Testing OpenAI Codex App for Local Development

I've been testing the new OpenAI Codex app for a week. For those catching up, it's a GUI on top of the codex harness focused on the annoying problem of agentic local development: git (worktrees, branches, commits, PRs) and working across local + the cloud. # What makes Codex different? ## Very focused Well first, it's not a TUI. It's not a chatbot. It's not an IDE. It's... something else. Codex strips away everything but the primitives it's focused on, which are: • Projects (basically, a repo) • Threads (chat) • Git things (worktrees, branches, diffs, commits, PRs) • Automations • and Skills I guess they realized everyone is yelling at their coding agent "SORRY WRONG BRANCH" and packaged up some of the core workflows that AI-powered engineers use into a graphical interface. ## Skills - now with icons! Codex also addresses my primary gripe with skills, which is they are very powerful but seem less cool when presented as markdown + code files (or worse, a .zip) In the Codex app, skills are they're own page, and installable with a button. It's obvious in the UI when you've tagged in a skill to an agent, similar to how you'd @ mention a file or branch in Cursor. The library of skills is a mix of technical and non technical use case, which point to the discussion I always have with these platforms which is: "who are we building for"? To be honest, I'm still partial to a npx skills add from skills.sh, and I suspect most software engineers are, too. ## Automations Now this, I love. Because codex can work in the cloud, the ability to automate work (yes, while you sleep) and using nice UI to set it up is awesome. A lot of these automations are things I know many developers are already doing via github actions or some other agent, and it's nice to see this get first class treatment, instead of being a tagged on hacked use case. # My coding experience with Codex I tried a few things with Codex including • Restructure of a components folder • Reviewing complex work executed by Claude Code for quality • Drafting a technical spec for a personal agent • Checking the codebase for a specific inconsistency OpenAI has great coding models. Everyone has great coding models! We're spoiled for choice. I love GPT-5.2-Codex for backend things. I use it to address all my Bugbot feedback. It's a great model. Unfortunately, my experience with the Codex harness blunts my appreciation for the power of the model. You all know I'm a @cursor_ai daily driver, and it's clear that they've filed away a lot of the sharp edges of using a lot of different models in their app. Codex isn't as polished here, which meant coding with Codex felt so. dang. slow. Not because the model was particularly slow (though there was some of that), but because the harness kept getting in the way of me going yolo. "Always run in this session" didn't always run commands within the session. The app was waiting on me.... a lot. The integration between terminal, browser, etc. was behind other platforms and it made my overall developer experience sub par. Also, this is a personal thing, but she's so dry. I love an efficient communicator but if I'm going to spend the day with you, I need some hints of friendly engagement. But alas, I've worked with senior software engineers enough to know heart and head are not always aligned 🙃 It also wasn't driving me enough to next steps. For example, after I finished some nits in a technial spec, it never said: "ready to build"? These sorts of harness / prompt optimizations will make a huge difference when people decide what to reach for when coding. # Again, who are we building for? So Codex is a nice, opinionated app, focused on getting all the code that Codex models are capable of spinning out actually in production without too much babysitting, focused on the core development workflow of git branches and PRs. But as someone who seems to be the only person in the world who uses github desktop, I can't name another developer who has 15 worktrees going locally and also doesn't have a) claude code managing git via CLI or b) has memorized every git CLI command down cold. Codex Github GUI - better than GH Desktop App? Hi, it's me, the only user of the Github Desktop App Codex is a lovely if you're new to development and trying to learn the basics of code, git, skills, and automations. It might be a nice interface for PMs or designers pushing code. But OpenAI will have to "friendly" up the harness to match that persona, who has been trained to love that little lollygagging mascot in their terminal. # Build fast, learn things All that being said, the @OpenAIDevs is the best at engaging developers on early releases. Every bug, nit, weird response gets ack'd by the team, who was shipping iterative releases all through the week I tested. Sorry I love to test software I suspect we'll see Codex evolve over time, and I'm keeping an eye on harness updates. In the meantime, the Codex models will be staying in my collection of faves, and the Codex app can continue to try to lure me from the dock with every update.