It is baffling how these AI companies, with billions of dollars, cannot build native applications, even with the help of AI. From a UI perspective, these are mostly just chat apps, which are not particularly difficult to code from scratch. Before the usual excuses come about how it is impossible to build a custom UI, consider software that is orders of magnitude more complex, such as raddbg, 10x, Superluminal, Blender, Godot, Unity, and UE5, or any video game with a UI. On top of that, programs like Claude Cowork or Codex should, by design, integrate as deeply with the OS as possible. This requires calling native APIs (e.g., Win32), which is not feasible from Electron.
The situation for Desktop development is nasty. Microsoft had so many halfassed frameworks and nobody knows which one to use. It’s probably the de facto platform on Windows IS Electron, and Microsoft use them often, too.
On MacOS is much better. But most of the team either ended up with locked in Mac-only or go cross platform with Electron.
This is another common excuse.
You don't need to use microsoft's or apple's or google's shit UI frameworks. E.g. see https://filepilot.tech/
You can just write all the rendering yourself using metal/gl/dx. if you didn't want to write the rendering yourself there are plenty of libraries like skia, flutter's renderer, nanovg, etc
Customers simply don't care. I don't recall a single complain about RAM or disk usage of my Electron-based app to be reported in the past 10 years.
You will be outcompeted if you waste your time reinventing the wheel and optimizing for stuff that doesn't matter. There is some market for highly optimized apps like e.g. Sublime Text, but you can clearly see that the companies behind them are struggling.
The various GPU-accelerated terminal projects always make me chuckle
Not seeing complaints doesn't mean they don't exist. Not to mention ui latency that is common in electron apps that is just a low-level constant annoyance.
I have complained about literally every Electron based app I have ever used. How would you know there are no complaints?
How is File Pilot for accessibility and for all of the little niceties like native scrolling, clipboard interaction, drag and drop, and so on? My impression is that the creator is has expertly focused on most/all of these details, but I don't have Windows to test.
I insist on good UI as well, and, as a web developer, have spent many hours hand rolling web components that use <canvas>. The most complicated one is a spreadsheet/data grid component that can handle millions of rows, basically a reproduction of Google Sheets tailored to my app's needs. I insist on not bloating the front-end package with a whole graph of dependencies. I enjoy my NIH syndrome. So I know quality when I see it (File Pilot). But I also know how tedious reinventing the wheel is, and there are certain corners that I regularly cut. For example there's no way a blind user could use my spreadsheet-based web app (https://github.com/glideapps/glide-data-grid is better than me in this aspect, but there's no way I'm bringing in a million dependencies just to use someone else's attempt to reinvent the wheel and get stuck with all of their compromises).
The answer to your original question about why these billion dollar companies don't create artisanal software is pretty straightforward and bleak, I imagine. But there are a few actually good reasons not to take the artisanal path.
> You don't need to use microsoft's or apple's or google's shit UI frameworks. E.g. see https://filepilot.tech/
That's only for Windows though, it seems? Maybe the whole "just write all the rendering yourself using metal/gl/dx" is slightly harder than you think.
"native" is used for different things, from "use the platform's default gui toolkit" to "compile to a machine code binary". the former is a bit of a mess, but the latter is strictly better than wrapping a web view and shipping an entire chrome fork to display and interpret it. just write something in qt and forget about native look and feel, and the performance gain will be enough to greatly improve the user experience.
Should just use javafx or swing. Take a leaf out of intellij which while it as it's own performance problems (although not from the fact of the ui framework) has a fantastic ui across Mac / windows / nix
Do not give a shit about how they excuse doing a bad job. If their tools make them that much more productive, and being the developer of those tools should allow you to make great use of them.
Use native for osx Use .Net framework for windows Use whatever on Linux.
Its just being lazy and ineffective. I also do not care about whatever "business" justification anyone can come up with for half assing it.
As I outlined in a sibling comment. You can still use React and your JS developers. Just don't ship a whole browser with your app.
May be an app that is as complex as Outlook needs the pixel-perfect tweaking of every little button that they need to ship their own browser for exact version match. But everything else can use *system native browser*. Use Tauri or Wails or many other solutions like these
That said, I do agree on the other comments about TUIs etc. Yes, nobody cares about the right abstractions, not even the companies that literally depend on automating these applications
microsoft also uses react native for the start menu and also bricked that during a recent upgrade apparently... along with breaking other stuff.
This. Even Linux is nasty. Qt and GTK are both horrible messes to use.
It would be nice if someone made a way to write desktop apps in JavaScript with a consistent, cross-platform modern UI (i.e. swipe to refresh, tabs, beautiful toggle switches, not microscopic check boxes) but without resorting to rendering everything inside a bloated WebKit browser.
- Video games often use HTML/JS-based UI these days.
- UE5 has its own custom UI framework, which definitely does not feel "native" on any platform. Not really any better than Electron.
- You can easily call native APIs from Electron.
I agree that Electron apps that feel "web-y" or hog resources unnecessarily are distasteful, but most people don't know or care whether the apps they're running use native UI frameworks, and being able to reassign web developers to work on desktop apps is a significant selling point that will keep companies coming back to Electron instead of native.
I have been building Desktop apps with Go + Wails[1]. I happen to know Go, but if you are ai-coding even that is not necessary.
A full fledged app, that does everything I want, is ~ 10MB. I know Tauri+Rust can get it to probably 1 MB. But it is a far cry from these Electron based apps shipping 140MB+ . My app at 10MB does a lot more, has tons of screens.
Yes, it can be vibe coded and it is especially not an excuse these days.
Microsoft Teams, Outlook, Slack, Spotify? Cursor? VsCode? I have like 10 copies of Chrome in my machine!
I've looked into Tauri and Wails, but they don't seem realistic for a cross-platform app with wide distribution across multiple platforms and platform versions.
One of Electron's main selling points is that you control the browser version. Anything that relies on the system web view (like Tauri and Wails) will either force you to aggressively drop support for out-of-date OS versions, or constantly check caniuse.com and ship polyfills like you're writing a normal web app. It also forces you to test CSS that touches form controls or window chrome on every supported major version of every browser, which is just a huge pain. And you'll inevitably run into bugs with the native -> web glue that you wouldn't hit with Electron.
It is absolutely wasteful to ship a copy of Chrome with every desktop app, but Tauri/Wails don't seem like viable alternatives at the moment. As far as I can tell, there aren't really any popular commercial apps using them, so I imagine others have come to the same conclusion.
If the web-interface of a website can serve same HTML to all browsers, these UI can as well. I don't think we have IE6 level incompatibility these days. I have no idea what specific incompatibility you are talking about. I am writing my 4th desktop app since early 2025.
But sure, you could have some specific need, but I find it hard to believe for these simple apps.
I am in love with wails. having python and JS background with no go experience. I pm'ed Ai agents to create a fairly complex desktop app for my own use. it is day and night in terms of performance compared to lightest electron app.
Wow Wails looks interesting! Hadn't heard of it before.
Given that OpenAI managed to single-handedly triple the price of RAM globally, people will very much care about their chat application consuming what little scraps are left over for them, even if they don't know enough about anything to know why their system is running poorly.
My main take is exactly the opposite. Why not build everything with a simple text interface (shell command) so the models learn to use these tools natively in pretraining. Even TUI like codex-cli or claude code are needless abstractions for such use cases and make full automation hard. You could add as many observability or input layers for humans as you want but the core should be simple calls that are saved in historical and documentation logs. [the headless/noninteractive modes come close, as do the session logs]
I agree. I like using Claude or Codex in VM on top of the tmux. Much more flexibility that way. I open a new tmux window for each issue/task big enough to warrant it, issue a prompt to create a worktree and agents and let them go to town. I actually use claude and codex a the same time. I still get observability because of tmux and I can close my laptop and let them cook for a while in yolo mode since the VM is frequently backed up in proxmox pbs. I am a retired hobbyist but this has been a nice force multiplier without devolving a complete viby mess. I hope these new orchestration tool support this like vs code remote development does. Same for cloud. I want them to support my personal "cloud" instead of laggy github mess.
