Codex, leveled up: a major upgrade to OpenAI's coding agent in TransistorKit

TransistorKit lets you pick the AI engine that builds your app — Claude, Grok, local models, and OpenAI's Codex among them. Codex is a favorite for a lot of you, so we spent real time making it a first-class citizen. Here's what changed and why it matters.

Codex now runs the whole build — not just the conversation

The biggest one. When you chose Codex before, it was great at talking through a plan… and then it would stop short of actually doing the work. You'd get the strategy without the build.

That gap is closed. Pick Codex and it now does what every other TransistorKit engine does: it breaks your idea into real tasks and puts builders to work across your platforms — iOS, macOS, web, and the rest — until you have something that runs. One prompt in, a building app out.

Usage and cost you can actually trust

If you're paying per token, you want to see where it goes. Codex's usage now shows up in your usage view, correctly attributed to Codex, so the running cost estimate is accurate while you build. No more numbers quietly going missing.

Always your latest Codex

TransistorKit now uses the Codex you've installed — for most people that's the Homebrew version — instead of an older copy bundled alongside it. The practical payoff: you get the newest models and options the moment they ship, and those cryptic "unknown option" version errors that could stop a build mid-flight are gone.

Built for GPT-5's enormous context

Codex's newer models manage their own context automatically — they keep the conversation healthy without you babysitting it. TransistorKit now understands that and gets out of the way, instead of flashing a confusing "you're over the limit" meter on a session that's perfectly fine. Long, ambitious builds just keep going.

It plays nicely with fallback

Mid-build, a provider runs low on credits — it happens. With your model fallback order set in Settings, TransistorKit can hand the work off to Codex (or hand it back) automatically and keep moving. You set the priority once; the build doesn't stall waiting on you.

The upshot

Codex went from "a good chat partner" to "an engine that ships." If you tried it earlier and bounced off because it wouldn't pull the trigger on the actual work, it's worth another look.

To try it: update to the latest TransistorKit, then choose Codex under Settings → Models, and describe what you want to build. Same flow you already know — now with OpenAI's coding agent doing the heavy lifting.

As always, we built a lot of this in response to what you told us. Keep it coming.