Custom Providers

Connect an endpoint using the adapter that matches its actual API protocol. A similar model name does not make incompatible request formats interchangeable.

Provider value	Protocol	Default path
custom-openai	OpenAI-compatible chat completions	/v1/chat/completions
custom-gemini	Native Gemini content/parts	/v1beta/models/{modelId}:generateContent
custom-anthropic	Anthropic Messages	/v1/messages

All three have implemented providers. Compatibility still depends on the upstream endpoint’s supported options, tool schemas, streaming, and model capabilities.

Model configuration

{
  name: 'internal-chat',
  modelId: 'your-model-id',
  provider: 'custom-openai',
  endpoint: 'https://models.example.com',
  apiKeyRef: 'internal_model_key',
  metadata: {
    customProvider: {
      auth: { type: 'bearer' },
      // pathOverride: '/v1/chat/completions',
      // extraHeaders: { 'X-Org': 'example' },
    },
  },
}

Configuration is stored on each model row, so models can use different endpoints and credentials.

Authentication type	Additional setting	Request
bearer	None	Authorization: Bearer ...
header	headerName	Named header containing the key
query	paramName	Named query parameter containing the key

apiKeyRef accepts env:VARIABLE_NAME for an environment lookup or a system-vault secret name. Provider-specific fallback variables are CUSTOM_OPENAI_API_KEY, CUSTOM_GEMINI_API_KEY, and CUSTOM_ANTHROPIC_API_KEY. Prefer the vault for persisted credentials; query authentication can expose credentials in upstream access logs.

Native Gemini and Vertex

The Gemini adapter sends contents, systemInstruction, generationConfig, and tools.functionDeclarations, and reads candidates[].content.parts[]. Streaming uses the corresponding streamGenerateContent path. The former gemini-blocks-config request envelope is no longer supported; migrate proxies to a supported protocol or implement an adapter.

For Google Vertex AI, use the first-class vertex provider when service-account authentication is required. It handles JWT/bearer authentication and Vertex resource paths; a custom Gemini API-key configuration is not a substitute for that authentication flow.

Verify before routing work

Use the Models page’s connectivity test, then test the features you need: text completion, streaming, tools, and any media inputs. A connectivity success alone does not establish feature compatibility. Custom adapters do not provide a general-purpose embeddings implementation.

See Model Management and LiteLLM Proxy.