Custom-model publication is currently rejected with
custom_llm_execution_unavailable. A successful connection test does not publish an agent or start a conversation.Prepare the candidate
Cloudflare documents the account-specific base URL and bearer-token authentication in its API compatibility reference. Use a text-generation model, not an embedding model. Do not paste the full
/chat/completions endpoint into Server URL; Threetone appends that path.
Run a controlled probe
1
Confirm account access
Check the account ID, model access, and token permissions with the account owner. Keep the token in approved secret storage. Do not place it in an agent prompt or a public example.
2
Enter the custom configuration
Open the agent’s LLM selection, choose Custom LLM, and fill the fields above. Follow the custom-model workflow. Secret environment references resolve against production during the probe, so verify that value before testing.
3
Test and inspect
Test connection makes a real provider request and can consume provider usage. The probe requests a streaming text response with
stream_options.include_usage: true. Verify that the selected endpoint accepts this request and produces assistant text; ordinary JSON or a tool-only stream does not pass.Frequently asked questions
Can I use a Workers binding as Server URL?
Can I use a Workers binding as Server URL?
No. The probe calls a public HTTPS API base URL. A Workers binding is a different execution interface.
Does Cloudflare Responses API support make it usable by this probe?
Does Cloudflare Responses API support make it usable by this probe?
No. The current Threetone probe accepts Chat Completions only. Provider support for another API does not change that boundary.