Reverse engineer Cursor API calls

TensorZero reverse-engineered Cursor’s API calls to see its actual prompts and swap models under the hood

some interesting stuff in here (check first reply for the link)

- step-by-step setup to proxy Cursor through your own gateway (they share the code)

- cursor’s surprisingly minimal 642-token system prompts that power the whole experience

- discovered cursor uses a “less intelligent apply model” hierarchy for code edits

- live a/b testing between claude, gpt-4, gemini while keeping full cursor functionality

- you can optimize cursor for your specific coding patterns vs the general population

- technical challenges: cursor routes through their servers + electron cors issues they solved

https://www.tensorzero.com/blog/reverse-engineering-cursors-llm-client/