The scene creation process
What happens when you build a scene yourself: uploading and sizing your model, reviewing what the AI understood, adding tours and states, and publishing it to a live URL and an MCP endpoint.
When you build a scene yourself, the work happens in two phases. First you upload and size your model — a short sequence that ends with you approving the cost. Then you walk seven steps in the studio, reviewing what the AI understood and adding the interactive parts, before publishing to a live URL and an MCP endpoint.
Most of the time is spent in the middle steps, and most of the quality comes from them too. The upload is mechanical; the review is where a scene becomes genuinely useful.
Two phases: a short upload that ends with you approving the cost, then seven steps in the studio.
Phase one — uploading your model
What you can upload
Drop in a GLB, GLTF, OBJ, or FBX file, up to 500 MB. Anything that isn’t already GLB is converted for you, and GLB files are re-exported to clean them up, so you don’t need to convert anything yourself.
You see the cost before you commit
Your model is analysed in the browser the moment you drop it — nothing is sent anywhere yet. That analysis counts the parts, identifies duplicate bodies, judges how meaningful the existing part names are, and estimates how many groups the assembly will produce. From that you get a price, before you’ve committed to anything.
Nothing exists on our side until you press Proceed. Up to that point the file has not left your machine.
Setting the real-world size
A 3D file has no idea how big the real object is — a housing and a hydraulic press look identical to it. So you’re asked to anchor the model to reality: pick a part in the viewer, pick an axis, and type the true measurement in millimetres.
This is required, and it is the single highest-leverage thing you do in the whole upload. Every dimension the finished scene quotes, every feature size, and every spatial judgement it makes depends on it. Get it right and the scene can tell a technician a bolt is 12 mm; get it wrong and every measurement it ever gives is wrong by the same factor.
The sizing anchor ties the model to millimetres. Every measurement the finished scene quotes depends on it.
One part, one axis — the quickest option. Measure something you know.
Between two parts — the distance between two components along an axis.
All three axes — the full bounding size of one part.
Marking the front
Next you click which face of the model is the front of the product. Only the four side faces are offered — top and bottom are never a front.
This is what lets the scene talk about your product the way a person would. Without it there is no meaning to “the panel on the right”, “round the back”, or “the front intake” — the AI would be describing your product from an arbitrary direction that matches nobody’s mental model.
Marking the front face is what lets the scene say “the panel on the right” and mean it.
Naming the scene
You give the scene a name and a one-sentence description of what the product is. These are not labels for your own filing — they are sent with the model and used to ground the AI’s first pass over it. A description of “industrial coffee grinder, commercial café use” produces meaningfully better part naming than “Scene 4”.
Choosing a tier
Two tiers are shown side by side with their exact prices before you commit. Cost is built the same way in both: a base fee, plus an amount per unique body, plus an amount per group — so a bigger assembly costs more, and you can see the arithmetic.
Basic — AI mesh naming, body grouping, and format optimisation.
Advanced — everything in Basic, plus AI-written descriptions, smart tours, and priority processing. This is the default.
Basic is unavailable when your model’s part names carry no meaning. If everything is called Body1 through Body400, there is nothing for the cheaper path to work from — Basic reads your existing names, and Advanced works out what parts are from the geometry itself.
Pressing Proceed uploads the model, which usually takes twenty to thirty seconds. While that runs, a snapshot of your model is analysed to suggest a name, category and cover image — anything suggested this way is badged so you can tell what was guessed, and the badge disappears as soon as you edit the field.
Uploading takes twenty to thirty seconds, while a snapshot is analysed to suggest a name, category and cover image.
Phase two — the seven steps
1. Basic info
Confirm the name and description, choose a product category and target audience, and set the cover image. Your uploaded model is shown here locked — swapping the model means starting a new scene, because everything downstream is built on it.
This step also takes supporting material, and it’s worth spending time on: manuals, spec sheets and service documents (PDF, Word, text, CSV, Markdown), reference URLs, and free-form notes. Anything you add here becomes searchable knowledge the finished scene can answer from, so a scene with the service manual attached answers questions a scene without it simply cannot.