It would be cool if I didn't have to worry about whether I was "in" or "out" of the AI TUI. Right now, I need at least two terminals open: One running my shell, that I use to issue shell commands, and one running Claude, where I can't. It would be nice if it could just be my shell, and when I wanted to invoke claude, I'd just type:
Right in the shell.c Do this programming task for me.-p
Oh wow, nice. Does it remember context from run to run?
> even with the help of AI.
This is what you get when you build with AI, an electron app with an input field.
Doesn't have to be. I just revived one of my C++ GLFW app from 2009. Codex was able to help me get it running again and added some nice new features.
I guess you get an Electron app if you don't prompt it otherwise. Probably because it's learned from what all the humans are putting out there these days.
That said.. unless you know better, it's going to keep happening. Even moreso when folks aren't learning the fundamentals anymore.
I've done way, way more than that, as I'm sure others have too.
This is just bad product management.
So where is all this amazing software that you and others built with AI?
All I see is hype blog posts and pre-IPO marketing by AI companies, not much being shipped though.
You won't see it because it's mostly personal software for personal computers.
I've got a medical doctor handwriting decipherer, a board game simulator that takes a PDF of the rulebooks as input and an accounting/budgeting software that can interface with my bank via email because my bank doesn't have an API.
None of that is of any use to you. If you happen to need a similar software, it will be easier for you to ask your own AI to make a custom one for you rather than adapt the ones I had my AI make for me.
Under the circumstances, I would feel bad shipping anything. My users would be legitimately better off just vibe coding their own versions.
quality takes time. I'm still in stealth
There is a guy on twitter documenting his progress with moltbot/openclaw: https://x.com/Austen/status/2018371289468072219. Apparently he has already registered his bot for an LLC so he can make money w/ it.
Some guy on Twitter selling an AI coding boot camp is an interesting example. Also it's literally just a post of him looking for a real developer to review his vibe coded bot???
What does the bootcamp have to do w/ anything? He is using AI slop to make money, that's all that matters in a socio-economic system wherein everyone & everything must make profits to persist. Edit: found another example from coinbase: https://x.com/0xEricBrown/status/2018082458143699035.
Edit: I'm not going to keep addressing your comment if you keep editing it. You asked for an example & I found two very easily. I am certain there are many others so at this point the onus is on you to figure out what exactly it is you are actually arguing.
Your first example is just a twitter post of a guy asking for a developer to review his vibe coded bot. Nothing shipped.
The second example is twitter post of a crypto bro asking people to build something using his crypto API. Nothing shipped.
Literally nothing shipped, just twitter posts of people selling a coding bootcamp and crypto.
Their goal is to ship as fast as possible b/c they don't care about what you care about. Their objective is to gather as much data as possible & electron is good enough for that.
I work at OpenAI, and I get the concern. From our side, this was a careful tradeoff: Electron lets us iterate faster and makes it possible to bring the Codex app to Windows and Linux very soon. That doesn’t mean performance or UX don’t matter—we’re actively paying attention to both.
Would genuinely love your thoughts if you try it. Early users have been surprised by how native it feels!
Claude code is perfectly capable of writing low level rendering and input code and it can be equally as mindless as vibe coding web apps.
E.g. just say "write a c++ gui widget library using dx11 and win32 and copy flutters layout philosophy, use harfbuzz for shaping, etc etc"
At the end of the day LLMs just reproduce the consensus of the internet, so it makes sense that a coding agent would spit out software that looks like most of what's on the internet.
LLM output is called slop for a reason.
Why would widgets and buttons be better than a console, and or voice?
What if they kept the 'good stuff' from us? Does that seem likely here?
Unpopular opinion: why would you want a "native app"? On macOS, basically every app Apple makes themselves is worse in terms of design, usability, and performance than a popular Electron-based alternative.
For example, I tried opening a 200MB log file in Apple's Console.app and it hung. Opened right up in VS Code.
People's mileage may vary, but in my instance, this was so bad that I actually got angry while trying to use it.
It's slow and stupid. It does not do proper research. It does not follow instructions. It randomly decides to stop being agentic, and instead just dumps the code for me to paste. It has the extremely annoying habit of just doing stuff without understanding what I meant, making a mess, then claiming everything is fine. The outdated training data is extremely annoying when working with Nuxt 4+. It is not creative at solving problems. It dosent show the thinking. The Undo code does not give proper feedback on the diff and if it actually did "undo." And I hate the personality. It HAS to be better than it comes off for me because I am actually in a bad mood after having worked with it. I would rather YOLO code with Gemini 3 flash, since it's actually smarter in my assessment, and at least I can iterate faster, and it feels like it has better common sense.
Just as an example, I found an old, terrible app I made years ago for our firm that handles room reservations. I told it to update from Bootstrap to Flowbite UI. Codex just took forever to make a mess, installed version 2.7 when 4.0.1 is the latest, even when I explicitly stated that it should use the absolute latest version. Then it tried to install it and failed, so it reverted to the outdated CDN.
I gave the same task to Claude Code. Same prompt. It one-shotted it quickly. Then I asked it to swap out ALL the fetch logic to have SPA-like functionality with the new beta 4 version of HTMX, and it one-shot that too in the time Codex spent just trying to read a few files in the project.
This reminds me of the feeling I had when I got the Nokia N800. It was so promising on paper, but the product was so bad and terrible to use that I knew Nokia was done for. If this was their take on what an acceptable smartphone could be, it proves that the whole foundation is doomed. If this is OpenAI's take on what an agentic coding assistant should be—something that can run by itself and iterate until it completes its task in an intelligent and creative way.... OpenAI is doomed.
Genuinely excited to try this out. I've started using Codex much more heavily in the past two months and honestly, it's been shockingly good. Not perfect mind you, but it keeps impressing me with what it's able to "get". It often gets stuff wrong, and at times runs with faulty assumptions, but overall it's no worse than having average L3-L4 engs at your disposal.
That being said, the app is stuck at the launch screen, with "Loading projects..." taking forever...
Edit: A lot of links to documentation aren't working yet. E.g.: https://developers.openai.com/codex/guides/environments. My current setup involves having a bunch of different environments in their own VMs using Tart and using VS Code Remote for each of them. I'm not married to that setup, but I'm curious how it handles multiple environments.
Edit 2: Link is working now. Looks like I might have to tweak my setup to have port offsets instead of running VMs.
I have the $20 a month subscription for ChatGPT and the $200/year subscription to Claude (company reimbursed).
I have yet to hit usage limits with Codex. I continuously reach it with Claude. I use them both the same way - hands on the wheel and very interactive, small changes and tell them both to update a file to keep up with what’s done and what to do as I test.
Codex gets caught in a loop more often trying to fix an issue. I tell it to summarize the issue, what it’s tried and then I throw Claude at it.
Claude can usually fix it. Once it is fixed, I tell Claude to note in the same file and then go back to Codex
The trick to reach the usage limit is to run many agents in parallel. Not that it’s an explicit goal of mine but I keep thinking of this blog post [0] and then try to get Codex to do as much for me as possible in parallel
[0]: http://theoryofconstraints.blogspot.com/2007/06/toc-stories-...
Telling a bunch of agents to do stuff is like treating it as a senior developer who you trust to take an ambiguous business requirement and letting them use their best judgment and them asking you if they have a question .
But doing that with AI feels like hiring an outsourcing firm for a project and they come back with an unmaintable mess that’s hard to reason through 5 weeks later.
I very much micro manage my AI agents and test and validate its output. I treat it like a mid level ticket taker code monkey.
My experience with good outsourcing firms is that they come back with heavily-documented solutions that are 95% of what you actually wanted, leaving you uncomfortably wondering if doing it yourself woulda been better.
