· 7 min read

OpenAI Just Turned Codex Into an App Builder That Deploys. If You Sell 'We'll Build Your Internal Tool,' Read This First.

AI Tools Stack

On June 2, OpenAI previewed Codex Sites: describe what you want, and Codex builds the app, runs it to test, deploys it, and hands you back a hosted, shareable URL. No deploy pipeline, no hosting setup, no handoff. It shipped alongside a set of role-specific Codex plugins aimed at analysts, marketers, sales, designers, and more: non-engineers, in other words. OpenAI's own framing is telling: Sites isn't chasing Squarespace and the public-website market. It's aimed at the internal apps every company builds, half-finishes, and abandons because nobody wants to maintain them.

That's the exact job a lot of solo devs and two-person agencies sell. "We'll build you an internal dashboard / a CRUD tool / a little admin panel" has been reliable $2K-to-$5K starter work for years. The floor under that work just moved, and it's worth being clear-eyed about which part moved and which part didn't.

What actually shipped, and what didn't

First, scope it accurately, because the panic version is wrong. Codex Sites is in preview, and as of launch it's available on ChatGPT Business and Enterprise plans only. If you're a solo builder on Plus or Pro (where most indie hackers live), you can't even touch it yet. So this is not "OpenAI just deleted your business this week." It's "OpenAI just told you, in a product, where the next two years are heading."

What it does inside that preview is the part to take seriously. A non-engineer at a company that already pays for ChatGPT Enterprise can now type "build me a dashboard that tracks our support tickets by team and flags anything over 48 hours" and get a working, hosted thing back, shareable by URL, without filing a ticket with IT or hiring you. The apps it's pitched for (dashboards, planners, project boards, simple internal tools) are precisely the low-complexity end of the contract work that solo devs use to land clients.

The part of the market that's getting squeezed

Be honest about which jobs these are. The vulnerable work is the self-contained internal app with clean requirements and no nasty integration: a form that writes to a table, a dashboard over a CSV or a single clean data source, an admin view a founder sketches on a whiteboard. That work was never hard. It was valuable because the client couldn't do it themselves and didn't want to learn. Both halves of that just weakened. The client increasingly can do it themselves, and the tool to do it is sitting inside software they already pay for.

If that specific job, the simple, self-contained internal tool, is the core of how you make money, treat this launch as a two-year warning shot, not a someday-maybe. The preview gate and the Enterprise-only limit are buying you time, not granting you immunity.

Where the work moves instead of disappears

Now the other half, because "the floor moved" is not "the building collapsed." Prompt-to-app gets you the 80% that was always easy. It does not get you the 20% that was always the actual job: the integration with a messy legacy system that has no clean API, the data that's wrong in three places and has to be reconciled before any dashboard means anything, the access controls and compliance requirements a regulated business can't hand-wave, the edge cases that only show up against real production data. That's the work a business genuinely cannot self-serve from a prompt, because the hard part isn't writing the app. It's understanding the specific mess the app has to survive.

I've shipped enough of these to know the secret: clients were rarely paying me for the code. They were paying me to absorb the ambiguity: to sit with their actual operations, figure out what they meant instead of what they asked for, and own the thing when it broke at 9pm. A prompt can write a CRUD app. It can't sit in a meeting and notice that the "simple" ticket dashboard has to reconcile two systems that disagree about what a "closed" ticket is.

What I'd actually do

Stop selling the artifact. The app is now the cheap part, and pricing yourself as "the person who produces the app" puts you in a race against a feature inside ChatGPT Enterprise. Sell the outcome and the ownership instead. Charge for the integration, the data correctness, the "it keeps working and someone is accountable for it": the parts that don't fit in a prompt. Treat the generated app the way you'd treat AI-generated code anywhere: as a fast first draft that still needs someone who knows what they're doing to make it production-real and keep it that way.

And reposition the cheap starter project. If your funnel depended on a $3K dashboard to land a client you'd later upsell, that funnel is leaking. Move the entry point to something the prompt can't deliver (a discovery engagement, a data-cleanup project, an integration audit) so the first thing you sell is judgment, not output.

The honest counter-take

Let me push against my own warning. It's very possible Codex Sites stays mediocre at the exact thing that matters (the messy, integrated, real-world app) for a long time, the way a lot of "build an app from a prompt" tools demo beautifully and fall apart on contact with an actual business. The category has overpromised before. And the Enterprise-only gate could persist, keeping this firmly in the "big companies' internal IT" lane and nowhere near the small-business clients many solo devs actually serve. If that's how it plays out, the simple-dashboard job survives longer than I'm suggesting.

But I wouldn't bet my pricing on the tool staying bad or staying gated. The cheaper bet, the one that's right whether Codex Sites is great or a dud, is to make sure the thing you charge for is the judgment and the ownership, not the keystrokes. The keystrokes were always going to get automated. Be the person clients call when the automated version isn't enough, because that person is the one with a business in two years.

Author

Lukas

Sources

More AI articles →

Stay in the Loop

Get new posts delivered to your inbox. No spam, unsubscribe anytime.

Newsletter coming soon. Set PUBLIC_CONVERTKIT_FORM_ID in .env to activate.

Related Posts