official DeepSeek Harness project is a developer preview and may make breaking configuration changes. Follow the official project for current behavior.

Install and run

Run the web UI locally:

npx @deepseek-ai/dsh web

The default web UI runs on your local machine. Because the project is in developer preview, check the current Node.js prerequisite in its official README before installing.

Configure in the Web UI

Open Settings → Models → Add a custom provider.

Set the Provider ID to lowercase gateway.

Set the Base URL to http://127.0.0.1:11113/v1 and the API protocol to openai-completions.

Enter the platform key, use Fetch available models or manually add the complete model ID copied from the Models page, and set the default model.

Configure with YAML

$DSH_HOME defaults to ~/.dsh. Credentials can be provided through an environment variable or .credentials.yaml. For a shell environment, export:

export AI_GATEWAY_API_KEY="YOUR_API_KEY"

Put the following settings in the DSH settings YAML file:

agent-default-model:
provider: gateway
model: <FULL_MODEL_ID>

llm-pi-ai:
providers:
  gateway:
    displayName: AI Gateway
    apiKeyEnv: AI_GATEWAY_API_KEY
    api: openai-completions
    baseURL: http://127.0.0.1:11113/v1
    models:
      - id: <FULL_MODEL_ID>
        name: <FULL_MODEL_ID>

Alternatively, create ~/.dsh/.credentials.yaml:

AI_GATEWAY_API_KEY: YOUR_API_KEY

Protect the directory and credentials file:

chmod 700 ~/.dsh && chmod 600 ~/.dsh/.credentials.yaml

Never commit the credentials file. Vision models require input: [text, image] on the model entry; configure this only according to the capability shown on the Models page.

Verify

In the Web UI, start a new session, select the provider and model, send a text task, and then perform a read-only file tool action. Check Request Logs.

Troubleshooting

MISSING_CREDENTIAL: The credential name must match exactly.
404: The base URL must include /v1.
Protocol error: The protocol must be openai-completions.
Model error: Use the complete model ID from /models.
After an update: If the schema changes, check the official documentation.