August 2026

August 19, 2026

Assigned materials no longer disappear when a project is edited further

Improved

• The project health check now reports a clear error when a variant's material data has been lost, naming the material that was assigned so you know exactly what to re-assign.

Fixed

• Reloading a project in the editor no longer strips the shared material data that assigned materials rely on — the state that showed placeholder swatches and a white model can no longer occur.
• A save made from a broken reload can no longer erase material assignments: the server now keeps any material data the save lost but the stored project still has.
• Re-assigning a material through the AI Assistant on a variant that had been saved before no longer risks silently reverting to the previous material on the next reload.

August 19, 2026

Pick 3D parts from a list instead of typing their names

Improved

• The parts the AI already thinks are right come pre-ticked, so the usual answer is a single confirming click.
• Long part lists get a search box, plus "select all" and "clear" that respect what you have searched for.

Fixed

• The part question now always shows a selectable list, however many parts the model has — the ten-part cut-off that pushed you into typing names is gone.
• You can select several parts. Previously the list let you choose exactly one part, or all of them, so a material meant for three of seven parts could not be described at all.
• A mistyped part name no longer ends the build. The run continues using the suggested parts and tells you to check the targets in the editor, instead of failing and losing the work done so far.

August 19, 2026

The AI Assistant can now actually add content from the Simplio3D library

Improved

• The assistant can search the curated library itself by keyword, so it can find a starting model or a realistic finish before suggesting you upload your own.
• Copying the same curated item twice reuses the copy you already have instead of creating a duplicate.

Fixed

• Asking the assistant for a curated model or material now really adds it, instead of occasionally showing a green "Done!" card that changed nothing.
• The assistant knows a curated item has to be copied into your own library first, and that the copy is what gets used — so it no longer tries to apply a library item directly and fail.
• Before copying several materials at once, the assistant now tells you that a curated import cannot be undone from the chat: to remove one, delete it in Dashboard → Assets or Dashboard → Materials.

August 19, 2026

AI changes can no longer be lost to an editor save

Fixed

• After the AI applies a change, the editor now always refreshes from the live project — an applied change appears in the panels immediately, instead of sometimes showing the state from before it.
• A save from an editor that loaded before another change landed is refused instead of applied, with a clear message ("This project was updated outside this editor") and an automatic reload. The newer change survives; before, it was silently erased.
• Keeping the editor open while a connected AI app or the developer API changes the project is now safe — the editor notices and reloads rather than overwriting the app's work on its next automatic save.
• The assistant's old one-click "Add … Block" buttons no longer appear when the AI's real project tools are available. Those buttons created content in the editor's memory only, could not carry values (a requested base price arrived as 0), and could conflict with the changes the AI was applying properly. Content changes now go through the AI's validated, undoable project tools; navigation buttons remain.

August 19, 2026

The AI Assistant no longer replies with nothing

New

• Every reply now contains a written answer. If the assistant runs out of room part-way through, it stops and writes up what it found or changed, what is still outstanding, and the next step — instead of returning an empty message.
• A Continue button appears under a reply that stopped early. One tap and the assistant picks up from where it left off; nothing is lost, and anything already applied to your project stays applied.

Improved

• When the assistant asks which parts to change and you answer, that answer now gets a full working allowance — acting on it is the real work, and it used to be treated as a trivial follow-up and run out of room.
• Replies lead with the outcome — Done, Partly done or Blocked — and anything that could not be done is stated at the top with the reason, rather than left for you to spot further down.
• The assistant acts on your answer straight away instead of confirming it back to you, and no longer says it is waiting for something you have already given it.
• When a change the assistant makes reports a problem — a material option with no material attached yet, which would show as a plain placeholder — it now resolves that in the same reply or tells you it is outstanding.

August 19, 2026

Run AI workflows from Claude, ChatGPT, Cursor or VS Code — and pause them

New

• Connect Claude, ChatGPT, Cursor or VS Code under Dashboard → Integrations → AI Connections and let it run workflows for you. The work happens on our side, so a run survives the chat being closed, the app disconnecting, or the conversation being restarted — pick it up later from your dashboard.
• A connected app can never invent an answer to a workflow question: an answer that is not one of the offered choices is rejected. And no AI, in any app, can approve a pricing, SKU or deletion change — it can only show you where to approve it.
• Workflows can now be paused and resumed. A paused run stays exactly where it is until you press Resume, however long you leave it.

