Hi — looking for clarification from people who’ve run into this, and ideally Cursor staff if possible.
I recently tried Oh My Pi (omp) with the built-in Cursor provider. It authenticates via OAuth / access token and talks to Cursor’s private agent endpoints (same general approach as the IDE/CLI client protocol, not the public Cloud Agents REST API). omp documents this as a normal provider (cursor / CURSOR_ACCESS_TOKEN) and I don’t see a ToS warning in their docs.
Separately, I’m considering a small local OpenAI-compatible proxy (for LibreChat / other OpenAI clients) that would translate POST /v1/chat/completions into that same Cursor Connect-RPC/protobuf backend — i.e. reuse the subscription chat models with streaming, similar to what omp already does natively.
Before I rely on either setup, I want to understand the policy risk.
What I found on Cursor’s side:
Terms of Service — Use Restrictions (§1.5) Cursor · Terms of Service
Includes restrictions on reverse engineering / deriving the underlying structure of the Service, probing the Service, and harvesting/extracting data from the Service.
Staff reply on the forum (unofficial proxy to private client endpoints / using subscription models outside official clients): Using Cursor Frontier Models like Composer 2.5 in External Harnesses (e.g. Codex)
Staff there say unofficial proxies that call private, non-public client endpoints go against those Use Restrictions, and that the supported path outside the IDE is the Cursor CLI / Agent SDK (not a DIY OpenAI-compatible bridge to internal APIs).
Questions:
Is omp’s Cursor provider considered the same kind of ToS issue as an unofficial proxy, or is it treated differently in practice?
Would a personal, local-only OpenAI-compatible proxy to the same private endpoints also be considered a violation?
If both are out of policy, is the only supported way to use Cursor models outside the IDE the official CLI/SDK (accepting agent-harness latency / behavior)?
Not asking anyone to help bypass anything — just trying to understand whether personal use of omp / a local proxy is something Cursor considers against the Terms, given that omp ships this as a first-class feature.
Hey, thanks for laying it all out with links. I’ll reply point by point.
The cursor provider 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:
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.
Thanks for your clear answer, so I did build a proxy, but to be honest it’s slow and doesn’t fit most cases, in human terms it’s like the agent doesn’t understand where he is
I just thought it’s a shame that my subscription has an api that I can’t find a use case for…
I hope that an endpoint can be done, yet I’m not sure it’s such a good idea for cursor, there are many many possibilities for cheap endpoint, maybe the point here is to rethink the included 20$ api for those that don’t need it…
Anyway I will delete my experiment and remove cursor from the Oh My Pi providers, I don’t want to do anything against tos, I like cursor.
Thanks again.
I’ve posted on reddit too, and somebody answered with interesting links
Thanks for the clarification. Just came across this thread. I’ve been using omp with cursor provider for a couple of days until now. Won’t do that again. Hopefully my account won’t be suspended/flagged. Btw, will the Cursor team support public API endpoint in the future?
Hey, thanks for checking the thread and stopping by. Stopping using omp with the cursor provider is the right move.
On your two questions:
I can’t comment on the status of a specific account here, since that isn’t something we handle on the forum. Since you’ve stopped using the third-party provider and you’re using the official clients, you’re within the ToS. If you’re specifically worried about your own account, email [email protected] and they can review your case.
There isn’t a public OpenAI-compatible /v1/chat/completions endpoint right now, and I don’t have a firm timeline for it. There’s an open feature request you can upvote and follow here: OpenAI-compatible /v1/chat/completions for Cloud API
For now, the official and safe ways to use Cursor models outside the IDE are the Cursor CLI cursor-agent, the Agent SDK or headless, and the public Cloud Agents API. Key detail: all of them always run the Cursor agent harness, they don’t expose the raw model.
