Cursor cloud agents seem perfect with one little downside. You cannot run an agent without a repo set up. I’m building a company that relies heavily on LLMs, and the simplest use case feels like an edge case for the cursor. I could not believe that using the cursor cloud api did not have a basic, old-school LLM input/output flow where you don’t need to set up a VM, repo, etc., just to access a model.
I’m a bit sad to have to switch to another LLM provider for the simplest use case… This is why I’m not using a cursor for everything, as I do with OpenAI.
If that happens, I’ll cancel my OpenAI and my Anthropic subscription.
Hi @Romain_G Thanks for reaching out! That’s definitely something for us to consider and we appreciate you sharing your thoughts here. We’re always looking for ways to improve our the quality of our service and experience that we deliver. I do a couple workarounds in the meantime to share you can do right now to before/if this becomes a feature.
If you specifically want cloud agents: spin up an empty GitHub repo (just an initialized main) and point Background Agents at it. The repo requirement is really just a place to branch from — it does not need to contain anything meaningful. That takes a few minutes, and then you can fire prompts at cloud-hosted agents normally.
If you want raw LLM input/output without a VM or PR flow**:** you can explore the Cursor SDK and REST API. With @cursor/sdk you can do Agent.prompt("...") against a small scaffold / dummy directory. Docs: Cursor SDK | Cursor Documentation . The /v1/agents REST endpoints work similarly for non-TypeScript clients.
I think cloud agents with a scaffold repo might be the best option for your use case.
Thanks a lot for the details, and just making it clear for anyone who reads it. The endpoint `/v1/agents` requires the GitHub repo, but the SDK does not. And yes, the /v1/agents work great, even if not directly an old school LLM input/output flow.