I’m not fully sure what’s worse, something close to garbage with a short shelf life anyone can see, or something so close to usable that it can fully bite me in the ass…
I fully believe that if I didn’t review its output and ask it to clean it up it would become unmaintainable real quick. The trick I’ve found though is to be detailed enough in the design from both a technical and non-technical level, sometimes iterating a few time on it with the agent before telling it to go for it (which can easily take 30 minutes)
That’s how I used to deal with L4, except codex codes much faster (but sometimes in the wrong direction)
It’s funny over the years I went from
1. I like being hands on keyboard and picking up a slice of work I can do by myself with a clean interface that others can use - a ticket taking code monkey.
2. I like being a team lead /architect where my vision can be larger than what I can do in 40 hours a week even if I hate the communication and coordination overhead of dealing with two or three other people
3. I love being able to do large projects by myself including dealing with the customer where the AI can do the grunt work I use to have to depend on ticket taking code monkeys to do.
Moral of the story: if you are a ticket taking “I codez real gud” developer - you are going to be screwed no matter how many b trees you can reverse on the whiteboard
Moral of your story.
Each and everyone of us is able to write their own story, and come up with their own 'Moral'.
Settling for less (if AI is a productivity booster, which is debatable) doesn't equal being screwed. There is wisdom in reaching your 'enough' point.
If you look at the current hiring trends and how much longer it is taking developers to get jobs these days, a mid level ticket taker is definitely screwed between a flooded market, layoffs and AI.
By definition, this is the worse AI coding will ever be and it’s pretty good now.
I will say that doing small modifications or asking a bunch of stuff fills the context the same in my observations. It depends on your codebase and the rest of stuff you use (sub agents, skills, etc)
I was once minimising the changes and trying to take the max of it. I did an uncountable numbers of tests and and variations. Didn't really matter much if I told it to do it all or change one line. I feel Claude code tries to fill the context as fast as possible anyway
I am not sure how worth Claude is right now. I still prefer that rather than codex, but I am starting to feel that's just a bias
I don’t think it’s bias: I have no love for any of these tools, but in every evaluation we’ve done at work, Opus 4.5 continually comes out ahead in real world performance
Codex and Gemini are both good, but slower and less “smart” when it comes to our code base
I hit the Claude limit within an hour.
Most of my tokens are used arguing with the hallucinations.
I’ve given up on it.
Do you use Claude Code, or do you use the models from some other tool?
I find it quite hard to hit the limits with Claude Code, but I have several colleagues complaining a lot about hitting limits and they use Cursor. Recently they also seem to be dealing with poor results (context rot?) a lot, which I haven't really encountered yet.
I wonder if Claude Code is doing something smart/special
In my case I've had it (Opus Thinking in CC) hit 80% of the 5-hour limit and 100% of the context window with one single tricky prompt, only to end up with worthless output.
Codex at least 'knows' to give up in half the time and 1/10th of the limits when that happens.
I don't want to be That Guy, but if you're "arguing with hallucinations" with an AI Agent in 2026 you're either holding it wrong or you're working on something highly nonstandard.
Your goal should be to run agents all the time, all in parallel. If you’re not hitting limits, you’re massively underutilizing the VC intelligence subsidy
https://hyperengineering.bottlenecklabs.com/p/the-infinite-m...
I have a found Codex to be an exceptional code-reviewer of Claude's work.
Same here. From my experience, codex usually knocks backend/highly "logical?" tasks out of the park while fairly basic front-end/UI tasks it stumbles over at times.
But overall it does seem to be consistently improving. Looking to see how this makes it easier to work with.
Hey thank you for calling out the broken link. That should be fixed now. Will make sure to track down the other broken links. We'll track down why loading is taking a while for you. Should definitely be snappier.
Is this the only announcement for Apple platform devs?
I thought Codex team tweeted about something coming for Xcode users - but maybe it just meant devs who are Apple users, not devs working on Apple platform apps...
Cool, looks like I'll stay on Cursor. All alternatives come out buggy, they care a lot about developer experience.
BTW OpenAI should think a bit about polishing their main apps instead of trying to come out with new ones while the originals are still buggy.
(I work on Codex) One detail you might appreciate is that we built the app with a ton of code sharing with the CLI (as core agent harness) and the VSCode extension (UI layer), so that as we improve any of those, we polish them all.
Any chance you'll enable remote development on a self-hosted machine with this app?
Ie. I think the codex webapp on a self-hosted machine would be great. This is impotant when you need a beefier machine (with potentially a GPU).
This should be table stakes by now. That's the beauty of these cli tools and how they scale so well.
Working remotely with the app would truly be great
Interested in this as well.
Any reason to switch from vscode with codex to this app? To me it looks like this app is more for non-developers but maybe I’m missing something
Good question! VS Code is still a great place for deep, hands-on coding with the Codex IDE extension.
We built the Codex app to make it easier to run and supervise multiple agents across projects, let longer-running tasks execute in parallel, and keep a higher-level view of what’s happening. Would love to hear your feedback!
ok , 'projects' but this would make a lot more sense if we could connect remotely to the projects which works without a problem using the IDE plugin, so right now I don't see any advantage of using this
Awesome. Any chance we will see a phone app?
I know coding on a phone sounds stupid, but with an agent it’s mostly approvals and small comments.
The ChatGPT app on iOS has a Codex page, though it only seems to be for the "cloud" version.
Looks like another Claude App/Cowork-type competitor with slightly different tradeoffs (Cowork just calls Claude Code in a VM, this just calls Codex CLI with OS sandboxing).
Here's the Codex tech stack in case anyone was interested like me.
Framework: Electron 40.0.0
Frontend:
- React 19.2.0
- Jotai (state management)
- TanStack React Form
- Vite (bundler)
- TypeScript
Backend/Main Process:
- Node.js
- better-sqlite3 (local database)
- node-pty (terminal emulation)
- Zod (validation)
- Immer (immutable state)
Build & Dev:
- pnpm (package manager)
- Electron Forge
- Vitest (testing)
- ESLint + Prettier
Native/macOS:
- Sparkle (auto-updates)
- Squirrel (installer)
- electron-liquid-glass (macOS vibrancy effects)
- Sentry (error tracking)
The use of the name Codex and the focus on diffs and worktrees suggests this is still more dev-focused than Cowork.
They have the same stack of a boot camper, quite telling.
Is the integration with Sentry native or via MCP ?
What does Sentry via MCP even mean? You want the LLM to call Sentry itself whenever it encounters an error?
> this just calls Codex CLI with OS sandboxing
The git and terminal views are a big plus for me. I usually have those open and active in addition to my codex CLI sessions.
Excited to try skills, too.
It's basically what Emdash (https://www.emdash.sh/), Conductor (https://www.conductor.build/) & CO have been building but as first class product from OpenAI.
Begs the question if Anthropic will follow up with a first-class Claude Code "multi agent" (git worktree) app themselves.
oh i didn't know that claude code has a desktop app already
It isn’t its own app, but it’s built in to their desktop, mobile and web apps.
And it uses worktrees.
I am not sure if multi agent approach is what it is hyped up to be. As long we are working on parallel work streams with defined contracts (say an agreed upon API def that backend implements and frontend uses), I'd assume that running independent agent coding sessions is faster and in fact more desirable so that neither side bends the code to comply with under specified contracts.
Usually I find the hype is centered around creating software no one cares about. If you're creating a prototype for dozens of investors to demo - I seriously doubt you'd take the "mainstream" approach.
Maybe a dumb question on my side; but if you are using a GUI like emdash with Claude Code, are you getting the full claude code harness under the hood or are you "just" leveraging the model ?
I can answer for Conductor: you're getting the full Claude Code, it's just a GUI wrapper on top of CC. It makes it easy to create worktrees (1 click) and manage them.
I don't think this is true. Try running `/skills` or `/context` in both and you'll see.
yeah, I wanted a better terminal for operating many TUI agent's at once and none of these worked because they all want to own the agent.
I ended up building a terminal[0] with Tauri and xterm that works exactly how I want.
0 - screenshot: https://x.com/thisritchie/status/2016861571897606504?s=20
Emdash is inducing CC, Codex, etc. natively. Therefore users are getting the raw version of each agent.
