Growing with cursor and trying the codex cli
Growing with Cursor
I’ve been using Cursor Pro ($20) since February 11, 2025. I felt like maybe I started with code gen a little earlier, but maybe that was experimentation on free plans? Or maybe just the chatgpt copy paste loop. In any event, close enough. That’s about 19 months ago.
At the time, Cursor was just a VSCode fork, and Claude Code and Codex weren’t even released yet. Historically I’ve been a vim and cli guy, so half the friction was just using a new to me IDE.
I started with Tab model, for sophisticated completion, and sometimes used the chat mode to try and generate an entire function. I remember learning that I could write a comment inline, and then use Tab to complete the block of code. It was all amazing. I got better and better at generating whole funcitons. Slowly the scope which I could manage, or the models could handle, grew.
Then for a while I grappled with things like generating intermediary markdown plans, figuring out where to put artifacts, figuring out the balance of monorepo and smaller repos, etc. I remember for a long time I was very explicit about which items to tag in to the context in the chat window.
Claude Code Frenzy
Not long after I started using Cursor, Claude Code was released. It slowly then furiously grew in popularity. I observed lots of discussion and commentary on its usage. Of special note was how many people seemed to significantly burn through usage limits and token volumes. As Codex grew behind it, similar stories arose.
At some point, Cursor released their Agent oriented app, and I switched over from the VSCode setup. The chat window is front and center, and diffs and such are optional viewing on the right hand pane. Their cloud agents came along as well, and I would use those when I was on the phone away from my desktop or laptop.
Along the way, people really got into the ideas around MCP, Skills, and general agent coordination. It seemed a bit much to me, and still does, but it also seemed right inline with how the general software community approaches problems.
Somewhere in there I used the Amp free plan a little, just to make sure I experienced a cli approach. That service didn’t last too long though.
Over the last few months, some of the shine seems to have worn off of Claude Code in the community. Codex is, or has, caught up. Things are stabilizing. Models change and waves of commentary come with them.
My simple patterns now
Back to my Cursor usage.
These days, the models and harness are so much better. Ideas like intermediary plans are built in. I rarely need to explicitly tag in a file (the search is better?). Overall things are easier to use and the scope of a single conversation with an agent is more broad.
My usage of these tools has become pretty simple. If I’m working locally, I start a conversation against a single repo. It might be a one shot, or a short back and forth. I get some diffs, review them lightly, and decide if I want to keep the changes. If I do, I then switch to a terminal and use git manually to commit and push etc. If I’m using cloud agents, then there’s a PR from the agent work. I review and then usually squash and merge.
All that to say, I rarely have the agent doing git commands. I’ve never installed any MCP servers or skills in cursor. Along with that, I’ve found the Cursor Composer 2.5 model to be very effective and token efficient.
Let’s try Codex CLI and getting distracted
I’m getting what I need from Cursor. But I still wonder a bit about switching back to a cli tool. And I want to experience some of the other tools myself.
So today I’m trying out the codex cli, and planning to subscribe to the $20 Plus plan. There’s some mild free tier usage, so I’m starting there.
The first thing I noticed was how /mcp showed me a
ton of things getting loaded as codex_apps. Most of
it was github api stuff.
Wow. Seems like that would be a waste of tokens, sending all that info about those tools to the model. And as noted before, I don’t really have the models do github stuff.
So I grappled with things and eventually found someone recommend this command:
codex features disable apps
Which resulted in this in the config.toml:
[features]
apps = false
And this when I queried the list:
Looking at the tools
While working on the mcp stuff, I saw someone just ask codex for its tool list. So I tried that out on both Cursor and Codex, just out of curiosity.
The default codex cli tools (codex-cli 0.147.0):
And the tools from my cursor app (mac os app, 3.15.6):
Back to work
I’ll try the codex cli for a while, and see how things go. Should be fun to compare and contrast.