3D AR Studio MCP Server

3D AR Studio

Drop a full augmented-reality studio into any web page.

Place as many 3D models as you like in your real room through the camera, describe a new one and watch it appear, arrange everything by hand, then share the whole scene as a link, a QR code, or a live room someone else can build in with you.

Live demo · npm · MCP server

<script type="module" src="https://unpkg.com/3d-ar-studio/dist/ar-studio.min.js"></script>
<ar-studio></ar-studio>

That is a working AR studio. No build step, no API key, no account. It comes wired to a free library of a few hundred public-domain models and a free, keyless text-to-3D lane; point it at your own catalogue with one option when you are ready.

---

Why this exists

Every web-AR drop-in places exactly one model and then hands off to a native viewer, which ends the session. This one keeps the whole scene in your page:

• Many models, one room. Place, drag, pinch-resize, twist-rotate and duplicate as many models as you want in a single live camera view.
• Generate without leaving the camera. Type "a brass desk lamp" into the dock. The generation runs behind the live view and the finished model drops into the room.
• Real WebXR where it exists. An always-armed hit-test reticle, one XRAnchor per placed model, real-world light estimation, and depth occlusion so models hide behind your furniture.
• Real ARKit and ARCore everywhere else. iPhones have no WebXR, so an iPhone gets Apple's AR Quick Look instead: true plane detection, true scale, true occlusion. The GLB is converted to USDZ on the device (a real conversion via three.js's USDZExporter, about a second for a typical prop, no server involved). Android without WebXR gets Scene Viewer. Desktop gets a grid preview and a QR hand-off to a phone.
• Scenes are links. Models, positions, rotations and scales round-trip through the URL. Compose on a laptop, scan the QR, it reopens exactly on your phone.
• Build together, live. Open a room, share a six-character code, and every add and move syncs to everyone in it.
• Characters actually move. Any humanoid GLB with no baked animation gets an idle clip retargeted onto its own skeleton. No rig allow-list, no T-poses.
• Agents can drive it. A bundled MCP server lets Claude, ChatGPT or your own agent generate a model, compose an arrangement, and hand a person one link that opens it in their room.

The rendering ladder, anchor lifecycle, retargeting pipeline, scene format and shared-room protocol are extracted from the AR surfaces running in production on three.ws, and generalized so they work on your site with your models.

---

Install

npm i 3d-ar-studio three

three is a peer dependency, so you keep one copy of it and pick the version. The CDN bundle (dist/ar-studio.min.js) has three.js inside it and needs nothing else.

import { createArStudio } from '3d-ar-studio'

const studio = createArStudio('#stage', {
  branding: { title: 'Acme AR', accent: '#00b894' },
})

studio.on('add', ({ placement }) => console.log('placed', placement.title))

The studio fills its host element absolutely, so give the host a height (any positioned box with a real height works; a <div> with no height gets a sensible 70vh default rather than rendering invisibly).

Scaffold a deployable page

npx 3d-ar-studio create my-ar-site     # a folder you can publish as-is
cd my-ar-site
npx 3d-ar-studio dev                   # look at it locally
npx 3d-ar-studio deploy                # push it and turn on GitHub Pages

deploy prints every git and gh command before it runs it. It needs git and the GitHub CLI; without them it tells you the three manual steps instead of failing silently.

Templates: static (one HTML file, no build), vite, react.

Both 3d-ar-studio and ar-studio run the CLI. The MCP server is a separate binary in its own package, so npx 3d-ar-studio-mcp resolves cleanly: see MCP server.

---

Your own models

The tray is filled from three.ws by default: a few hundred public-domain (CC0) props, free for commercial use, served with open CORS. Swap in your own with the assets option.

A JSON file anywhere. Five common shapes are read without reshaping:

createArStudio(el, { assets: 'https://cdn.acme.com/models.json' })

// Any of these work:
[ { "url": "https://cdn.acme.com/chair.glb", "name": "Aero chair" } ]
{ "items":     [ … ] }
{ "objects":   [ … ] }   // three.ws object library
{ "creations": [ … ] }   // three.ws forge gallery
{ "models":    [ … ] }

Per entry, the model URL is read from the first present of src, url, glb, glb_url, glbUrl, file or model; the label from title, label, name or prompt; and the thumbnail from poster, thumb, thumbnail, image or preview_image_url. Anything that is not an https (or site-relative) URL is dropped rather than handed to the loader.