Improved

• Only one workflow at a time may change a given project. Starting a second one is refused with the name of the run already working on it, so two jobs can no longer edit the same pricing at once. Read-only runs such as the audit are never blocked.
• Every workflow now reports a clear summary of what it found, proposed, changed and left outstanding — the same information whether you are reading it in the assistant, on the AI Workflows page, or through a connected app.
• Existing AI connections keep exactly the access you granted them: running workflows is a new permission, so an older connection can watch runs but not start them until you re-authorize it.
• The AI Assistant window now uses the width of your screen instead of staying in a narrow column, so long answers, applied-change cards and your own messages are no longer cut off on the right.
• The assistant identifies itself simply as AI rather than naming the model behind it. Which model answers is ours to manage — and can differ between one answer and the next — so it is no longer surfaced as if it were a setting of your project.
• Generate with AI is clearer that it hands you a first draft rather than a finished configurator: it says so before you build, on the button, and again on the result — which now lists what to check first and offers both ways to carry on, by hand in the editor or with the AI Assistant inside the project. The button that opens it reads "Review draft in editor".

August 19, 2026

Price shown vs price charged, and settings that quietly did nothing

New

• Pricing blocks your formula never references are now flagged, because they add nothing to the displayed price but are still charged at checkout. Custom sliders like Width get their own warning: unreferenced, their raw value was being charged as money.
• Project settings with a fixed set of values are validated when written. The clearest case: tax mode accepts "inclusive", and the near-miss "included" was accepted and then treated as exclusive, adding tax on top of a price that already included it.
• Add-to-Cart buttons created by the AI now point at whichever store you actually have connected. Previously they always defaulted to WooCommerce, so on a Shopify-only store the button could never reach checkout — and because it replaces the Request-a-Quote button, the form was left with no working action at all.

Improved

• A "Save Configuration" field can no longer be marked required — doing so made the quote form impossible to submit, with no visible error.
• A price table using a checkbox for one of its axes is reported: checkbox selections cannot be read by the price-table lookup, so every price in that table was ignored.
• Switching off the options sidebar on a project that still has options is now reported as an error — it leaves shoppers with no way to configure anything.
• Rules that can never fire, and modular rules aimed at something they cannot affect, are now surfaced instead of sitting silently in the project.
• Scenery camera settings left inactive because no scenery models are selected are now flagged when a change causes it — including when it is the model list that was cleared.
• Declining to link a store product during an AI project build no longer adds an Add-to-Cart button that cannot check out.

August 19, 2026

The AI stops creating options that look right but do nothing

New

• Options that replace each other (door counts, body styles, with/without a shelf) are now built correctly: you tell the AI which parts each choice represents, and it works out the show/hide rules itself.
• Project Health and validation now report configuration that cannot do anything — an option hiding the geometry it represents, options that never hide each other so nothing disappears, or a "show" applied to parts that were never hidden.
• Number Input scaling gets the same checks: a scaling axis with no parts, parts with no axis, a part told to both stretch and stay fixed, or a range whose minimum equals its maximum. It also flags when a part will grow in both directions because the model has no custom pivot — the usual reason a width control resizes the wrong way.

Improved

• The AI inspects your actual 3D model before targeting parts, instead of guessing names from a pattern.
• A part name that does not exist anywhere in the project is now rejected outright rather than saved and silently ignored.
• Mixing whole-object and individual-part targets in one visibility rule is refused — previously it saved, then ignored the parts.
• When a request implies mapping options onto geometry, the AI now completes the mapping in the same reply instead of creating the option and describing what is left to do.

August 19, 2026

The AI now truly understands your 3D scene — and can move, resize and rename parts

New

• Ask about your scene and get real answers: the assistant reads the complete model hierarchy, part sizes, positions and spatial relationships ("the worktop is the large flat part at the top", "these two doors are mirror images"), instead of guessing from part names.
• It can search big imported models — asking for "door" finds Left_Door, Right_Door and Upper_Cabinet_Door with enough context to pick the right one, even in models with hundreds of parts.
• Move, rotate and scale by asking: single-object changes apply immediately and can be undone; changes that touch several parts at once ("make the cabinet wider but don't stretch the handles") are presented as a complete before/after plan you approve first.
• Rename badly-named parts safely. When the AI renames "Cube001" to "Kitchen_Left_Door", every option block target, conditional rule, animation and saved part adjustment that referenced the old name is migrated in the same change — nothing silently stops working.
• Before assigning option-block targets or building show/hide rules, the AI can now check what a part actually is — its position, size and surroundings — rather than trusting a name.