I never heard of Emdash before and I am following on AI tools closely. It just shows you how much noise there is and how hard is to promote the apps. Emdash looks solid. I almost went to build something similar because I wasn't aware of it.
They have Claude Code web in research preview
It seems the big feature is working agents in parallel? I've been working agents in parallel in Claude Code for almost 9 months now. Just create a command in .claude/commands that references an agent in .claude/agents. You can also just call parallel default Task agents to work concurrently.
Using slash commands and agents has been a game changer for me for anything from creating and executing on plans to following proper CI/CD policies when I commit changes.
To Codex more generally, I love it for surgical changes or whenever Claude chases its tail. It's also very, very good at finding Claude's blindspots on plans. Using AI tools adversarially is another big win in terms of getting things 90% right the first time. Once you get the right execution plan with the right code snippets, Claude is essentially a very fast typer. That's how I prefer to do AI-assisted development personally.
That said, I agree with the comments on tokens. I can use Codex until the sun goes down on $20/month. I use the $200/month pro plan with Claude and have only maxxed out a couple times, but I do find the volume to quality to be better with Claude. So far it's worth the money.
I'm a Claude Code user primarily. The best UI based orchestrator I've used is Zenflow by Zencoder.ai -- I am in no way affiliated with them, but their UI / tool can connect to any model or service you have. They offer their own model but I've not used it.
What I like is that the sessions are highly configurable from their plan.md which translates a md document into a process. So you can tweak and add steps. This is similar to some of the other workflow tools I've seen around hooks and such -- but presented in a way that is easy for me to use. I also like that it can update the plan.md as it goes to dynamically add steps and even add "hooks" as needed based on the problem.
Always sounds so interesting and then I do a search only to found out it's another product trying to sell you your 20th "AI credit package." I really don't see how these apps will last that long. I pay for the big three already - and no I don't want to cancel them just so I can use your product.
Aren't there 500+ aggregator services?
The landing page for the demo game "Voxel Velocity" mentions "<Enter> start" at the bottom, but <Enter> actually changes selection. One would think that after 7mm tokens and use of a QA agent, they would catch something like this.
It's interesting, isn't it? On the one hand the game is quite impressive. Although it doesn't have anything particularly novel (and it shouldn't, given the prompt), it still would have taken me several days, probably a week, working nonstop. On the other hand, there's plenty of paper cuts.
I think these subtle issues are just harder to provide a "harness" for, like a compiler or rigorous test suite that lets the LLM converge toward a good (if sometimes inelegant) solution. Probably a finer-tuned QA agent would have changed the final result.
It's also interesting how the functionality of the game barely changes between 60k tokens, 800k tokens, and 7MM tokens. It seems like the additional tokens made the game look more finished, but it plays almost exactly the same in all of them.
I wonder what it was doing with all those tokens?
How about us, Linux users? This is Mac only. Do they plan to support CLI version with all the features they are adding to desktop app?
Hi! Romain here, I work at OpenAI. The team actually built the Codex app in Electron so we can support both Windows and Linux very soon. Stay tuned!
Nice, thank you for sharing!
Let me guess, you use MacOS yourself?
not only is it mac only, it appears to be arm only as well. App won't launch on my intel mac
Yeah, I'm having the same issue. Disappointing limitations.
Guess MacOS gives you pass for early-access stuff, right? /s
From a developer's perspective it makes sense, though. You can test experimental stuff where configurations are almost the same in terms of OS and underlying hardware, so no weird, edge-case bugs at this stage.
OpenAI, ChatGPT, Codex
So many of the things that pioneered the way for the truly good (Claude, Gemini) to evolve. I am thankful for what they have done.
But the quality is gone, and they are now in catch-up mode. This is clear, not just from the quality of GPT-5.x outputs, but from this article.
They launch something new, flashy, should get the attention of all of us. And yet, they only launch to Apple devices?
Then, there are typos in the article. Again. I can't believe they would be sloppy about this with so much on the line. EDIT: since I know someone will ask, couple of examples - "7MM Tokens", "...this prompt initial prompt..."
And why are they not giving the full prompt used for these examples? "...that we've summarized for clarity" but we want to see the actual prompt. How unclear do we need to make our prompts to get to the level that you're showing us? Slight red flag there.
Anyway, good luck to them, and I hope it improves! Happy to try it out when it does, or at the very least, when it exists for a platform I own.
> truly good (Claude, Gemini) to evolve
Claude yes, but Codex is much better than Gemini in every way that matters except speed in my experience.
Gemini 3 Flash is an amazing model, but Gemini 3 Pro isn't great. It can do good work, but it's pretty random if it will or it will go off the rails and do completely the wrong thing. OTOH GPT 5.2 Codex with high thinking is the best model currently available (slightly better than Opus 4.5)
Not sure when you last evaluated the tools, but I strongly prefer Codex to Claude Code and Gemini.
Codex gets complex tasks right and I don't keep hitting usage limits constantly. (this is comparing the 20$ ChatGPT to the 200$ Claude Pro Max plans fwiw)
The tooling around ChatGPT and Codex is less, but their models are far more dependable imo than Antropic's at this very moment.
I don’t hit Codex limits because it’s so much slower, is what I’ve found personally.
I am not sure how those TUI are going to fare against multi providers ones like opencode.
The main thing I noticed in the video is that they have heavily sped up all the code generation sections... seems to be on 5x speed or more. (because people got used to how fast and good Sonnet, and especially Gemini 3.0 Flash, are)
I can't speak to the typos, but launching first for MacOS not something new for OpenAI. They did the same with their dedicated desktop client.
To me, the obvious next step for these companies is to integrate their products with web hosting. At this point, the remaining hurdle for non-developers is deploying their creations to the cloud with built-in monetization.
We have been working on this, letting any coding agent define infrastructure so we can define it effortlessly: https://specific.dev. We aren't just targeting non-developers though, we think this is useful to anyone building primarily through coding agents.
Just tell it to use your gcp/aws account using the cli, makes it infinitely powerful in terms of deployment. (Also, while I might miss some parts of programming that I have given to AI, I certainly don't miss working with clouds).
> Just tell it to use your gcp/aws account using the cli
Please don't.
People burning through their tokens allowance on Claude Code is one thing.
People having their agent unknowingly provisioning thousands of $ of cloud resources is something completely different.
This is also on the cloud providers for not giving us good tools to manage costs.
How about, "tell the agent to write instructions for cloud deployment with a cost estimate"
and specifically, the big companies, in a way that people notice. Claude Artifacts, AI Studio, etc. all kinda suck. If you have used Manus or connected your own CF, GCP, AWS, etc. you see how easy it could be if one of the big guys wanted it to be (or could get out of their own way).
the big boys probably don't want people who don't know sec deploying on their infra lol.
Deploying from Antigravity is as easy as say connecting the Firebase MCP [1] and asking it "deploy my app to firebase".
[1] https://firebase.google.com/docs/ai-assistance/mcp-server
Replit already does this
I dont think these are made for non-devs, Lovable and other which are built for non-devs already provide hosting.
interestingly opencode's first product was an IaC platform... seems to be where this is all going.
Something similar but for any ACP server: https://github.com/inercia/mitto
Looks like they forgot the part of the code editor where you can… edit code. Claude Code in Zed is about the most optimal experience I can imagine. I want the agent on the side and a code editor in the middle.
That’s not really a negative for me as I can easily jump into vscode where I already have my workspace for coding set up exactly as I like it. This being a completely separate app just to get the agentic work right is a good direction imo
Yeah but its annoying to find the file the agent just edited without any IDE/editor integration, you have to add that command which opens the file in vscode after editing.
It would be nice to have an integrated development environment.
How dare you want your IDE to be integrated!
Usage like this is becoming a rarity. Most people are editing significantly less and "agent interfaces" are slowly taking the focus.
"most" people aren't even using AI yet
Of those that are, most are not vibe coding, so an editor is still required at many points
For greenfield apps you can vibecode it. For existing complex apps, where existing products where customers pay us a lot of money for working software, understanding the changes and context surrounding them in the code is critical or else nobody knows how the system works anymore and maintenance and support becomes impossible.
