I heard one of the guys on the Codex team say on a podcast that he never has to fill out a form anymore. He just has Codex do it in his logged-in browser.
I knew Codex could use a browser. I just didn't realize the new browser control could work this cleanly in my existing browser session, as opposed to launching a separate one.
I've done a decent amount of Playwright automation, but it generally runs in its own browser session. It's great for stuff I'm going to do repeatedly, but it doesn't really help when I need to sign up for some random thing once.
I hadn't upgraded the ChatGPT app in a while and didn't realize how much had been added. Now when I run into a form, I can just tell Codex to do it in the browser I'm already using.
I'm working on a bank import thing using Plaid and had to go through their whole signup process. I barely even looked at it. Codex just did it.
It's glorious.
There is a crazy amount of cognitive overhead in going to some website, figuring out its registration flow, filling everything in, and clicking through all the little steps. None of it is hard. It just completely interrupts whatever you were actually doing.
I may actually never fill out a form again.
@gregbarbosa Gotcha yeah I just checked and see that looks like maybe Claude optimized the way it handles MCP since I last checked, I mostly use Codex. They’re probably not doing as many optimizations.
Initially, I hated MCPs because they were this abstract thing everyone was talking about and nobody was actually using.
Now I use a handful of them, and I hate them for an entirely different reason.
Most of the time I'd rather have a good old-fashioned API or CLI. The docs are already perfectly human-readable and machine-readable. I can point an agent at them and it works.
MCPs add loading overhead to every session. They bloat the context. Installing one usually means restarting Codex or Claude. All for something an API reference and a few lines in AGENTS.md can often do just as well.
The one part I actually love is authentication.
Let me connect a brand-new service without opening its stupid website, clicking through seven onboarding questions, finding the developer settings, and hunting down an API key.
Honestly, that's the part I want standardized. Let the coding agent handle authentication, then get out of the way. I want Codex to be the interface for everything. I don't want to learn another dashboard just to use a service headlessly.
So yeah, I still hate MCPs. I just finally hate them from experience.
@galengidman This is exactly the kind of thing I was thinking that I probably wasn’t thinking of thanks the only issue here is that I tend to usually work reclined with my feet up and then I also like to walk around sometime so I kind of want something mobile
I use voice-to-text so much now that I'm starting to wonder if I need a dedicated physical button for it.
I already have a little one-handed keyboard for my left hand, which has been nice. But I'm imagining something even easier. Maybe a button attached to my hand? A ring? Something I can press without reaching for a keyboard.
I did some searching and found something called Flic, but it didn't exactly inspire a ton of confidence.
Has anyone else gone down this rabbit hole? What did you end up using?
I'm starting to think the most powerful thing about AI is that it lets me overcome technical debt.
It's cool that I can launch a site. It's cool that I can make a voice agent call me and have that magical holy-shit-this-works moment.
But looking back at the things I've built, launching was almost never the problem.
The wheels started falling off when technical debt built up.
Yesterday I found a bug in one of my dashboards that caused it to send multiple emails to people when it shouldn't have. Codex didn't just patch the bug. It refactored the underlying workflow so the same class of problem would be harder to create again.
Same thing with the voice-to-text tool I use constantly. It's an open source version of Whisper Flow that I've modified heavily.
When I first started modifying it, I thought it was cool. Then something broke the next day and I immediately thought, shit, I need to go back to the paid one.
This is something I rely on. It can't just randomly stop working.
But Codex is good enough to fix that stuff now. And not just get it working for the moment. It can usually fix it in a durable way.
Yes, I learned the word durable from AI. Yes, I know it's an AI tell. But sometimes AI is just smarter than me and I learn a useful word.
Being able to build cool things is great. But being able to keep them working, refactor them, and prevent the technical debt from slowly killing them is probably the bigger deal.
@kalenjordan as maintenance engineer sounds better than caretaker, or software engineer sounds better than programmer, or … Tarzan, King of the Jungle, sounds better than crazy long-haired hippie that delights in swinging from tree vines and screaming nonsense.
@gregbarbosa Does it show them all loading for each new session? I think codec does that can’t remember if Claude does. I have like 3 of them.
This is another reason I prefer Apis or clis they can be consumed deterministicaply out of the box.