I love Cursor, but it feels like a ticking time bomb with extensions not being updated at the same rate as VSCode.
Also another issue I am starting to see is the lack of shared MCP servers. If I have VSCode, Cursor, and Claude open, each one is running its own instance of the MCP server. You can imagine that with a dozen or so MCP's, the memory footprint becomes quite large for no reason.
I think about this daily. More devs are starting to pick up on Claude Code. The initial “not an IDE!” scare is usually diminished within the initial session.
I don’t think the future of agentic software development is in an IDE. Claude Code gives me power to orchestrate - the UX has nothing to do with terminal; it just turns out an agent that lives on the OS and in the filesystem is a powerful thing.
Anthropic can and will evolve Claude Code at a pace cursor cannot evolve IDE abstractions. And then yea - they are designing the perfect wrapper because they are also designing the model.
Note: the text of that article itself is AI generated.
> Automate Content: Like this very post. I use Wispr Flow to talk with Claude, explain the topic and tell it to read my past blog posts to write in my style.
Now I have the mental image of the owner of that blog tearing his hair out trying to get back into his computer, while the AI that locked him out is happily posting on his blog trying to convince other gullible humans to hand over control of their computers.
How are people using Claude Code day to day without spending a lot? I tried it on a moderately complex task and it chewed through tokens at an alarming rate. I quickly spent $2 and hadn’t even arrived at an adequate solution yet. I’ve heard other people say they’ve spent $10-20 in a coding session. I don’t see how that’s sustainable for me, so I’ve stuck with my $20/month Cursor subscription.
Pro isn't a static plan. Pro subs can access Claude Code but are paying via API metering. I have it setup at home and, while I haven't used it much, it can quickly add up.
What I did do, because my personal projects aren't too complex, is moved the model from Opus to Sonnet which is about 1/5 the cost.
For day-to-day stuff I have ProxyAI (on IntelliJ, continue.dev works for this too) pointed at Mistral's Codestra for auto-complete and to Claude 4 for chat.
Claude Code is just for giving the bot tasks to do without me being deeply involved in the work.
Normally to make a new smallish feature it costs me about $0.40.
The core suggestion is to point specifically at the files you want it to read and use as a reference, otherwise it might go read some huge file for no reason. Also the tokens used depend on the size of the project.
Generally, if I'm doing something I can box, I'd use chatgpt and copy it in myself. If I want something matching the style, I'd use a guided Roo Code.
Until Claude Code becomes manageable price wise, I don’t think Cursor really sees them as their competition. I can burn the whole cursor subscription price in a single day with Claude Code.
Ha, same. How do you use it? I tried all the fancy context management stuff multiple times, but I mostly just have a chat buffer open and copy paste stuff manually. Text wrangling is so damn efficient in Emacs. I pay around 10$ to Anthropic per month in API tokens for pretty heavy usage. With deliberate context management (I found keeping it small and focused vastly improves responses), cost is really not an issue.
Didn't try anything agentic within Emacs yet, don't find that helpful enough so far.
> tried all the fancy context management stuff multiple times, but I mostly just have a chat buffer open and copy paste stuff manually.
As of last week you can insert a link to a plain-text file in a chat buffer to include its contents in the prompt. It must be on a line by itself. In Markdown it looks
[like this](/path/to/file)
with Org links in Org chat buffers.
This feature is disabled by default to minimize confusion. To enable it you can flip the header line button that says "ignoring media" to "sending media". This works for sending images and other media too, if the model supports it.
I'm surprised nobody is mentioning how cheap copilot pro is. $20 and you get all you can eat inference without using your own api key for the models on vs code agent mode.
Copilot pales in comparison to Cursor Pro. I've trialed it three or four times in the last two years and stopped using it after a few days each time. Honestly, I have no idea why anyone pays for it given the alternatives.
My only wish is that Cursor had partnered with Zed. vscode isn't enjoyable.
I’m on Pro+ and get rate limited heavily. 1-2 hours of semi heavy use and the brakes kick in. I can’t stay productive in it because this always rips me out
Yeah, if you're a heavy user of Claude code, you pretty much need to use it with a Max subscription rather than a BYOK approach. But that starts at $100 / month so it's a significant bump from cursor.
For any professional SWE 1/200 a month is basically nothing in terms of the value it delivers. They just rolled it out to the 20/month plan with limited usage but as soon as people get used to it I have no idea why they wouldn't upgrade unless they are just entering the field and don't have a job yet.
Anyone pulling that a month not working for themselves doesn't have to think about the cost. That kind of salary is paid by corps with strict privacy policies.