- looks like OpenAIs answer to Claude Code Desktop / Cowork
- workspace agent runner apps (like Conductor) get more and more obsolete
- "vibe working" is becoming a thing - people use folder based agents to do their work (not just coding)
- new workflows seem to be evolving into folder based workspaces, where agents can self-configure MCP servers and skills + memory files and instructions
kinda interested to see if openai has the ideas & shipping power to compete with anthropic going forward; anthropic does not only have an edge over openai because of how op their models are at coding, but also because they innovate on workflows and ai tooling standards; openai so far has only followed in adoption (mcp, skills, now codex desktop) but rarely pushed the SOTA themselves.
Also interesting that they are both only for macOS. I’m feeling a bit left out on the Windows and Linux side, but this seems like an ongoing trend.
We did train Codex models natively on Windows - https://openai.com/index/introducing-gpt-5-2-codex/ (and even 5.1-codex-max)
I appreciate this (as a Windows user) but I'm also curious how necessary this was.
Like I notice in Codex in PhpStorm it uses Get-Whatever style PowerShell commands but firstly, I have a perfectly working Git-Bash installed that's like 98% compatible with Linux and Mac. Could it not use that instead of being retrained on Windows-centric commands?
But better yet, probably 95% of the commands it actually needs to run are like cat and ripgrep. Can't you just bundle the top 20 commands, make them OS-agnostic and train on that?
The last tiny bit of the puzzle I would think is the stuff that actually is OS-specific, but I don't know what that would be. Maybe some differences in file systems, sandboxing, networking.
my guess is that openai/anthropic employees work on macOS and mostly vibe code these new applications (be it Atlas browser or now Codex Desktop); i wouldn't be surprised if Codex Desktop was built in a month or less;
linux / windows requires extra testing as well as some adjustments to the software stack (e.g. liquid glass only works on mac); to get the thing out the door ASAP, they release macos first.
A lot of companies that use Windows are likely to use Microsoft Office products, and they were all basically forced to sign a non-compete where they can't run other models- just copilot.
I'm so sick and tired of the macOS elitism in the AI/LLM world.
Bit of a buried lede:
> For a limited time we're including Codex with ChatGPT Free
Is this the first free frontier coding agent? (I know there have been OSS coding agents for years, but not Codex/Claude Code.)
That depends on whether Gemini CLI counts. I've had generally bad experiences with it, but it is free for at least some usage.
Google also has aistudio.google.com which is Lovable competitor and its free for unlimited use. That seems to work so much better than gemini CLI even on similar tasks
Mac only. Again.
Apple is great but this is OpenAI devs showing their disconnect from the mainstream. Its complacent at best, contemptuous at worst.
SamA or somebody really needs to give the product managers here a kick up the arse.
Hi! Romain here, I work on Codex at OpenAI. We totally hear you. The team actually built the app in Electron specifically so we can support Windows and Linux as well. We shipped macOS first, but Windows is coming very soon. Appreciate you calling this out. Stay tuned!
Only thing i'd add re windows is it's taking us some time to get really solid sandboxing working on Windows, where there are fewer OS-level primitives for it. There's some more at https://developers.openai.com/codex/windows and we'd love help with testing and feedback to make it robust.
Electron? Why can't Codex write, or at least translate, your application to native code instead of using a multi-hundred-mb browser wrapper to display text? Is this the future of software engineering Codex is promising me?
Curios why electron not native?
Wouldn’t native give better performance and more system integration?
When you're a trillion dollar company that burns more coal than Bangladesh in order to harness a hyperintelligent Machine God to serve your whims, you don't have the resources to maintain native clients for three separate targets.
He literally says why electron in his comment that you are replying to
Going cross platform doesn’t sound the main reason (or I hope not). For a company that size, is it really hard to hire specialised small team?! It would be a good show case for their Codex too
They presumably use codex to build this. LLMs output is non-deterministic. Harder to keep the same logic across.
Would I love to see swiftui on macos, wpf/winui om windows, whatever qt hell it is on linux? Sure. But it is what it is.
I am glad the codex-cli is rust and native. Because claude code and opencode are not: react, solidjs and what have you for a tree layer.
—
Then again, if codex builds codex, let it cook and port if AI is great. Otherwise, it’s claim chowder
It is hard because this product will likely be obsolete next year based on how quickly AI is changing and evolving. Speed is king when you're on the frontier
Kudos to the OpenAI reps for responding to my comment and doing so politely.
My ire was provoked by this following on from the Windows ChatGPT app that was just a container for the webpage compared to the earlier bells and whistles Mac app. Perceptions are built on those sorts of decisions.
Because of that windows had thinking budget selectors for months before ios and macos (those got this only last week)
If you were going to release a product for developers as soon as it was ready for developers to try, such that you could only launch on one platform and then follow up later with the rest, macOS is the obvious choice. There's nothing contemptuous about that.
Windows is almost ready. It's already running but we are solving a few more things before the release to make sure it works well.
This will actually work well with my current workflow: dictation for prompts, parallel execution, and working on multiple bigger and smaller projects so waiting times while Codex is coding are fully utilized, plus easy commits with auto commit messages. Wow, thank you for this. Since skills are now first class tools, I will give it a try and see what I can accomplish with them.
I know/hope some OpenAI people are lurking in the comments and perhaps they will implement this, or at least consider it, but I would love to be able to use @ to add files via voice input as if I had typed it. So when I say "change the thingy at route slash to slash somewhere slash page dot tsx", I will get the same prompt as if I had typed it on my keyboard, including the file pill UI element shown in the input box. Same for slash commands. Voice is a great input modality, please make it a first class input. You are 90% there, this way I don't need my dictation app (Handy, highly recommended) anymore.
Also, I see myself using the built in console often to ls, cat, and rg to still follow old patterns, and I would love to pin the console to a specific side of the screen instead of having it at the bottom and pls support terminal tabs or I need to learn tmux.
> "Localize my app and add the option to change units"
To me this still feels like the wrong way to interact with a coding agent. Does this lead people to success? I've never seen it not go off the rails in some way unless you provide clear boundaries as to what the scope of the expected change is. It's gonna write code if you don't even want it to yet, it's gonna write the test first or the logic first, whichever you don't want it to do. It'll be much too verbose or much too hacky, etc.
The better models can handle that prompt assuming there is an existing clean codebase and the scope of the task is not too large. The existing code can act as an implicit boundary.
Weaker models give your experience, or when using a 100% LLM codebase I think it can end up in a hall of mirrors.
Now I have an idea to try, have a 2nd LLM processing pass that normalizes the vibe-code to some personal style and standard to break it out of the Stack Overflow snippet maze it can get itself in.
I've had no issues with prompts like that. I use Cursor with their plan mode, so I get a nice markdown file to iterate on or edit myself before it actually does anything.
100%
First phase: Plan. Mandatory to complete, as well as get AI feedback from a separate context or model. Iterate until complete.
Only then move on to the Second Phase: make edits.
Better planning == Better execution
Until a few days ago (when I switched to Codex), I would have agreed. My workflow was "thoroughly written issues" -> plan -> implement. Without the plan step, there is a high likelyhood that Claude Code (both normal or with GLM-4.7) or Cursor drift off in a wrong direction.
With Codex, I increasingly can skip the plan step, and it just toils along until it has finished the issue. It can be more "lazy" at times and ask before going ahead more often, but usually in a reasonable scope (and sometimes at points where I think other services would have gone ahead on a wrong tangent and burnt more tokens of their more limited usage).
I wouldn't be surprised that with the next 1-2 model iterations a plan step won't be worth the effort anymore, given a good enough initial written issue.
I still use tons of non-plan mode edits with cursor too. The example prompt above I'd plan it out first just to make sure it does it in a way I want since I personally know there are tons of ways to implement it. But for simple changes or when I don't want a plan on purpose I just use a normal agent.
