Skip to main content
A custom model configuration identifies an external model endpoint and its credential references. It differs from selecting a managed model in the workspace catalog.
The dashboard has an implemented connection probe for an unsaved custom configuration. Current custom-model publication is rejected with custom_llm_execution_unavailable. A successful probe does not publish the configuration or start an agent conversation.

​
Configure the candidate

1

Choose the model mode

Open an editable agent’s LLM selection and choose Custom LLM. Keep the intended agent and workspace selected.
2

Set the endpoint

Enter Server URL and Model ID. The current probe accepts a public HTTPS Chat Completions base URL. Review endpoint compatibility.
3

Prepare credentials

Select a workspace secret or Secret environment variable for API Key, or None when the endpoint needs no key. Environment references resolve against production for the probe. Review request headers. Current probe policy rejects auth connections, API-version parameters, and dynamic headers.
4

Test the connection

Use Test connection with an approved provider account. This makes a real model request. Inspect the returned result and any error before changing other fields.
5

Preserve the candidate

Field edits update the editor’s local working copy; closing the sheet only dismisses it. For durable review, create a complete server draft and read it back. Do not treat a rejected Publish action as a saved version.

​
Configuration relationship

Under conversation_config.agent.prompt, custom selection requires llm: "custom-llm", a custom_llm definition, and backup_llm_config.preference: "disabled". Managed model selection requires custom_llm to be null. Preserve the rest of the selected snapshot when authoring a draft.

​
Frequently asked questions

No. Endpoint probing, draft storage, publication, and conversation execution are different operations.
The probe is dashboard-only. Its implementation is not a public API-key workflow or a reason to copy dashboard credentials into application code.