Improved

• When the AI proposes renames it lists every affected reference in the approval plan, including subtle knock-on renames among duplicate part names, so you can see exactly what will change before saying yes.
• Rotation values are always stated in both degrees and radians in AI answers, so there is never ambiguity about units.

Security

• Part renames only ever modify model files uploaded directly to the project being edited. Files that came from your Asset Library, a template or the curated library are shared across projects and are refused with an explanation — so a rename in one project can never break another.

August 14, 2026

A faster, cheaper AI Assistant — and a clear AI allowance on trials

New

• The assistant answers faster and picks tools more accurately, because it is now given only the tools relevant to what you asked rather than all of them every time. If it turns out to need something else mid-answer, it fetches it and carries on in the same conversation.
• Requests that need deeper reasoning are automatically escalated to a more capable AI on accounts with an active subscription — with no model picker and nothing for you to configure.
• Long conversations stay fast. Older turns are condensed while every project, block, material and variant already discussed is kept exactly, so the assistant does not forget what it has already done or re-propose work it has finished.
• Trial accounts now have a stated AI allowance, tracked separately for the AI Assistant and for Generate with AI — so using up one still leaves the other available.
• Generate with AI keeps working even when its AI allowance runs out: the build plan stays fully editable by hand and the Build button still produces the project, because the build itself never used AI.

Improved

• Reaching an AI allowance is now shown as a clear, plain-language message instead of an error — nothing is broken, and the message says exactly what still works.
• When a request is beyond what the AI available on your account can reliably complete, the assistant says so and suggests breaking it into smaller steps, rather than returning a half-answer.
• Answers are more concise by default. The assistant no longer narrates each internal lookup it made before getting to the point.

Fixed

• Fixed the assistant occasionally returning an empty reply on a complex request instead of either answering or explaining that it could not.

August 14, 2026

Generate with AI: a proper workspace, and a Build button that tells you what it needs

New

• The window is much wider on desktop and scales with your screen, so the conversation and the build plan both have room. On a phone or a narrow window they become tabs — Conversation and Build plan — and each gets the full screen.
• The Build button explains itself. When it is disabled, the exact reason is listed above it: no project name, a Viewer project with no 3D model, or nothing to build at all. The plan badge shows how many things need fixing.
• New amber warnings for builds that would surprise you — most importantly materials planned with no 3D model, where the swatches appear but clicking one does nothing. Warnings never block a build; they just tell you what will happen.
• A new "What will happen when I build" list shows all 13 build steps up front, with the ones that will be skipped struck through and the reason given.
• The plan is now editable: rename the project or an option block, add and remove finishes, set a base price and currency, add and remove form fields, and delete a rule — without asking the assistant. Renaming a block updates the rules that refer to it.
• The assistant can now look at your actual workspace while you talk — your 3D models and their real part names, your materials and categories, your projects and templates, your store and its products, and the official documentation. Replies say what it checked, so "the closest model in your library is X" is something it verified rather than guessed.
• It can only READ. It has no ability to create, change or delete anything: everything it decides goes into the build plan for you to review, and pressing Build my configurator is still the only thing that creates a project.
• Ask it a question and it answers from the real documentation, linking the page inside its answer, so it tells you plainly when something is not supported instead of inventing a setting.
• Every reply now says what it changed in the plan and what is most worth doing next, instead of repeating "Updated the build plan." on every turn — and it stays accurate even if the AI service is briefly unreachable, because the summary is written from the plan itself.
• Replies are formatted properly (headings, lists, bold, working links) and show a small summary of what changed in the plan on that turn.
• One-tap follow-up suggestions appear under its replies based on what your plan is still missing.

