Skip to the article
3dCMS Knowledge Base
Contact Open the editor
Build with AI: Claude, ChatGPT and MCP

Your first 3D website from a chat: a 10-minute walkthrough

From an empty account to a published 3D website by chatting with Claude or ChatGPT: four phrases, the tools, the design review, and publishing.

Open in the interactive manual →

The interactive version has full-text search, a table of contents, and light/dark themes.


This is the whole loop, start to finish, in any connected client: Claude, ChatGPT or one of the others. You need a connected client and nothing else; a new 3dCMS account is created the first time you sign in.

The four phrases

SayWhat happens
"list my scenes"list_scenes shows what you have
"make a scene called X with ..."create_scene, then apply_ops builds it
"in <scene>, add / move / recolor ..."get_scene reads it, then apply_ops edits it
"publish <scene>"publish_scene, after it asks you first

The walkthrough

  1. Check the connection (1 minute)

    Ask "list my scenes". An empty list on a new account is a pass.

  2. Say what the site is for (1 minute)

    Give it a name and a purpose, not just objects: "Make a scene called Studio: a portfolio for a ceramics artist. Calm and natural. A home page with my name and one line about my work, a gallery page, a contact page, and signs linking all three. The point is getting people to email me." The assistant records that as the scene's brief (intent, mood, the one action the page is for) and builds to it.

  3. Let it build (3 minutes)

    It calls create_scene, then one or more apply_ops batches: pages, text, a floor, lights, the arrival camera and the signs between pages. Each batch is atomic and checkpointed, so you can roll any of it back in the editor under File ▸ Version history.

  4. Let it look (2 minutes)

    It renders the scene with get_scene_screenshot and checks it, then runs design_review against the brief and probe_framing on the arrival camera, and fixes what those name first: unreadable text, a sign out of view, a camera that crops the page. If it skips this, ask: "Review the design and fix what matters most."

  5. Change something (1 minute)

    "In Studio, make the home heading terracotta and move the gallery page two metres to the left."

  6. Open it yourself (1 minute)

    Every result carries an editorUrl. Open it in your browser while signed in and the scene is there in the editor, to walk through and edit by hand. The first time you open a scene your assistant built and have not published yet, the editor greets it with one card instead of the usual first-run tour: This is the scene your AI built. It is private until you publish it, with a Publish button that opens the publish dialog and a Look around first button that dismisses it. The card shows once per scene per browser; the guided tour stays under Help ▸ Interactive tutorial. Anything you save there, the assistant sees next time.

  7. Publish (1 minute)

    Until you publish, the scene is private: only you and your assistant can see it, and the assistant says so after the first build. Say "publish Studio". The first call does not publish: the assistant gets a notice of what publishing means (a public address, submitted to search engines, listed on Explore) and has to ask you. Say yes, and it comes back with the site URL and a URL per page.

What each tool does, in one line

ToolIn one line
list_scenesLists the 3D scenes saved in your account.
get_sceneReads one scene: its objects, positions and materials.
create_sceneMakes a new scene: empty, or a copy of one of yours (copyFrom, a Save As to experiment on while the original and its published site stay as they are).
apply_opsEdits a scene: add, move, restyle, delete and link objects, pages and lights.
list_assetsSearches the images, 3D models and sounds in your library and the shared public library.
upload_assetPuts a file (a model, an image, a sound) into your library, or copies in a shared one.
compose_surfaceBuilds a printed surface: a canvas, poster, sign or label.
measureMeasures sizes, gaps, clearances and the opening of a frame or doorway.
get_scene_screenshotRenders the scene on the server and returns the image.
probe_lightingNames everything lighting one point.
probe_framingJudges a camera as a photograph, at desktop and phone shapes.
design_reviewA designer's critique against the brief, with the five fixes that matter most.
publish_scenePublishes to a public, crawlable URL, after asking you.

The details of each are in Tools & quick phrases.

What the published site gives you

The URL opens the 3D site for visitors, and the same URL serves a full HTML version of every page for search engines and AI assistants, with a plain-text /text/ twin, a sitemap and an llms.txt. Nothing extra to set up: see AIO: how AI assistants read a 3dCMS site. Want the words to rank? Ask for real copy on each page, not just headings: a crawler reads only what the pages say.

Prompts that work well Say who the site is for and the one thing a visitor should do. Name pages, not coordinates. Ask for one change at a time once it is built. And ask it to look ("screenshot it and review the design") before you publish.
Spotted something off, or want a topic covered? Tell us, a person reads every message.