Unless you do nothing else with your time I'm not sure how you'd utilize the $100/mo plan fully.
the pricing is for token usage in 5/hr windows not monthly caps. if you use it intensely a couple times a month within a 5hr window it's not hard to hit the cap and want to upgrade. Personally I just work on some side projects during work on another monitor and just every half an hour so throw something at it and that's been very valuable for me.
225 messages every 5 hours? You hit that on the side while you're doing your day job? I suppose if you push all work to Claude and do nothing else all day it could be a concern but I don't think it would be a very effective way to work in it's current state unless you want to be left with a giant mess.
I admit their transparency around limits is a bit questionable. Why don't they just list out the tokens and remaining time?
Yeah I hit it again today, refactoring can use a ton. It did make a bit of a mess but planning it out making sure everything is tested and having it run through checking the tests making sure they all still pass uses lots of tokens but passively doing that while I'm working is way faster than doing it manually.
Sometimes I'll just experiment with weird stuff and end up reverting it or delete it afterword's. Also fun to build really nice data visualizations for data I'm working with.
> For any professional SWE 1/200 a month is basically nothing in terms of the value it delivers.
If it is delivering that value indeed, then 100-200 dollars each month is exactly what that professional SWE is worth.
That SWE is going to have to pivot into something adjacent to provide value-add, and if the only value-add is "I can review and approve/reject", then it's only a matter of time[1] that there are no SWEs who can review and then approve/reject.
And that is assuming that the LLM coding performance remains where it is now, and does not improve to the review, orchestration, design and planning levels.
> Pro ($20/month): Average users can send approximately 45 messages with Claude every 5 hours, OR send approximately 10-40 prompts with Claude Code every 5 hours. [0]
One of the first features we added to FastMCP 2.0 [1] was server composition and proxying (plus some niceties around trying to reuse servers whenever possible) for exactly this reason, so that you can ideally run a single instance of your preferred servers without replicating the setup over every application. In some ways, MCP standardized how agents talk to APIs but introduced a new frontier of lawless distribution! This is something that has to improve.
Have you tried the Jetbrains agent, I think it’s called Jennie?
I am trying it right now and it seems decent enough but I haven’t tried Cursor as I don’t really like vs code.
have you compared Windsurf (previously codeium, had to look it up) to Github Copilot agent or ChatGPT? I'm mostly still using ChatGPT's App integration. It can see my currently open files in webstorm and can directly suggest and apply changes from the app. And I find the suggestions better then JetBrain's AI integrations. But curious about Windsurf / CoPilot Agent
Jetbrains Junie is pretty good and comparable to Cursor in my experience. And since it is already included in my Jetbrains license, I have had no need for Cursor.
have you tried the CoPilot agent tab in Github copilot chat in Webstorm? I wonder if/why you find using it through VSCode better in any way then using it directly in WebStorm.
I don't understand. Is this meant to run locally? Because I tried to deploy my agent using GitHub MCP server to K8s. I can't ask my agent to run docker command in a pod.
Cursor and other VSCode forks connect to Open VSX [1] for extensions. Barring some of the Microsoft extensions, I've found that pretty much all the extensions I use are available and kept up to date on Open VSX. Cursor seems to have enough funding to support their own variants of the Microsoft extensions, like Python and C++.
The one issue I've run into is that the VSCode version Cursor uses is several months old, so we're stuck using older extensions until they update.
for me, I’ll keep doing work on the codebase in a separate vscode while Cursor’s agent is wiling away at it, so as to not be distracted or interrupted by the agent activity in the corner of
my screen. and then i’ll have a claude or aistudio tab open doing bigger analysis or planning tasks, reading papers together, etc.
What would you say to somebody who doesn't trust Cursor to "take the reigns off" and go "agent mode" and do such large changes? Even with "checking/approving/having the final say" as a user, I feel there is time lost if the AI does not do the changes the right way/makes small pesky mistakes.
Autocomplete: yes
Asking questions: yes
I know everybody is all into agent mode and it isn't a Cursor specific thing, I'm just confused why everybody is into it.
I just accept that when I return to the Cursor window, it is going to have made some mistakes and I’m going to have to spend some time fixing things. I agree it’s a delicate balance, and sometimes it feels more like I’m exploring methods of steering the AI than I am doing anything about the code.
my usecases have been building relatively self-contained, well-specified components of distributed systems out of academic papers and my own design notes / prototype skeletons, in rust. there’s a lot of context for the agent to ground against, and it can blow out ideas into their implications relatively well ime.
the experience of fixing up lots of pesky mistakes that you yourself would never make is kinda annoying.
Also another issue I am starting to see is the lack of shared MCP servers. If I have VSCode, Cursor, and Claude open, each one is running its own instance of the MCP server. You can imagine that with a dozen or so MCP's, the memory footprint becomes quite large for no reason.