And then
> gh-address-comments address comments
Inspiring stuff. I would love to be the one writing GH comments here. /s
But maybe there's a complementary gh-leave-comments to have it review PRs for you too.
OT: I never liked about codex how it didn't ask for confirmations before editing. While Claude has auto accept off by default I never understood why codex didn't have it. I want to iterate on LLMs edit suggestions.
Did they fix it?
Otherwise I'm not interested.
At least codex inside pycharm has auto edit off by default.
This looks interesting and I use Codex a fair bit already in vscode etc, but I'm having trouble leaving a 'code editor with AI' to an environment that sort of looks like it puts the code as a hidden secondary artefact. I guess the key thing is the multi agent spinning plates part.
(I work on Codex) I think for us the big unlock was GPT-5.2 and GPT-5.2-Codex, where we found ourselves needing to make many fewer manual edits.
I find that the case too. For more complex things my future ask would be something that perhaps formalized verification/testing into the AI dev cycle? My confidence in not needing to see code is directly proportional in my level of comfort in test coverage (even if quite high level UI/integration mechanisms rather than 1 != 0 unit stuff)
I'm still waiting for the big pivotal moment in this space, I think there is a lot of potential with rethinking an IDE to be Agent first, and lots of what is out there is still lacking. (It's like we all don't know what we don't know, so we are just recycling UX around trying to solve it)
I keep coming back to my basic terminal with tmux running multiple sessions. I recently though forked this https://github.com/tiann/hapi and been loving using tailscale to expose my setup on my mobile device for convenience (plus the voice input there)
There is little to no integration between deterministic IDE features(like refactorings) and LLMs. For example I don't want a statistical tool to rename a method by predicting tokens, I want it to use IDE features and not via another higher abstraction protocol like mpc, I want deeper integration. Sometimes I look at comments in code and think "why can't I have an agent checking if the content of a comment actually reflect the code below?" I feel like we're light years away from a killer integration
This might actually be another area language servers shine. As I understand it, the TS Language Server can do renames. Ergo, we ought to be able to have the LLM ask the lang server to do the rename instead of trying to do it itself. That'd be easier than trying to integrate with each IDE individually. (Whereby "IDE" seems to be synonymous with "VSCode" lately...)
Agree, another improvement i'd like along the lines or renames is lsp suggestions for method names, enums, functions, etc The llm should be able to autocomplete given lsp available symbols, this way it would avoid far less hallucinated methods
Claude, at least, already supports LSP servers though. https://code.claude.com/docs/en/plugins-reference#lsp-server...
Or maybe, hear me out, we don't need any of this ""agent"" first shiny thingy
yeah, TUIs for AI is just lazy work imho. I'm glad at least this time it's a macOS app, but it's still just a shitty chat interface
also this feels like a unique opportunity to take some of that astronomical funding and point it towards building the right tooling for building a performant cross-platform UI toolkit in a memory-safe language—not to mention a great way for these companies to earn some goodwill from the FOSS community
> For a limited time, Codex will also be available to ChatGPT Free and Go users to help build more with agents. We’re also doubling rate limits for existing Codex users across all paid plans during this period.
Is there more information about it? For how long and what are the limits?
They are probably providing it for free for 1 month.
no
I really look forward to using this. I tried Codex first time yesterday and it was able to complete a task (i.e. drawing Penrose tilings) that Claude Code previously failed at. Also a little overwhelmed by all the new features that this app brings. I feel that I'm behind all the fancy new tools.
It would be nice if it didn't have to be all local. I'd love a managed cluster feature where you could just blast some workloads off to some designated server or cluster and manage them remotely, share progress with teammates, etc. (Not "cloud" though; I'd still want them on the internal network). I imagine something like that is in the works.
I do it with ssh and tmux. I suppose tools could make it better.
Interesting timing for me personally as I just switched from running Codex in multiple tabs in Cursor to Ghostty. It had nicer fonts by default, better tab switching that was consistent with the keyboard shortcut to switch to any tab on Mac, and it had native notifications that would ping when Codex had finished. Worktrees requiring manual configuration was probably the one sticking point, so definitely looking forward to this.
Genuinely curious if people would just let this rip with no obvious isolation?
I’m aware Mac OS has some isolation/sandboxes but without running codex via docker I wouldn’t be running codex.
(Appreciate there are still risks)
Shameless plug, but you can sandbox codex cli without a container using my macOS app: https://multitui.com
This is a really nice tool! (Also, I love the old school animated GIFs in the site's footer.)
I wouldn't trust it. I'm moving to always running AI coding in a full VM.
(I work on Codex) We have a robust sandbox for macOS and Linux. Not quite yet for Windows, but working on that! Docs: https://developers.openai.com/codex/security
When can I get remote access in the iPhone app? Start on my laptop, check results using Tailscale/VPN and add follow up’s on the mobile to run on the computer. Know many that would love this feature.
ChatGPT can’t even write me a simple working AutoHotKey script so I’m not sure why I’d trust it with any actual coding. As I’ve done for about the past year with OpenAI showcases like this, this elicited an ‘Oh, that’s kinda neat, I’ll just wait for Gemini to do something similar so it will actually work’ from me.
I don't know you, but apart from ai tools race fatigue(feel pretty much like frameworks fatigue), all I see is mouse traveling a lot between far distant small elements, buttons and textareas. AI should have brought innovation even in UIs we basically stopped innovating there
Is everything OpenAI do/release now a response to something Anthropic have recently released?
I remember the days when it was worth reading about their latest research/release. Halcyon days indeed.
This is the 5th OpenAI product called Codex if I'm counting correctly
This is an ode to opencode and how openai, very strangely, is just porting layout and feature of real open-source.
So much valuation, so much intern competetion and shenanigans than the creatives left.
No.
I am glad to not depend on AI. It would annoy me to no ends how it tries to assimilate everything. It's like systemd on roids in this aspect. It will swallow up more and more tasks. Granted, in a way this is saying "then it was not necessary to have this things anymore now that AI solves it all", but I am skeptical of "the praised land" here. Skynet was not trusted back in 1982 or so. I don't trust AI either.
I'm the same way but I've got the gloomy sense that folks like us are about to be swept aside by the flood if we don't "adapt."
I got invites to seven AI-centered meetings late last week.
Same. And indeed, it's here. The genie is not going back into the bottle, so we have to learn how to live in this new world.
Eric Schmidt has spoken a lot recently about how it's one of the biggest advances in human history and it's hard to disagree with him, even if some aspects make me anxious.
One of the biggest advances in human history, and yet the owners of the technology with access to an unlimited number of "agents" using frontier models still can't release a desktop chat application without using Electron to bring in several hundred mb of bloat for displaying text. Someone's going to have to explain this one to me because the math is not mathing.
Exactly. If AI really worked, they would've released a native app. And it wouldn't take much to also get a Windows and a Linux native app, wouldn't it?
Apparently, the Codex app itself is proof that AI is not that good at doing what people think it does.
How come if I download code from GitHub, rename some stuff, and republish it under another license I’m a bad guy, but if I ask ChatGFY to do it for me I’m a 10x Chad? … someone is gonna figure that part out in court. I remember what code SCO used to make hay, and I know what side the MPAA, RIAA, Google, and NVidia Are gonna be on at the end of the day.
Replacing workers with things you can’t beat, sue, intimidate, or cajole? Someone is gonna do something to make that not cool in MBA land. I think if one of my employees LL-MessedUp something, and I were upset, watching that same person stealing my money haplessly turn to an LLM for help might land me in jail.
I kinda love LLMs, I’ve always struggled to write emails without calling people names. There’s some clear coding tooling utility. But some of this current hype wave is insano-balls from a business perspective. Pets.com X here’s-my-ssh-keys. Just wild.
I think a lot of AI talk doesn't explain where it shines the brightest (imo): Write the code you don't want to write.
I've recently had an issue "add VNC authentication" which covers adding vnc password auth to our inhouse vnc server at work.
