The streamed response must use SSE data with a nonempty assistant text delta. An ordinary JSON response or a stream containing only tool calls does not satisfy the current probe.
For example,
https://models.example.com/v1 illustrates a base URL. It is not an operational Threetone endpoint. Review the provider’s current Chat Completions documentation for the model identifier and response shape.
The probe resolves environment references using the workspace’s production environment. Review that value before testing; the probe does not select an environment from a future deployment.
Recording pending: Inspect a custom-model connection probe. Follow the written guide below.
Provider review
Cloudflare Workers AI, Groq, Together AI, SambaNova, or a self-hosted endpoint may expose different authentication, routing, model, and parameter conventions. Select a provider only after checking the required base URL, available protocol, credential scope, rate limits, and data-processing terms. Provider compatibility must be demonstrated; it is not guaranteed by this list. Keep a controlled probe result with the endpoint configuration and model identity. Retest after changing the endpoint, credential reference, model ID, or headers. A previous success belongs to that earlier configuration.
Screenshot pending: Custom LLM endpoint fields with synthetic values and the bounded probe result. Hide every credential.
Frequently asked questions
Does a schema alternative mean the probe supports it?
Does a schema alternative mean the probe supports it?
No. The broader authoring schema includes alternatives that current executable probe policy rejects. Use the requirements above for that probe.
Does the endpoint probe validate every tool and speech feature?
Does the endpoint probe validate every tool and speech feature?
No. It tests a bounded model request. Agent publication, tool behavior, voice processing, and live execution require separate support.