Cursor Agent always routes requests through its own backend servers (streamFromAgentBackend) regardless of custom base URL settings, so the resource_exhausted error is a Cursor-side quota issue — not your LiteLLM gateway — as confirmed by zero POST /v1/ e

Hey, this is a known issue. Sub-agents currently don’t inherit the custom OpenAI base URL and BYOK credentials. The main agent correctly uses your LiteLLM gateway, but sub-agents fall back to Cursor servers, which triggers resource_exhausted.

So your analysis is right. Cursor isn’t sending requests to your endpoint for sub-agent calls.

There are already threads about the same bug:

Sadly, there’s no workaround yet. The team is aware, and your report helps with prioritization. I’ll update the thread once there’s progress.