This is not hard, but just a bit of tedious work getting the plumbing done, adding some UI for the settings, fiddle with some bits according to the spec.
But it's (at least to me) not very enjoyable, there is nothing to learn, nothing new to discover, no much creativity necessary etc. and this is where Codex comes in. As long as you give it clearly scoped tasks in an environment where it can use existing structures and convetions, it will deliver. In this case it implemented 85% of the feature perfectly and I only had to tweak minor things like refactor 1-2 functions. Obviously I read and understood and checked everything it wrote, that is an absolute must for serious work.
So my point is, use AI as the "code monkey". I believe most developers enjoy the creative aspects of the job, but not the "type C++ on your keyboard". AI can help with the latter, it will type what you tell it and you can focuse on the architecture and creative part of the whole thing.
You don't have to trust AI in that sense, use it like autocompletion, you can program perfectly fine without it but it makes your fingers hurt more.
i wonder if the skills will divide a bit. That there will be those who still program by hand - and this will be a needed skill, though AI will be a part of their daily toolset to a greater or lesser degree.
Then there will be the AI wranglers who act almost like DevOps engineers for the AI - producing software in a different way ...
Good luck.
I feel the same way about using the Internet or books to code. I'd rather just have the source code so that I'm not dependent on anything other then my own brain.
How does Codex mac app compare with Cursor? If anyone who tried both can explain here?
My experience with Cursor is generally good and I like that it gives me UX of using VS Code and also allows selection of multiple models to choose if one model is stuck on the prompt and does not work.
Coding agents with full automation like this require a different workflow that is almost purely conversational compared to Cursor/Windsurf/VS Code. It requires more trust in the model (but you can always keep Cursor open off to the side and verify its edits). But, once you get into the right rhythm with it, it can be _really_ powerful.
Is it open source? Do they disclose which framework they use for the GUI? Is it Electron or Tauri?
lol ofc not
looks like the same framework they used to build chatgpt desktop (electron)
edit - from another comment:
> Hi! Romain here, I work on Codex at OpenAI. We totally hear you. The team actually built the app in Electron specifically so we can support Windows and Linux as well. We shipped macOS first, but Windows is coming very soon. Appreciate you calling this out. Stay tuned!
Hey, that's great OpenAI. Now add about 6 zeroes to the end of the weekly token limit for your customers and maybe we could use the app
I guess the next it was meant to happen...I tried Google's Antigravity and found it quite buggy.
May give a go at this and Claude Code desktop as well, but Cursor guys are still working the hardest to keep themselves alive.
We are certainly approaching the point where a high end MacBook Pro for development isn’t required. Feels very close to just being able to use an iPad? My current workplace deploy on Vercel, we already test actively on feature branches and the models have gotten so good that you can reliably just commit what they’ve done (with linting and type check hooks etc) and in the rare event something is broken, follow up with a new commit.
Judging from the performance issues with Claude Code, you won't be able to run a decent agentic cli or desktop workflow orchestrator on anything other than mbp.
* Replace MacBook with any high-end machine that requires a bunch of configuration for local dev. And replace iPad with any thin client that’s much more portable and convenient, eg glasses that display a screen and Bluetooth keyboard. Why shouldn’t I be able to describe the changes I want to an LLM from a coffee shop?
More simple and similar app: vibe-kanban
Is there any marked difference or benefit over Claude Code?
It’s possible to run up to 4 agents at once vs. Claude Code’s single thread. Sometimes I’ll find meaningful quality differences between what agents produce.
Interesting. Has anyone found running multiple parallel agents useful in practice?
You'd save time compared with running them in serial obviously?
I’ve been using codex regularly and it’s pretty good at model extra high with pretty generous context.
From the video, I can see how this app would be useful in:
- Creating branches without having to open another terminal, or creating a new branch before the session.
- Seeing diff in the same app.
- working on multiple sessions at once without switching CLI
- I quite like the “address the comments”, I can see how this would be valuable
I will give it a try for sure
Wow, this is nearly an exact copy of Codex Monitor[1]: voice mode, project + threads/agents, git panel, PR button, terminal drawer, IDE integrations, local/worktree/cloud edits, archiving threads, etc.
Codex Monitor seems like an Antigravity Agent Manager clone. It came out after, too.
Bunch of the features u listed were already in the codex extension too. False outrage it its finest.
I have both Codex Monitor and this new Codex app open side by side right now; aside from the theme, I struggle to tell them apart. Antigravity's Agent Manager is obviously different, but these two are twins.
I have a very hard time getting worked up over this. There are a ton of entrants in this category, they all generally look the same. Cribbing features seems par for the course.
Antigravity is a white labeled $2B pork of Windsurf, so it really starts there, but maybe someone knows what windsurf derived from to keep the chain going?
cursor?
from what I can tell, the people behind windsurf were at it first
oh, codeium? that was them?
Maybe github copilot then
Bugs me they treat MacOS as first class. Do people actually develop on a Mac in 2026? Why not just start with Linux?
I mean if they were targeting "software engineers" in general then Windows would be the obvious choice in 2026 as much as in 2006. But these early releases are all about the SF bubble where Mac is very much dominant.
Really? I frankly don’t know anyone who’s not on Linux. If you do any AI/ML you basically find yourself on a Linux box eventually. Perhaps I live in a bubble.
Surely it varies a lot and everyone is in an industry bubble to some extent, but from my experience in some non-tech industries (healthcare, manufacturing), Linux workstations were nonexistent and working with the Linux servers was more a sign of an ops role. People who wrote code for a living didn't touch them directly. Last StackOverflow survey [1] puts it at something like 50% use Windows at work, 30% Mac, 20-40% Linux (breakdown of WSL and Ubuntu as categories seems confusing, maybe the raw data is better).
[1] https://survey.stackoverflow.co/2025/technology/#1-computer-...
Can't build iOS apps on anything else sadly.
Maybe it's because I'm not used to the flow, but I prefer to work directly on the machine where I'm logged in via ssh, instead of working "somewhere in a git tree", and then have to deploy/test/etc.
Once this app (or a similar app by Anthropic) will allow me to have the same level of "orchestration" but on a remote machine, I'll test it.
I typically bounce between Claude Code and Codex for the same project, and generally enjoy using both to check each other.
One cool thing about this: upon installing it immediately found all previous projects I've used with Codex and has those projects in the sidebar with all of the "threads" (sessions) I've had with Codex on these projects!
Does anybody know when Codex is going to roll out subagent support? That has been an absolute game changer in Claude Code. It lets me run with a single session for so much longer and chip away at much more complex tasks. This was my biggest pain point when I used Codex last week.
It's already out.
Can you explain how to use it? I’ve tried asking it to do “create 3 files using multiple sub agents” and other similar wording. It never works.
Is it in the main Codex build? There doesn’t seem to be an experiment for it.
I'm excited to try this out, it seems like it would solve a lot of my workflow issues. I hope there is the ability to review/edit research docs and plans it generates and not just code.
This does look like it would simplify some aspects of using Codex on Mac, however, when I first saw the headline I thought this was going to be a phone app. And that started running a whole list of ideas through my brain... :(
But overall, looks very nice and I'm looking forward to giving it a try.
I don't know why any frontier model lab can't ship a mobile app that doesn't use a cloud VM but is able to connect to your laptop/server and work against local files on there when on the same network (e.g.: on TailScale). Or even better act as a remote control for a harness running on that remote device, so you couldn't seamlessly switch between phone and laptop/server.
I'm also so baffled by this. I had to write my own app to be able to do seamless handoff between my laptop/desktop/phone and it works for me (https://github.com/kzahel/yepanywhere - nice web interface for claude using their SDK, MIT, E2E relay included, no tailscale required) but I'm so baffled why this isn't first priority. Why all these desktop apps?
i've been using ai vibe coding tools since Copilot was basically spicy autocomplete, and this feels like the next obvious step: less “help me type” and more “please do this while I watch nervously.” The agent model sounds powerful, but in practice it’s still a lot of supervision, retries, and quiet hope it doesn’t hallucinate itself into a refactor I didn’t ask for.
