Add an unknown provider#

Harnez can use a provider that is not built into its model catalog when the provider exposes an OpenAI-compatible Chat Completions API.

Add the provider#

Add a named provider to ~/.config/harnez/settings.json:

{
  "providers": {
    "my-provider": {
      "type": "openai-compatible",
      "baseUrl": "https://llm.example.com/v1",
      "auth": "api-key",
      "models": ["my-coder-model", "my-chat-model"]
    }
  },
  "model": {
    "provider": "my-provider",
    "model": "my-coder-model"
  }
}

The provider name is your own stable ID. models contains the exact model IDs accepted by the endpoint. Keep API keys out of this file.

Restart Harnez after editing the settings file:

harnez server restart

Authenticate and select a model#

Open the provider's authentication flow:

/login my-provider

Choose API-key authentication and enter the secret. Harnez stores it in ~/.config/harnez/auth.json, separately from other providers.

Select a configured model from the picker:

/model my-provider

Or select it directly:

/model my-provider my-coder-model

Use a keyless local provider#

For a local server that does not require authentication, use auth: "none":

{
  "providers": {
    "ollama": {
      "type": "openai-compatible",
      "baseUrl": "http://localhost:11434/v1",
      "auth": "none",
      "models": ["qwen3-coder:30b"]
    }
  },
  "model": {
    "provider": "ollama",
    "model": "qwen3-coder:30b"
  }
}

No /login is required. Harnez sends the placeholder header Authorization: Bearer unused, which the local server must tolerate.

Compatibility requirements#

The endpoint must support streaming POST /v1/chat/completions. It must also support OpenAI-style tool calls for Harnez to use workspace tools. Responses API endpoints and automatic model discovery are not currently supported.

See Configuration for project-level overrides, validation rules, fast-cycle configuration, and credential locations.