Share what you made in ChatGPT, with or without Sites

ChatGPT Sites is OpenAI's own hosting: you describe a website or lightweight app, ChatGPT builds a private preview, and you deploy it to a URL from the same conversation. You can start the same product from Codex.

• Who gets it: Plus and Pro accounts, and Business and Enterprise workspaces. It isn't on Free or Go, or in the ChatGPT Classic app.
• Where it starts: Work on the web, or Work or Codex in the desktop app. Use the word "website" in your prompt, or mention @Sites.
• What it can run: web apps with server-side code, a D1 database of up to 10 GB per site, R2 file storage, secrets and Sign in with ChatGPT. OpenAI says some frameworks, databases and background services aren't supported.
• Who can open it: at first only the owner and workspace admins. Depending on your plan and settings you can widen that to selected people, the workspace, named outside viewers or anyone on the internet. Public publishing is off by default on Enterprise.

Use Sites if your plan has it and you want ChatGPT to build and host in one place. Use host0 when the file already exists, you're on Free or Go, or you want to share privately with people who don't have ChatGPT accounts.

Sites wins outright if you need server code, Sign in with ChatGPT, visitors' own connected apps, a tunnel to a private network, or scheduled updates. host0 has none of those. If you're already paying for ChatGPT and the site lives entirely in ChatGPT, you may never need us.

Hosting is free through December 31, 2026, and building or editing a site counts toward your Codex usage. From January 1, 2027, OpenAI plans to meter hosting and charge only for use above your plan's monthly allowance.

OpenAI calls these rates proposed, and all the credit amounts approximate. It also says one page visit typically makes 30 to 50 requests. Enterprise follows its own agreement, and hosting charges are waived there for now.

Custom domains work where they're available, not in Enterprise workspaces at launch. You bring a domain you already own and add the DNS records Sites gives you. On host0: Pro includes 3 custom domains per account; Free has none.

Mainly three things: every deployment is live, restoring a version doesn't bring back data, and some kinds of data aren't allowed on a site at all. These come from OpenAI's docs and hold whether or not you ever use host0.

• To review a build before it goes live, ask ChatGPT to save a version without deploying it. Deploying an existing site updates what visitors see.
• Publishing can change the database before the new version is live. If publishing fails, some of those changes may already be saved, so look at the database before you retry.
• There's no built-in way to back up database records and uploaded files together and restore them into a new site. Codex can export the source code.
• A deleted site can't be restored. To take one down for now, restrict who can open it instead.
• Sites must not process protected health information or payment-card data, apart from cards handled entirely by a third-party payment processor.
• On Plus and Pro, conversations about your site can train OpenAI's models if "Improve the model for everyone" is on. Business and Enterprise conversations aren't used by default.

No, and canvas is on its way out. OpenAI's release notes say it's no longer in GPT-5.5 Instant or GPT-5.5 Thinking, and that paid users can keep it for a limited time on legacy models until those are retired.

A canvas was an editing surface in the conversation, and sharing one worked much like sharing the conversation, including rendered HTML or React. If you still have a page in an old canvas, copy the code into a file named index.html. From there it can go to host0, or into a new prompt for Sites.

Download what ChatGPT made and drop it on host0 yourself. host0 has no way to deploy from inside a ChatGPT conversation, so asking in the chat won't publish anything.

• A page: ask for one self-contained HTML file with its CSS and JavaScript inline. OpenAI documents downloads in Work for documents, spreadsheets, presentations and PDFs but doesn't promise HTML on the web, so if you get a code block instead, paste it into a plain-text editor and save it as index.html.
• Several files: ask for a zip with index.html at the top. Your browser unpacks it, and it's checked like a folder.
• A PDF: Work can create one for you to download. Drop it on the PDF page.
• Images that never arrived: if the code points at pictures ChatGPT didn't give you, ask it to inline them or take them out, or they'll show as broken.

The first drop asks you to sign in. You get a random URL you can rename, and it's private until you change that. On Free, a file can be up to 3 MB.

We don't offer that from Work, because nobody here has run it. Codex in the desktop app is the closest thing that should work today, and we haven't run that either.

• Work on the web runs shell commands in a managed cloud environment and can use skills, so the host0 skill might run there. It would have to reach host0.ai and host0.app, and cloud work is limited to approved websites, with network access under your workspace admin's control.
• Codex in the desktop app runs commands on your computer inside a sandbox, and it asks when a command needs to go beyond that. That's the same flow as our Codex page, so expect an approval for the upload.
• Chat is for back-and-forth, so the setup prompt at the top of this page is no use there: it's written for agents that can run commands.

Yes, if ChatGPT writes it for host0: an h0.json declaring the collections, and a page that reads and writes /.h0/data/<collection>. It can't guess that contract, so you have to tell it.

• Paste the h0.json shown further down this page into your prompt, and say that all saved data goes through relative /.h0/data/ URLs, not localStorage.
• Ask for a zip, with index.html and h0.json both at the top. A single HTML file can't carry the manifest.
• Don't judge it by ChatGPT's preview. Those URLs only exist once the app is on host0, so try it after the drop.

We haven't run this with a file from ChatGPT, though a drop doesn't care who wrote it. If the h0.json is wrong, the publish is refused with a message you can paste back into the chat. An app that needs server code or file uploads is a job for Sites, or another host.

It manages apps already on host0, and it never deploys. OpenAI gives full MCP support to Business, Enterprise and Edu, and read and fetch only to Pro. Its article doesn't mention Plus, Free or Go, so check your own Plugins page.

• Connect it on ChatGPT web. OpenAI says connecting a custom server isn't available on mobile.
• Select host0 with @ for each message that needs it. The selection covers that message, not the whole conversation.
• Expect a confirmation before anything changes, depending on your app permissions and the action.
• In a workspace, ChatGPT keeps a frozen snapshot of the tools from when an admin approved them. When host0 adds or changes a tool, an admin has to refresh it, and calls can fail until then.

These are OpenAI's documented steps and they should work, but we haven't run a full ChatGPT session ourselves.

Drop the new file on the app's Versions tab in your dashboard, not on the tool page. The tool page makes a new app with a new link, while the Versions tab publishes the next version at the same URL with the same access and shared data.

It shows what will change against the live version and waits for your click. A drop with no h0.json keeps your collections and their data. The connector can't upload the new file, but on plans where it can make changes, it can put an older version back.