A week with Cursor Agent CLI
It feels like every week, there's either a new LLM version or new AI coding tool or both. It's freakin' exhausting, so I try to be selective about which tools I dabble with. When I do venture from my usual stack, I try to commit to at least a solid week of use.
I jump between Cursor and Claude Code regularly, I still find the Anthropic pricing to be a bit prohibitive. I recently did a week of Claude Code, to test my own cost sensitivity. Got a lot done, but when I've hit 10 bucks of usage before lunch, I start to get a bit nervous.
I like a good deal
Something really nice about Cursor, is that when a new LLM version becomes available, they tend to do some discount period. With the newer Claude model, it was 25% less credits. With GPT-5 they did a free week or so.
It's already hard to compete with Cursor's pricing, as the cost of admission already gets you a bunch of "fast" credits. With these discounts, I've found that my previous estimates on usage-based costs were way too high.
I'd rather be using Vim
Well, Neovim to be exact. Seriously though, the biggest hurdle for me using these "modern" tools is that there is a heavy focus on IDEs, specifically Visual Studio Code.
I wrote about it over a decade ago, code editor proficiency is extremely important. A lot of the speed boost from using generative AI for coding gets lost when you're also trying to learn the editor.
Don't even think of mentioning "Vim mode" either. That shit's garbage, if you actually used Vim you'd realize it. It's cosplay, at best.
CLI agents fit better
It's not just about Vim, either. I live in the terminal.
Because of this, CLI agents do a really good job of fitting into my existing
stack. Armed with tmux, I can jump between a few projects, double and triple
fisting tasks across different repos with ease.
I'm not sure that these CLI agents will indoctrinate people to the terminal life, but it's nice to see that these tools are being prioritized.
When value meets interfacing
Fast forward to last week. I hung up Claude Code for the most part, and switched over to Cursor Agent CLI (Cursor CLI) to help me tackle this week's tasks.
There's no special pricing for using the CLI version of Cursor, best as I can tell. Also, GPT-5 was still in their promotional period.
Now a very affordable solution has the proper interfacing. LFG.
Musings after a week of use
A week of use went well enough. Cursor CLI didn't feel as bad as when I dabbled with Codex CLI. Far from perfect though, but I'm thinking some of that is more of some of the bigger picture issues I've been having with Cursor, rather than something specific to the CLI version.
I was able to get quite a bit done and took care of business in multiple languages across multiple projects. Generally speaking, things felt a lot slower, thanks to GPT-5's "thinking" models taking more time to process.
The slowness wasn't a bad thing, as it seemed to have improved the overall output. Spin outs were minimal, but I did have some trouble picking up where I had left off on projects.
Claude Code is really good at handling a prompt that starts with "Recently you
helped me with...". Cursor CLI always decided to try to run git to walk
through previous commits. In fact, it continually tried to run git for various
things, like trying to commit code.
I gave it an earful a few times.
Can I get a word in here?
I've been having decent success with these tools, I'm still very reluctant to adopt background agents. I don't think these tools are even remotely close to running unassisted. Not just that, I'd much rather try to right the ship in real time, rather than having to walk back a merge request.
Best as I can tell the folks over at Cursor are trying to train us to trust the robots more. The feedback loop with both Cursor and the CLI agent seems to be favoring doing more and asking less.
With every new version of Cursor, I struggle to get it to open files to let me review them. Cursor CLI exhibited the same behavior, doing a bunch of work, but not giving me an opportunity to review each step of the process.
I've coaxed the behavior back out, but it's been a struggle. While you can
review things after the fact, I found the interfacing to be less than ideal. At
least a few times, I found myself slapping CTRL+C to stop what felt like an
attempted uprising or coup.
Conclusion
While I do have some gripes with Cursor CLI, I still plan to continue using it. But, I don't consider it a replacement for Claude Code, rather a replacement for the Cursor IDE.
Next steps on my end will be to continue down the path of using more CLI agents.
As it's been a minute, it's probably time for me to give Codex CLI another shot. Also, I've yet to dig into things with Aider, even though I think it's probably the tool I'll end up with long term at some point.
God only knows what new tools will be zipping across the wire next week.