HTML Studio sits at /workbench/text/html, and it's the corner of x-rush I keep coming back to. If you landed here through an older link or bookmark: this is the HTML Studio guide, same tool and same address under the new name. It's an online studio for building web pages, H5 games, and interactive pieces with AI — start from a template, or describe what you want and let the agent write the first version. Both routes drop you into the same editor. I've used it for a client landing page, a friend's portfolio, and a pixel-art game I never finished. The workflow holds up, and the payoff is the part most people walk past: hit Publish and the page goes live on its own public URL that anyone can open with no login, ready to share or embed.
This is the 2026 version of the doc. The template library, the preview rule, the fork-or-generate decision, the editor's four zones, and how the Composer turns a request into a confirmed change are all below. Then the part I'd put on a billboard: what Publish actually hands you — one stable public page per workspace, a version history that keeps everything you've shipped, and no separate results page to manage.
The Template Library
Open the Templates view and you get a card wall organized into 11 categories, with search plus dropdown filters for category and tags. Clicking a card only previews it. Nothing touches your code until you explicitly apply, and if the workspace you're sitting in has unsaved edits, you'll be asked to save, discard, or cancel first.
Using a template is free and takes two clicks: choose the Project it should live in, confirm, and you get an independent HTML Workspace with the template's code, required files, and starter notes copied over in one atomic step. No AI runs, no credits move, and a failed copy leaves nothing half-built behind. The copy is a one-time starting point — if the template improves next month, your workspace stays exactly as you built it. This is the route I take for standard client landings, and it's the fast path most of the time.
Preview Before You Commit
Preview has one rule worth internalizing: previewing never changes code. From the template wall, a card opens a preview you can back out of with no side effects. Inside the editor, the live preview at the bottom renders your current draft as you work, with a device-width menu for desktop, tablet, and phone checks. Escape, a second click, or a click outside closes it.
I check the phone width more than I expected to. Full-size layouts lie at small widths in ways the code never warns you about, and the device menu is the cheap way to catch that before a visitor does.
Fork a Template or Generate from Scratch
Two ways in. The fork route is the one above: preview, apply, done. The generate route starts from a form with three fields. A name, required, and the agent can draft one from your description if you'd rather not invent it. Tags, optional, for filtering inside your Project. Then a plain-language description of what you want. That's the entire form.
The old type, style, and feature menus are gone, deliberately. A fixed list could never cover "H5 game with a leaderboard", so those requirements now live in the description where they belong: write "a dark, glassy landing page for a habit tracker, with a contact form" and every constraint sits inside the sentence. Staring at an empty field? Two helpers will suggest ideas or tighten your wording, and both refill the same field.
Then one button: Execute creation · 0.5 credits. The backend checks your balance before anything else happens. If the task isn't accepted, you keep the form exactly as typed — no workspace created, no credits charged. If it is, the workspace opens while code streams in; the file tree, editor, and preview fill in as it goes, so you're never staring at a spinner. Generation and every later edit run on x-rush's smart routing, which picks from the top-tier text models available for the task and swaps in new models as they ship.
Honest take: for a standard SaaS landing page the library still gets you further, faster. Generate when the library doesn't have what you need.
Inside the Editor
On desktop and tablet-landscape the editor runs four zones, each scrolling inside its own area:
- Left, file tree. New files, rename, delete, download, and uploads. Assets come in through the tree, not the AI panel. Uploads are free; a checkbox, off by default, decides whether a file also lands in your account media library (500 MB per account).
- Center, Monaco. The editor VS Code uses, so muscle memory transfers: multi-cursor, find-and-replace, the works.
- Right, the Copilot panel. One input box, one Send button. Covered next.
- Bottom, live preview. Your draft, rendered as you work, with fullscreen, popout, and the device-width menu.
On a phone the whole thing collapses into a full-screen chat, with Preview and Publish kept at the top. That's a different shape, not a shrunken editor — file surgery belongs on a bigger screen.
