Admin Dashboard: Models

Models are stored provider configurations used for either agent chat or document embeddings. Kind and provider type are separate: kind says what the model does; provider type says how SyntheticBrew calls it.

Model kinds

• Chat: used by agents. One can be promoted to the workspace default for agents without an explicit assignment.
• Embedding: used by Knowledge bases to vectorize documents and queries.

Kind cannot be changed after creation. Create a second model if the original kind is wrong.

Supported provider types

SyntheticBrew supports ollama, openai_compatible, openrouter, anthropic, and azure_openai. OpenRouter is accepted as a shortcut and saved as openai_compatible with its default base URL. OpenAI, Google Gemini, DeepSeek, Mistral, xAI, Z.ai, and other services with a compatible endpoint use openai_compatible.

Choose a supported provider

If the provider menu shows Google (Gemini) or Embedding Provider, do not select those shortcuts. For Gemini, choose OpenAI Compatible and enter Google’s OpenAI-compatible base URL. For embeddings, choose one of the supported providers and set Kind to Embedding.

Add and verify a model

1. Open Models, choose Add Model, and select Chat or Embedding.
2. Enter a lowercase URL-safe display name.
3. Select the provider and registry model, or type the provider model/deployment name.
4. Supply the base URL when required. Azure also needs its API version; embedding models need their vector dimension.
5. Save, open the model detail, and choose Verify. Saving does not run verification automatically.
6. For Chat models, optionally choose Set default.

The expected result is a verified model that appears in the relevant agent or Knowledge-base selector. API keys are write-only: the UI reports Configured but does not reveal the saved secret.

Operational cautions

• Ollama and compatible endpoints must expose the expected OpenAI-style routes.
• Tool-using agents require a model that reliably supports structured tool calls; test the exact workflow.
• An embedding dimension must match the provider response and existing vectors.
• Prompt-cache controls are not available in this form. Configure them through the model REST API or brewctl; see Prompt caching.
• Cloud egress policy can reject custom base URLs that resolve to private/internal addresses. Enterprise operators control their own approved endpoints.

Next steps

• Model selection guide
• Model Registry
• Knowledge
• Settings and BYOK