Models
Sidebar → Language Models → New
Fill in provider, API key, base URL, and model name. The same connection can be used by multiple agents.
Supported Providers
| Provider | Notes |
|---|---|
| OpenAI | Chat Completions-compatible OpenAI models |
| OpenAI Responses | OpenAI Responses API models |
| Anthropic | Claude 4.x, 3.x series |
| Google Gemini | Gemini 2.0 / 2.5 Pro, Flash |
| Gemini Image | Gemini image-generation models |
| Ollama | Local models via Ollama runtime |
| OpenAI-compatible | Azure OpenAI, Groq, Mistral, DeepSeek, Qwen, Together, etc. |
Any endpoint that implements the OpenAI chat completions API can be used as a provider — pick OpenAI-compatible and override the base URL.
Resilience
- Automatic retry with exponential backoff on transient failures (5xx, network, rate limit)
- Optional response caching with hit/miss metrics for repeated identical prompts
- Per-model token usage tracking visible in the Token Usage page
Advanced Options
- Context window / max tokens — override model limits when the provider does not report them clearly
- Max tools — cap how many tools are sent to the model
- Anthropic thinking — choose none, adaptive, or always-on thinking, with an optional budget
- Anthropic prompt caching — enable provider-side prompt cache hints
- Gemini API version — override the Gemini API version when needed
Embedding Models
Sidebar → Embedding Models → New
Embeddings are required for vector-based features (Notes, Wiki semantic search, Memory hybrid search). Supported: OpenAI, Google, Ollama, Cohere, VoyageAI.