The AI Assistant
The Copilot is one input box, and the distinction that matters is question versus change. Ask how to make the hero section sticky and you get an answer. Ask it to make the hero section sticky and you get a proposal: the specific edits it would make, against the workspace you actually have. Nothing is written yet, because sending never executes code. A stray Enter can't cost you a file.
Code moves only when you confirm, through the button that appears with the proposal: Execute change · 0.5 credits. It's the only paid action in the workbench that writes code, and it prices first generation and edit number forty the same. Editing your request invalidates the old proposal, so a stale instruction can't fire later. On desktop you can attach the current file's selection as the focus for a change; the selection steers the edit without fencing it in, so a fix that legitimately spans three files can touch three files.
Both answers and proposals draw from the shared assist pool: every 10 successful assists across x-rush, AI Chat included, cost 0.5 credits, and the count survives switching projects, sessions, or devices. You can't reset it with a new tab. I've tried.
Undo is scoped. Before every paid execution the system keeps a one-time restore point covering both the code and the private running notes about your piece; one Undo rewinds both, the restore point never appears in version history, and using it changes no billing facts. At first I expected undo to return me to some pristine starting state. It rewinds exactly one AI action instead, which turns out to be what you actually want.
One more constraint: a workspace runs one AI task at a time and keeps one conversation history across all your devices. The execute button refuses a second click while a task is running, so double-clicks and second tabs can't double-charge you.
Save, Publish, Version History
Saving is quiet. Each workspace holds one current draft, continuously auto-saved; manual edits and AI changes both update that draft and never create versions. Versions exist when you publish, and only then. Before your first publish, version history is empty.
Publish starts with a visibility choice: a public share link (not listed in the community) or community public, which adds the piece to Explore with your author card, likes, and Remix. Both routes pass AI-run safety checks before anything goes live; community pieces get a second pass against the content rules. What you get is one stable address on the x-rush.ai domain, /p/{id} — a real hosted page inside a security shell with a footer carrying the x-rush.ai mark and your public identity. Anyone with the link opens it fullscreen with no login. Send it in a chat, paste it into a post, drop it into an iframe; the person on the other end just sees a working page.
The address behavior is the feature I'd put on the billboard. Republish, and the same URL switches to the new version. Unpublish, and it stops serving but stays yours; restore later and it's the same link you already sent around. Every version you publish stays in the workspace's version history, and you can republish any old snapshot — the current draft stays untouched.
There's no separate results page, and that's not a gap. The Workspaces list, filterable by name, tag, and draft or published state, plus version history is the whole management surface. One identity per piece, not a pile of one-off generations.
If someone Remxes a community piece of yours, they get a one-time copy into their own Project: your public code, the files it needs, and (only if you switched on the public-final-prompt toggle) a copy of the prompt you chose to share. No lineage is recorded on either side, and their copy doesn't follow your updates or your deletion. The copy itself is free. Your private notes and chat never leave your workspace.
Pro Tips and Sharp Edges
- Sending is free; executing is 0.5 credits, flat. Proposals cost nothing beyond the shared assist pool, and every confirmed change is the same fixed price.
- Undo rewinds one AI action, code and notes together. Use it without ceremony; it never touches version history or billing.
- Republish freely. Same URL, new version, old versions kept and re-publishable.
- Applying a template over unsaved work asks first. Save, discard, or cancel — read it before you click.
- Remix is a clean copy. Nothing tracks back to you or the original afterward.
- Phones get chat, Preview, and Publish. Do file work on desktop or tablet landscape.
- Off-site links pass a stop first. Published pages route outbound clicks through a "you're leaving x-rush" screen that names the destination. Deliberate, and it protects whoever clicks.
That's HTML Studio. Templates get you shipping in minutes, and the Copilot scales from one-line fixes to real client work. Publish is the part that sticks: every piece ends up with a stable address, and there's no results pile to clean up afterward.