So if the agent struggles even when you are working with it, how will it be better working alone? This is why I never let the agent work by themselves. I'm very opinionated about my code.
Does it somehow gain some superpower from being left alone?
I'm managing context with codex inside VSCode using different threads. I'm trying to figure out if there are use cases where I'd rather be in this app.
Having dictation and worktree support built in is nice. Currently there is a whole ecosystem of tools implementing similar functionality for Claude Code. The automations look cool too!
Is this not just a skinned version of Goose: https://block.github.io/goose/
they are all copies of each other. Did you expect them to build something completely new? Software Development is stuck in an AI hole where we only build AI features.
In the end this and all other 89372304 AI projects are just OpenAPI/Anthropic API wrappers, but at least one has 1st party support which maybe gives it a slight advantage?
How is this better than vscode with the codex extension?
Not to rain on the parade, but this app feels to me ... unpolished. Some of the options in the demo feels less thought out and just put together.
I will try it out, but is this just me, or product/UX side of recent OpenAI products are sort of ... skipped over? It is good that agents help ship software quickly, but please no half-baked stuff like Altas 2.0 again ...
What else do you expect from vibecoding? Even the announcement for this app is LLM generated.
I don’t get why they announce it as a “Mac app” when the UI looks and feels nothing like a Mac app. Also electron… again.
Why not flex some of those codex skills for a proper native app…
These paid offerings geared toward software development must be a hell of a lot "smarter" than the regular chatbots. The amount of nonsense and bad or outright wrong code Gemini and ChatGPT throw at me lately is off the charts. I feel like they are getting dumber.
Yes they are, the fact that the agents have full access to your local project files makes a gigantic difference.
They do *very* well at things like: "Explain what this class does" or "Find the biggest pain points of the project architecture".
No comparison to regular ChatGPT when it comes to software development. I suggest trying it out, and not by saying "implement game" but rather try it by giving it clear scoped tasks where the AI doesn't have to think or abstract/generalize. So as some kind of code-monkey.
I don’t understand why we are getting these software products that want to have vendor lock in when the underlying system isn’t being improved. I prefer Claude code right now because it’s a better product . Gemini just has a weird context window that poisons the rest of the code generated (when online) ChatGPT Codex vs Claude I feel that Claude is a better product and I don’t use enough tokens to for Claude Pro at $100 and just have a regular ChatGPT subscription for productivity tasks .
> I don’t understand why we are getting these software products that want to have vendor lock in when the underlying system isn’t being improved.
I think it's clear now that the pace of model improvements is asymptotic (or at least it's reached a local maxima) and the model itself provides no moat. (Every few weeks last year, the perception of "the best model" changed, based on basically nothing other than random vibes and hearsay.)
As a result, the labs are starting to focus on vertical integration (that is, building up the product stack) to deepen their moat.
> I think it's clear now that the pace of model improvements is asymptotic
As much as I wish it were, I don't think this is clear at all... it's only been a couple months since Opus 4.5, after all, which many developers state was a major change compared to previous models.
Like I said, lots of vibes and hearsay! :)
The models are definitely continuing to improve; it's more of a question of whether we're reaching diminishing returns. It might make sense to spend $X billion to train a new model that's 100% better, but it makes much less sense to spend $X0 billion to train a new model that's 10% better. (Numbers all made up, obviously.)
It’s the inconsistency that gets me. Very similar tasks, similar complexity, same code base, same prompting:
Session A knocks it out of the park. Chef’s kiss.
Session B just does some random vandalism.
seems like I need to update my toolset for the 3rd time this week
Eh. Kicked the tires for a few minutes. Back to the old clunker app.
No worries. I'm not their target demographic, anyway.
Currently using opencode with Codex 5.2 and wondering why I should switch.
Does this support users who access Codex via Azure OpenAI API keys?
For pure code generation is ChatGPT 5.2 so much better than Claude opus 4.5 thinking to have me switch? I’m basically all in on Claude.
Sure I could move to open code and use them as commodities but I’ve gotten use to Claude code and like using the vendors first party app.
Built an open source lightweight version of this that works with any cli agent: https://github.com/built-by-as/FleetCode
This is so garbage. OpenAI is never catching up.
The inclusion of a live vibe-coded game on the webpage is fun, except the game barely works and it's odd they didn't attempt any polish/QA for what is ostensibly a PR announcement. It just adds more fuel to the fire to the argument that vibecoding results in AI slop.
To be fair the premise is that they 1 shotted it. I'd just be suspicious if it were any better (the POC is that is just about works)
I agree, if it had been polished I would have not trusted the demo at all, the fact it shows what you can potentially expect from a one-shot is cooler.
What are the max context sizes ?
Kind of embarrassing to demo "Please change this string to gpt-5.2". Presumably the diff UI doesn't let you edit the text manually? Or are they demonstrating being so AI-brained you refuse to type anything yourself?
> and we're doubling the rate limits on Plus, Pro, Business, Enterprise, and Edu plans.
I love competition
Maybe I'm just not getting it, but I just don't give a flying fuck about any of this crap.
Like, seriously, this is the grand new vision of using a computer, this is the interface to these LLMs we're settling on? This is the best we could come up with? Having an army of chatbots chatting to each other running basic build commands in a terminal while we what? Supervise them? Yell at them? When am I getting manager pay bumps then?
Sorry. I'll stick with occasionally chatting with one of these things in a sandboxed web browser on a single difficult problem I'm having. I just don't see literally any value in using them this way. More power to the rest of you.
OK.
It keeps offering me to "Get Plus" even though I am signed and already have a Plus plan.
Codex really grown on me lately. I re-signed to try it out on a project I have and it turned out to be really great addition to my toolkit.
It isn't always perfect and it's cli (how I mostly use it) isn't as sophisticated as OpenCode which is my default.
I am happy with this app, I am using Superset, terminal app which suprisingly is well positioned to help you if you work in cli like I do. But like I said, new desktop app seems like a solid addition.
> Work with multiple agents in parallel
But you can already do that, in the terminal. Open your favourite terminal, use splits or tmux and spin up as many claude code or codex instances as you want. In parallel. I do it constantly. For all kinds of tasks, not only coding.
But they don't communicate. These do.
Does the Codex app host MCP Apps?
I really want to like the native Mac app aesthetic but I kinda hate it. It screams minimalist but also clearly tells me it’s not meant for a power user. That ruggedness and sensitivity is missing.
No Linux support? :(
Another boring update from OpenAI. Why would I want an orchestration tool tied to one model? Part of the value of orchestration tools is using the most appropriate and/or cost effective model for the task, sub-task, etc
> We're also excited to show more people what's now possible with Codex . For a limited time we're including Codex with ChatGPT Free and Go, and we're doubling the rate limits on Plus, Pro, Business, Enterprise, and Edu plans.
Translated from Marketingspeak, this is presumably "we're also desperate for some people to actually use it because everyone shrugged and went back to Claude Code when we released it".
I dunno, feels like the models have different weak/strong points, sometimes I can sit with Claude Code for an hour with some issue, try it with Codex and have it solved in five minutes, and also the opposite happens. I tend to use Codex mostly when I care more about correctness and not missing anything, Claude when it's more important I do it fast and I know exactly what it needs to do, Codex seems to require less hand-holding. Of course, just anecdotal.
GPT models definitely seem stronger when they "get it" and in the types of problems they "get", while claude seems more holistic but not "as smart" as some of the spikes GPT can get.
Yeah this is clearly just a marketing re-release but if they've executed well i'm happy to try it
They also claim 2x usage from December (though 2x a tiny amount is still a tiny amount)
Given the prevalence of Opencode and its ability to use any model and provider I don't see reason why would anyone bother with random vendors half-assed tools.
For starters, money. There is no better value out there that I'm aware of than Claude Code Max. Claude Code also just works way better than Opencode, in my experience. Though I know there are those that have experienced the exact opposite.
Are you calling OpenAI a random vendor?
That's like calling Coca Cola a random beverage vendor