Using DeepSeek Harness with SoAI

DeepSeek Harness can use SoAI as a custom OpenAI-compatible Chat Completions provider.

Configure model access

In the SoAI WebUI, select the Settings button at the right of the header, choose Settings, and open OpenAI. This tab appears only to an Administrator with API key management (OPENAI_API_ADMIN). Select Create key and copy the key when it is shown. Open Models and copy the exact model ID; this requires Model read access (MODEL_READ).

Configure the provider

Add the provider to the Harness settings.yaml:

llm-pi-ai:
  providers:
    soai:
      displayName: SoAI
      apiKeyEnv: SOAI_OPENAI_API_KEY
      api: openai-completions
      baseURL: https://SOAI_HOST:5090/v1
      compat:
        supportsDeveloperRole: true
        maxTokensField: max_tokens
      models:
        - id: YOUR_MODEL_ID
          name: SoAI model
          contextWindow: 65536
          maxTokens: 4096

agent-default-model:
  provider: soai
  model: YOUR_MODEL_ID
export SOAI_OPENAI_API_KEY='YOUR_SOAI_API_KEY'
export NODE_EXTRA_CA_CERTS='/path/to/soai-ca.pem'
dsh

Connect to SoAI in the DeepSeek Harness Web UI

  1. Launch the official browser interface with npx @deepseek-ai/dsh web and open its local URL.
  2. Open Settings > Models and select Add provider.
  3. Create a custom provider with ID soai, display name SoAI, base URL https://SOAI_HOST:5090/v1, and protocol openai-completions.
  4. Paste the SoAI API key, add YOUR_MODEL_ID to the model list, and set its real context-window and maximum-output values in the advanced model fields. Select Apply.
  5. Use the conversation's model selector to choose SoAI > YOUR_MODEL_ID, then send a short verification prompt.

Fetch available models can query the server's OpenAI-compatible model list before applying the provider. The browser host must trust the server CA. DeepSeek Harness stores the key through its write-only credential service; settings.yaml keeps only the generated credential reference.

MCP status and connection checks

DeepSeek Harness does not provide a user-facing remote MCP server configuration. Use the model-provider connection above. A SoAI MCP token is not used by this integration.

Verify and troubleshoot

Send a short prompt and confirm that the server records a Chat Completions request for the selected model. A provider error usually means the base URL, key environment variable, or exact model ID is wrong. A certificate error must be fixed by trusting the server CA. Do not disable TLS.

Related documentation