A list you hold in code:

import { staticSource } from '3d-ar-studio/sources'

createArStudio(el, {
  assets: staticSource({
    label: 'Our furniture',
    items: [{ src: 'https://cdn.acme.com/chair.glb', title: 'Aero chair', poster: '…' }],
  }),
})

Several tabs at once, in the order you want them:

createArStudio(el, { assets: ['recent', myCatalogue, 'objects', 'link'] })

Built-in keys: 'three.ws' (the default set), 'recent', 'objects', 'community', 'link'.

Anything else. A source is an object with a list():

createArStudio(el, {
  assets: {
    id: 'search',
    label: 'Search',
    searchable: true,
    async list() {
      const rows = await fetch('/api/models').then((r) => r.json())
      return rows.map((m) => ({ src: m.glb, title: m.name, poster: m.thumb }))
    },
  },
})

Throwing from list() is fine: the tray renders a designed error state with a Retry button.

Your users can retarget it too, without touching your code: ?assets=https://…/manifest.json on the page URL. Only https URLs are accepted, and every model source is re-validated before it reaches the loader, so a hostile link can add a catalogue but can never smuggle a javascript: or data: model into the scene. Set allowUrlOverride: false to switch that off.

CORS

Models are loaded by the browser, so the host serving your .glb files has to allow cross-origin requests (access-control-allow-origin). If a model fails to load, that is almost always why, and the studio says so in the status line rather than failing silently.

---

Options

URL parameters

Methods

await studio.addModel({ src, title })        // place a model
studio.clear()                               // remove everything; returns what was there
studio.getScene()                            // [{ src, title, x, z, yaw, scale }]
await studio.setScene(items)                 // replace the arrangement
studio.shareUrl()                            // a link that reopens it exactly
await studio.generate('a brass desk lamp')   // text to 3D, into the room
studio.viewInYourSpace(src, title)           // open the hosted launch page for one model
await studio.startCamera()                   // needs a user gesture on iOS
await studio.enterAR()                       // best AR path for this device
await studio.placeInYourSpace()              // native viewer (Quick Look / Scene Viewer)
await studio.toggleImmersive()               // enter or leave WebXR specifically
await studio.openRoom()                      // returns the room code
studio.destroy()                             // releases camera, socket and GPU context

Events

studio.on(name, fn) returns an unsubscribe function. The same events also fire as ar-studio:<name> DOM events on the mounted element.

---

Web component

<ar-studio
  assets="https://cdn.acme.com/models.json"
  title="Acme AR"
  accent="#00b894"
  generate="true"
  rooms="true"
></ar-studio>

element.studio is the live instance. Importing 3d-ar-studio/auto (what the CDN bundle does) registers the element for you.

---

Keyboard and accessibility

Every control is a real button with an accessible name, the source tabs implement the full ARIA tablist contract, and each dialog takes and returns focus.

prefers-reduced-motion removes the spawn-in animation and every transition.

---

MCP server

npx 3d-ar-studio-mcp

{
  "mcpServers": {
    "3d-ar-studio": {
      "command": "npx",
      "args": ["-y", "3d-ar-studio-mcp"]
    }
  }
}

No API key. Every tool is free and keyless.

A session looks like this:

> Put a mid-century lamp and a potted fern in my living room.

  generate_3d_model  { prompt: "a brass mid-century desk lamp" }   → lamp.glb
  search_models      { query: "potted plant" }                     → fern.glb
  compose_ar_scene   { models: [{ src: lamp.glb, x: 0,   z: -1.4 },
                                { src: fern.glb, x: 0.9, z: -1.2 }] }

  → one link; open it on a phone and both objects stand in the room.

---

Device support

The AR button in the top bar always takes the best path the device has, and labels itself so it never promises the wrong one. On a phone without WebXR, each selected model also gets a Place in your space button that opens the native viewer directly.

Camera and WebXR both need a secure context: https:// or localhost.

---

Development

npm install
npm test                 # 55 unit tests, no browser needed
npm run build            # dist/ bundles
npm run build:site       # docs/ (the GitHub Pages site)
npm run test:browser     # 18 end-to-end checks in a real browser (needs Playwright)
npm run inspect          # MCP Inspector against the local server

The published site is committed under docs/ and served by GitHub Pages from the main branch. There is no CI workflow: npm run build && npm run build:site, commit, push.

---

Licence and credits

Apache-2.0.

The bundle includes three.js (MIT) and colyseus.js (MIT). The default model library is CC0 content from Poly Haven, and the default generation and animation lanes are hosted by three.ws. None of them is required: every one is a URL you can change.