2. Documentation — checking what the AI understood
This is the most important step in the process. The AI has divided your assembly into component groups and named them; here you check that work and correct it.
Rename any group the AI got wrong or named unhelpfully.
Merge groups that should have been one thing.
Create a group yourself by selecting parts in the viewer.
See where the AI was unsure and what alternatives it considered.
Ask for a group to be reanalysed with a hint when it has misread something.
Mark points of interest and camera positions — the tours step later builds on these.
Time spent here pays back everywhere else. Every answer the finished scene gives is phrased in terms of these groups and names, so a mislabelled assembly produces confidently wrong answers forever. If you do nothing else carefully, do this.
Documentation — check the AI's grouping and naming. Every answer the finished scene gives is phrased in these terms.
3. Enrichment
The AI writes descriptions for the groups you just confirmed, and can search for replacement parts across the assembly. Parts search runs in the background and takes longer than the rest — you can carry on while it works.
4. Interactive experiences
Three kinds of interactivity, all optional, all authored against the same model:
Guided tours — an ordered sequence of steps, each pointing at a part or a point of interest, with its own camera position, timing, and optional spoken narration. Best for procedures a person should be walked through.
Animations — movement. Demonstrate how something opens, disassembles, or operates, either by posing it yourself or by describing what you want and letting the AI build it.
States — a named, described configuration of the product: “fully disassembled”, “top flap open”, “service position”. You either mark a moment on an animation or pose the model directly and save it.
States are what let someone say “show it disassembled” and have it happen, or “show me how to put it back together” and watch it run in reverse. They are worth the effort on anything that comes apart.
States let someone ask to see the product disassembled, or watch the reassembly run in reverse.
Give animations and states real names and real descriptions. The description is how the AI finds them later — it is what gets matched when someone asks a question, not a label for your own reference. “Open the service hatch to reach the filter” is findable; “Animation 3” is not.
5. Customization
Lighting and materials — ambient and directional lights, shadows, bloom and ambient occlusion, and per-part material editing with presets and bulk edits. This is presentation only; nothing here changes what the scene knows. Available once processing has finished.
6. Review — including who can see it
A readiness summary showing what’s present and what’s missing: a schema completeness score, counts of components, tours, states and points of interest, and any problems that would block publishing.
This step is also where you decide access, with two linked settings:
Unauthenticated access — anyone with the direct link can open the scene without signing in.
Embedding — the scene can be placed in an iframe on your own or a partner’s site.
Embedding depends on unauthenticated access and cannot be switched on without it. Turning unauthenticated access off turns embedding off at the same time — there is no way to end up with a scene that is embeddable but private, which is deliberate.
Review — embedding depends on unauthenticated access, so a scene can never be embeddable and private at the same time.
7. Publish
Publishing validates the scene, builds thumbnails, optimises assets, prepares documentation and generates your URLs. You come out with an authenticated scene URL, a public URL and embed code if you enabled them, and ready-to-paste iframe markup.
Publishing produces a live scene URL, an optional public URL and embed code, and — from the published scene — an MCP endpoint.
After you publish — connecting it to AI
MCP is not set up during scene creation. Once a scene is published, open it from your scenes list and you’ll find its connector there, with two forms:
A share endpoint — for customers, technicians and colleagues. Browsing, tours, and questions.
An edit endpoint — adds tools for changing and publishing the scene from inside the assistant.
Copy the one you want and follow the install guide for Claude or ChatGPT in this section.
Workspace-wide MCP settings live separately, under your workspace configuration: which tool groups are exposed, and API keys for machine-to-machine use that can be scoped to specific groups. Visual styling of the MCP app — your colours, brand and typography — is set in your workspace customization settings.
Getting a good result
Measure something you’re certain of. The sizing anchor propagates into every measurement the scene will ever quote.
Write the description before you upload. It grounds the AI’s first pass, so a vague one costs you quality in every later step.
Attach the manual. Documents added in step one become answerable knowledge.
Fix the groups. Ten minutes in the documentation step is worth more than anything you can do afterwards.
Describe, don’t label. Animations and states are retrieved by their descriptions.