Last few months my expenses on Cursor skyrocketed and I’m trying understand why.
Is there any way to see tokens spent in a given agent chat? I’m trying to understand what changed - its either Cursor not managing context well or me. I refactored a few large files I suspected could be part of problem, but really without seeing what agent pushes to LLM its probably impossible to guess.
A “simple” tokens count or even a cost near each chat request, or maybe just for each chat as whole would help a lot. I love Cursor but last month I stopped at $1000 and switched to Claude Max to save some tokens.
Hey, valid concern, this comes up a lot from power users.
You can track your overall usage right in the app. Open Cursor Settings and you’ll see how much of your included usage you’ve consumed. For a more detailed breakdown by model and tokens, check the web dashboard at https://cursor.com/dashboard/usage.
That said, per-chat or per-request token and cost counters in the IDE don’t exist yet. This is one of the most requested features on the forum.
A few tips to keep costs down:
Start a new chat for each task. The full conversation history gets resent as context with every request, so long threads add up fast.
Only attach the files you actually need in context.
Use cheaper models, or Auto, for simpler tasks. Model choice is the biggest factor in burn rate.
Keep Cursor rules concise and disable MCP tools you’re not using.
With Ultra you get 400 USD of included API usage, so hitting 1 000 USD means significant on-demand charges. If you haven’t already, set a spend limit at https://cursor.com/dashboard/spending as a safety net.
It doesn’t help much.
It could be really big if Cursor can intelligently advice on saving tokens.
As we don’t directly see how agent creates context - its not always easy to guess.
In my case for example agent decided to use Supabase MCP to execute very large SQL files (migrations, seed). It had to include full content of these files into tool call multiple times within session. Using CLI would be faster and easier option.
I suspect a kind of a hints for things that not even require any ai - “this MCP tool call is very big” kind of hints.
Another thing is refactoring code, I may be wrong but feels like simple renaming of a symbol can eat up lots of tokens and it takes a lot of time. Last few ties I literally switched to WebStorm just to rename and move things around quickly. Agent probably should have better tools for such common dev tasks to search codebase properly and understand ast not just grepping things around.
Interesting enough - you can just ask llm and it’ll answer the best it can:
I see the screenshot, it’s an interesting idea to ask the LLM itself to analyze context usage.
On your specific points:
MCP tool calls with large files: yeah, this is one of the main reasons for overspending. When an agent sends a full SQL file through MCP, all that content goes into the context window. Until we have built-in warnings for large tool calls, the best workaround is to explicitly say in your prompt or Cursor rules that SQL and migrations should be done via the terminal tool instead of MCP. For example: “For executing SQL files, use the terminal to run psql/supabase CLI commands instead of passing file contents through MCP tools.”
Rename and refactor via AST: agreed. Right now the agent mostly uses text search and replace, which is inefficient for large-scale symbol renames. Switching to WebStorm for this kind of work is a reasonable approach. In Cursor, you can use the built-in VS Code refactor F2 to Rename Symbol instead of the agent, it’s instant and free.
Thank @deanrie , appreciate all the suggestions !
I do use rules and agents.md to guide Cursor to use CLI over MCP - that seems to work well most of time.
Currently I just had to settle on a strange schema where I use Cursor up to $500-600 total and switch to Conductor (Claude Code) after. You can imagine my surprise seeing new agents management UI for Cursor 3. I root for Cursor, just not sure how you guys can do anything about Anthropic being able to sell $3-4k worth of tokens for $200.
Any update on this? Just giving people some per-convo visibility would help a lot!
On the MCP side, when can we expect things like tool search instead of just dumping the entire MCP tool description into every chat? It really makes some MCP servers unusable, and you can’t toggle them on or off if you’re instantiating a cloud agent via slack or linear
@Ingvar - glad the rules workaround for the CLI helps. On pricing and competing with Claude Max, yeah, that’s a real challenge. The team is working on improving how efficiently the agent uses tokens.
@ryanw - on per-chat token visibility, there’s no concrete update yet. It’s one of the most requested features and the team is aware, but there’s no ETA. I know that’s not the answer you want to hear.
On MCP, toggling individual MCP servers already exists, but only in Glass mode for now. For cloud agents via Slack or Linear, that option isn’t available right now.
On tool search, great point, thanks for the link to Anthropic’s approach. Right now all tool descriptions are loaded into the context for every request, and tool search would solve that at the root. I’ll pass this to the team. This lines up with this request too: Quick MCP toggle from chat
