Hey, thanks for laying it all out with links. I’ll reply point by point.
-
The
cursorprovider in Oh My Pi is in the same category as an unofficial proxy. It authenticates using your access token and calls our private, non-public client endpoints. That goes against the Use Restrictions in the Terms of Service, specifically reverse engineering or accessing the internal structure of the Service, §1.5. The fact that omp presents it as a normal provider and doesn’t show a warning in their docs doesn’t change anything on our side. omp isn’t affiliated with Cursor and isn’t an authorized client. Using your subscription outside official clients can trigger abuse enforcement, up to and including an account ban. -
A “personal, local-only” proxy doesn’t change the analysis. The issue isn’t who else uses the proxy. It’s the fact of calling private endpoints outside official clients. So yes, a local OpenAI-compatible bridge to those same endpoints is the same case as in point 1.
-
Supported ways to use Cursor models outside the IDE:
- Cursor CLI
cursor-agentUsing Agent in CLI | Cursor Docs - Agent SDK or headless Cursor TypeScript SDK | Cursor Documentation
- Public Cloud Agents API Cloud Agents API | Cursor Docs
Important point: all of these always run the Cursor agent harness, they don’t give you a “raw” model. We don’t currently have a public OpenAI-compatible /v1/chat/completions endpoint that you can point a third-party harness at. That’s an open feature request, not a shipped product. So yes, if you need automation outside the IDE, the official and safe path is the CLI or SDK, with the harness behavior and latency in mind.
If you build a workflow around the CLI or SDK and something doesn’t line up, message me and I’ll help.