I have a few user and project rules… not much else… Can it use so much Tokens?
This is on Grok 4.6 and on Plan mode.
On “Agent” mode it uses 21.4K
Hey @AndrewC, thanks for the report! What you’re seeing is expected.
The Context Usage panel counts everything Cursor sends along with the conversation, not just your message. Whenever you talk to an AI, there’s a system prompt shaping its behavior. We keep ours as lean as we can, but it isn’t free.
All of that has to be in place from the very first request. Without the tool definitions, the model has no way to read files, run terminal commands, or make edits, and without your rules and project context, it has no idea how your codebase is meant to work. It’s the difference between an agent that can do real coding work and a chat window that can only guess.
Two things worth knowing:
-
Most of that setup is served from cache after the first request, so follow-up messages in the same conversation are far cheaper than the raw numbers suggest.
-
You can shrink the baseline by trimming rules you don’t need (yours are about 6.7K tokens in that screenshot) and disabling any MCP servers you’re not using.