Apologies if this isn’t the most polished. I’m cross-posting my feedback brain-dump from the stripe subscription cancellation page:
Why was the @codebase feature removed? It was the main feature that I used cursor for. I tried the new agent feature where it tries to grep for the code it wants and:
It’s way worse most of the time vs @codebase
It’s cumbersome to have all of these low information-density(“grepped for XYZ”) messages added to the conversation.
By all means, explore agent-based search, but don’t take away the stuff that works while you explore that.
Unfortunately, cursor/vscode don’t work very well for java so I’m stuck using both intellij and cursor. There isn’t a clean way to open a file in cursor from intellij and vice-versa. As a workaround, I setup cursor as the os-default program for opening java files and added a keyboard shortcut in intellij to open the current file with the os-default program. If cursor is not already open in the current project when the keyboard shortcut is pressed, files don’t open within a project at all. @codebase doesn’t work in this projectless workspace and chat history appears to be completely ephemeral(or at least there’s no way to access it again once I’ve opened the project workspace and I don’t see any “global” view of chat history).
There’s a number of small annoyances as well.
The “Auto” model feature doesn’t seem to mention which model is being used.
Clicking the @ symbol, and then typing @<code/codebase/docs/etc>, removes all of the options that are normally shown if you type “@” without clicking the symbol first.
Mostly I’m dropping cursor because of the removal of the @codebase feature which feels extremely user-hostile and seems suspiciously like you’re trying to save tokens at the expense of quality. I’d definitely consider re-subscribing if/when the cursor team adds that feature back with some explanation of why it was removed in the first place.
If you’re losing money based on each request because the context length is too long with @codebase enabled, just tell us that, adjust your pricing model and move on. Don’t remove core features from your product without a reasonable explanation!
Of course they are saving costs by limiting the context, models are expensive and we want to pay cheap.
If you want full context and direct access to Claude servers you can use Claude Code, but its cost in a single day can be more than a whole month’s subscription to Cursor.
But I also hope they can find a better balance, Agent is great.
It would be pretty bad choice if they cripple the whole purpose of the editor by limiting tokens to save a few bucks per day per user because they only need the editor because of having AI integration only.
From other discussions I would say that context limit combined with tool usage and providing a practical middle ground for most users is the actual limiting factor.
This is proved by screenshots of users who have the context limit reached when they enter one word into the prompt due to many MCP tools attached. When they remove some heavy MCP tools it runs normal again. Several users posted proof of this independently.
Comparing Cursor to Claude Code is apples and oranges.
For me 3.5 works amazing in 0.48. Gave it huge list of things to do and it did them all correctly. Yes i had to tweak prompt and rules for 0.47 and 0.48 but now simpler rules work better actually where before i had to regulate heavily.
As repeated many times, adding more context is not a way forward, I’m not going to paste links to scientific papers, its obviously difficult for someone to change the workflow from @codebase change this thing to a more detailed prompting and rule-specified workflow but you can get there with practice, this technology is continously evolving and we can’t ask for it to remain still, good luck finding any alternatives cost-wise.
I am renewing my subscription because I really want AI-assisted coding within an IDE, and it’s too early to give up on one of the leaders in this area. I use Claude Code for all major coding operations and Cursor for tweaks and small edits, and life is good. I would use Cursor for everything if I could control what I am using with complete transparency on the costs. I basically want Claude Code plus fine-grained API controls inside a decent IDE like VS Code, and Cursor is well positioned to provide this, even if it’s not the default mode of operation. Why not let us pay for what we want and make a small profit on top of that?
I agree whole heartedly. Thank you for sharing your perspective with the Cursor community. I hope the Cursor leadership team and finance wing sees the dumpster fire that Cursor is evolving into and changes course to avoid sinking this Titanic IDE.
FWIW, I updated cursor to 0.48 and even though the changelog claims “the @Codebase tool has been removed”, this only seems to be true for agent-capable models.
I’m not really sure what changed in that respect. The last few versions had also removed @codebase for models like sonnet.
If you use r1, @codebase still works, as does the cmd + enter keyboard shortcut.
While I don’t really see the purpose of stripping out @codebase for “agent-capable” models, the 0.48 changes were not quite as catastrophic for my workflow as the changelog would have led me to believe.
@codebase did not perform as well as we’d like, and basically ignored the incoming prompt from the user!
Agent has so far proven to be a better method of acquiring context, as it goes through the same workflow as a human would in reading the question, then searching for the answer with a thought process.
We are not fixed on this, though, and if we find a good solution for context gathering that doesn’t rely on the agent’s ability to find the right files, we’ll definitely add it!
.