Integrate local Ollama model support in Cursor to enable AI code completion locally, without relying on cloud services.
Why ?
Some user could need to rely on an off-cloud model because they can’t always have access to internet, their internet is slow, or they work in sensitive environments. Additionally, some could be concerned about privacy
What is unique ?
While Cursor’s current AI completions rely on the cloud, integrating Ollama allow users to keep their data local, this could also reduce latency, as code completion is done on the user machine, or a server on their networks.
What would it contain ?
Integration with Ollama API: Embed Ollama API within cursor for the user to chose models they prefer.
Configuration and requirement: Allow the user to choose local or a home/distant server running Ollama.
Merge with:
It could be merged with current existing AI feature that are cloud based of Cursor, the user could easily switch if needed to the cloud service, it should be seen as an alternative when people can’t or don’t want to use Cloud-based models.
This is a crucial feature, as some organizations prohibit sharing code with external AI tools. It will particularly benefit developers constrained by internal privacy policies, especially those working in government agencies and large financial corporations.
Yes this is becoming the next level for AI enabled IDEs. Its clear that some of Cursors magic is in their own models, the vectorization etc. but at least to offload some of the remote API work when e.g. Claude is overwhelmed it would help to have access to local models.
absolutely necessary. and I’m happy to keep paying for cursor. it’s not about the money. it’s about being able to code with a decent model with no wifi.
and it should work with composer too!
Hey, unfortunately, Cursor’s servers do a lot of the heavy lifting when it comes to communicating directly with the LLM, so it’s unlikely there will be offline support in the near future!
Sorry to be the bearer of bad news here thought, I do sympathise that there are workflows with Cursor that could benefit from offline / local LLM support!
I guess would reveal a lot of the secret sauce of how the LLM calls are being formatted ?
Perhaps it could be integrated with the mention that it wouldn’t have refined capabilities we would have going through Cursor backend. I’m sure it would drive a lot of new users to the platform. Right now codebase embeddings and composer features give cursor an edge to tools such as Continue.dev but that might change in the near future…
might allow better custom model integration, then people can set up their proxy server with their deployed model and still use them with cursor, they might run R1 or future models on their own server with the distil size they want and use them with the cursor agent / composer?
Any chance you can adda feature where we set a public URL for the LLM API to use for requests on a per account/user basis?
I would pay more for a feature that lets you set your own URL, so I can reverse proxy to my own liteLLM or ollama instance. This way you can keep all the backend features the same, this would a a drop in replacement for the calls your backend already made to other LLM APIs.
The issues is thats if we do that, the whole point of the existance of this integration is none existant now (the whole point of this integration would be privacy)
