Codex Sites vs v0, Bolt & Lovable: prompt-to-app

The short version: all four are prompt-to-app tools, but they differ on where they run, what backend they give you, whether you can take the code, and who they are built for. Codex Sites lives inside ChatGPT and Codex and hosts on OpenAI infrastructure, with D1 and R2 for data and a workspace-first access model — its sweet spot is internal tools. v0 generates front-end UI as React and Next.js code you can copy out or ship to Vercel. Bolt runs an editable full-stack app live in your browser. Lovable builds full-stack apps for non-developers on a Supabase backend with GitHub sync. Pick by stage and audience, not by hype.

They all start with a prompt — so what differs?

In 2026, "describe an app and get one" is table stakes. Every tool here does it. The interesting questions are the ones the demo videos skip:

• Where does it run and deploy? Inside an existing tool, on a specific cloud, in your browser, or on the vendor's own host.
• What is the backend and data story? A built-in database and storage, a partner like Supabase, or bring-your-own.
• Can you take the code? Copy-out, GitHub sync, or a more closed loop.
• Who is it for? Front-end engineers, full-stack tinkerers, non-technical founders, or workspace teams.

Those four axes separate the field more than raw generation quality does. Let's put them side by side, then take each tool on its own terms.

The four-way comparison

We have kept this fair. Where Codex Sites genuinely doesn't document something the others have, we say "not documented in preview" rather than scoring it a zero — and where the others have a clear edge, we say so.

Two rows deserve a note. On export and ownership, the three non-OpenAI tools all document a way to take your code with you today — v0 outputs standard React you can paste anywhere or sync to GitHub, Bolt gives you an editable project and GitHub export, and Lovable does bi-directional GitHub sync into a repo you own. The Codex Sites docs don't describe export, GitHub sync, or custom domains during preview, so we list those as not documented rather than asserting they're absent. You can confirm the current state on the official Codex Sites documentation.

Codex Sites: a hosted app from a prompt, workspace-first

Codex Sites is the Sites plugin for OpenAI Codex. You describe what you want inside ChatGPT or Codex, name the Sites plugin, and it can create, save, deploy, and inspect a hosted site, web app, or game — no separate deploy workflow. Under the hood it builds Cloudflare Worker-compatible output as ES modules, with D1 for relational data and R2 for object storage.

Two things make its shape distinct. First, every deployment URL is production — there is no throwaway preview URL. The review gate is to save a version without deploying it; a saved build ties to your source Git commit and is reviewable, but it isn't live until you deploy. Second, it is workspace-first: new sites start restricted to owner and admins, and you choose the audience (admins, the whole workspace, or a custom set) before sharing. Business workspaces have Sites on by default; Enterprise requires an admin to enable it via RBAC.

The honest caveat: it's in preview, free for now with pricing to be announced, and the portability story — export, GitHub sync, custom domains — isn't documented yet. For the fuller picture, see what Codex Sites is and how to use Codex Sites.

v0: generate the UI, ship it to Vercel

v0, by Vercel, is the front-end specialist of the group. You describe an interface and it produces React and Next.js components styled with Tailwind and shadcn — clean, idiomatic code you can copy straight into an existing project. Its 2026 updates added Git sync, an in-tool editor, and database connectivity, but the center of gravity is still UI generation, and the deploy path runs naturally to Vercel with shareable preview URLs.

Reach for v0 when the hard part is the interface, when you already have a codebase to drop components into, or when you're a front-end engineer who wants generated UI that matches how you'd hand-write it. It does less of the full backend wiring than Bolt or Lovable, by design.

Bolt: a full-stack app you can edit live in the browser

Bolt (bolt.new, by StackBlitz) runs the whole development environment inside your browser tab using WebContainers — Node.js via WebAssembly, no remote VM. Prompt it and you get a running, editable full-stack app you can poke at immediately, with integrations for backends and one-click deploys to hosts like Netlify, Vercel, or Cloudflare.

Bolt's edge is the live, hands-on loop: you see the app run, tweak code or re-prompt, and watch it update in place. That makes it strong for fast full-stack prototyping when you want to stay close to the code. If you'd rather describe an app and never touch the editor, a more guided tool may suit you better.

Lovable: full-stack apps for non-developers

Lovable is built for the non-technical founder, indie hacker, or PM who wants a working app without learning to code. It generates a full-stack app — frontend, backend, auth — commonly on a Supabase backend (Postgres, authentication, storage, edge functions), and deploys to its own hosting with one click. Crucially, it does bi-directional GitHub sync into a repo you own, so a developer can take over later.

Pick Lovable when the goal is shipping a real, shareable app with the least technical friction, and when "a developer can graduate it later" matters. It abstracts more of the stack than the others, which is the point.

How to choose

Decide by what you're actually making:

• A workspace dashboard or internal tool, from a prompt, by a team already in Codex? Codex Sites. The owner+admins default, RBAC, and built-in D1/R2 fit internal apps — just lean on save-without-deploy as your review gate, and accept the preview gaps.
• UI components or a front-end you'll fold into an existing app? v0, especially on a Next.js/Vercel stack.
• A full-stack prototype you want to edit live and deploy fast? Bolt.
• A complete app shipped by a non-developer, with code you can hand to an engineer later? Lovable.

The mistake we see is forcing one tool to do every job. And a step earlier than all of them: pressure-test the brief before you prompt. Ask GPT, Claude, and Gemini the same plan — pages, data model, auth, what "done" means — and compare answers in one place on oran.chat, branching the conversation instead of overwriting it so a sharper draft survives. oran.chat is a thinking tool, not a builder; it ends where these four begin.

Where this fits

This is one entry in our Comparisons cluster. If you're weighing prompt-to-host against general-purpose hosts, read Codex Sites vs Vercel and Netlify. For the broader market view, the pillar is the best ChatGPT alternatives in 2026. The throughline: all four start with a prompt, but they diverge on where they run, the backend they hand you, whether you can take the code, and who they're for. Match the tool to the stage and the audience.