OpenAI Codex quietly powers 70 apps, from terminals to design tools

OpenAI’s Codex, the natural language-to-code engine that debuted last August, has moved well beyond the GitHub Copilot hype cycle. The company just disclosed that the model is now baked into 70 different applications through its API, touching everything from Rust-based terminals to Figma design toolchains. It’s a quiet but significant milestone that suggests Codex is becoming infrastructure rather than a novelty.

Microsoft is also widening the pipes. At its Build conference, the company announced that Azure OpenAI Service—the enterprise gateway to Codex, GPT-3, and embeddings—is shifting from invite-only to a limited access preview. That means more healthcare and financial services firms will soon have a sanctioned, compliance-friendly on-ramp to the same tech that’s been fueling indie developer experiments for months. The enterprise play here is real, and it’s moving faster than a lot of skeptics predicted.

The 70-app figure includes the usual suspects like Replit, which uses Codex to explain code snippets in plain English for learners, and Warp, the Rust terminal that lets developers search for commands using natural language instead of spelunking through Stack Overflow. But the more interesting story is in tools like Pygma and Machinet. Pygma converts Figma designs directly into frontend code, matching a developer’s existing style. Founder Emile Paffard-Wray noted that features requiring flexible code-tweaking would have been “incredibly difficult to build in-house” without Codex. Machinet, meanwhile, pivoted from building its own ML models to using Codex for generating Java unit test templates, accelerating their development several-fold.

The common thread across these integrations is that Codex is doing the grunt work developers resent—scaffolding tests, explaining legacy code, converting visual designs into skeleton components. GitHub Copilot, which now suggests roughly 35% of the code its users write and moves to general availability this summer, set the template. But the broader ecosystem is discovering that the model’s real value isn’t writing novel algorithms. It’s eliminating the friction between intent and implementation. The terminal command you can’t remember, the boilerplate you’ve typed a hundred times, the design mockup that needs to become React components by Friday—Codex is eating those tasks. Whether that meaningfully boosts productivity or just shifts the bottleneck to code review and architecture is the question nobody’s answering yet.