Improved

• Closing the window no longer throws away your draft — re-open it and the conversation and plan are still there, with a Start over link when you want a clean slate.
• Your edits to the plan are now sent to the assistant as the authoritative version each turn, so it no longer re-adds an option block you just deleted.
• Starter prompts are clearer and lay out two-up instead of wrapping into six-line blocks.
• The base price field can be typed into and cleared normally, instead of fighting you as you type.

Fixed

• Fixed the window rendering at a fixed narrow width on every desktop screen, which was squeezing the conversation into a column barely wide enough for a sentence.
• Fixed renaming an option block losing your cursor after every character.
• Fixed an option block deleted from the plan leaving behind a show/hide rule that pointed at it.
• Fixed an edit made to the plan while the assistant was replying being overwritten by the reply.
• Fixed replies that said only "Updated the build plan." or "Nothing changed in the plan." with no explanation. Replies now describe what changed; when nothing could be changed they say why and offer the next step.
• Fixed the assistant occasionally producing a reply the builder could not read at all, which showed up as an empty answer next to an unchanged plan. Your description is also now used to fill in the project name and type immediately, whatever the assistant returns.
• Fixed replies being generic because the assistant could not see your workspace — it can now look things up before answering.
• Fixed "I could not find a matching 3D model" going unsaid — it now tells you, and offers an Upload a 3D model button in the conversation instead of quietly attaching nothing.
• Removed the "From the docs" list under replies; documentation links now appear inside the answer, where the sentence refers to them, and are clickable.

August 13, 2026

The assistant can now finish the job — and answers from the real documentation

New

• Ask for the finished thing, not the skeleton: "make the swatches bigger and use two columns", "add a Width parameter from 300 to 1600 mm that scales the panel", "let customers put their name on the back panel", "make the lid swing open", "hide the armrest meshes when Frame = Basic".
• Modular projects: the assistant can add placeable modules, choose which faces they snap on, cap how many can be placed, and set rules like "no wall may attach to the roof's top face".
• Conditional rules can now target individual 3D parts, whole 3D objects and materials — not just blocks and variants. Before writing one, the assistant reads your actual model file and refuses a mesh name that does not exist, suggesting the closest real one.
• New page: Dashboard → AI Changes. Everything awaiting your approval in one place, with the full before-and-after list, plus a history of applied changes you can undo. Previously the only Apply button lived inside the chat, so a change proposed by a connected AI app — or in a chat you had closed — quietly expired after 15 minutes.
• Ask "how do I…" and get a real answer with a link. The assistant searches the official Simplio3D documentation and cites the page; if the documentation does not cover something, it says so instead of inventing a setting.
• It can now answer "how big is this account?" accurately (real project, material, asset and seat totals), read your team seats, list saved shopper configurations and check whether an email actually went out.
• Starting from scratch is easier: the assistant can browse and import from the curated Simplio3D content library and from free, licence-cleared texture sources, create material categories, and start a project from a template.
• Attach a 3D model, texture or graphic directly in the assistant conversation — it lands in your normal Assets library and the assistant can use it straight away.

Improved

• Price estimates the assistant gives you are closer to reality: per-unit Number Input pricing is now included, and options hidden by your conditional rules are no longer priced. The live Preview and Share views remain the final word, and the assistant still lists exactly what its estimate could not account for.
• Reading a 3D model with deeply nested groups now returns a complete, flat list of every part with its full path — previously very deep models were cut off partway through.
• A new read tells you a model's saved pivot before you configure a dimension parameter, so you can tell whether a size change will grow in both directions or only upward.
• Materials the assistant creates now get the same reflectivity values the editor applies when you pick a preset, so a "Metal" material actually looks like metal instead of only carrying the label.
• Texture maps can be removed from a material by the assistant, not just added.
• Imports are safe to repeat: importing the same library item twice reuses the existing copy instead of creating a duplicate.

Fixed

• A change proposed by a connected AI app used to be invisible and unapprovable — it expired after 15 minutes and you were never told. It now appears at Dashboard → AI Changes.
• Six kinds of change cannot be undone from the history — duplicating a project, the three library imports, creating a project from a template, and updating a quote request's status — because they create something new rather than editing something existing. Each now says so up front and tells you how to revert it.
• The assistant no longer relies on an internal copy of the documentation that could drift from the real site; a documentation change is picked up by the next release.

Security

• The assistant still cannot upload a file, publish a project, connect or disconnect a store, email your customers, or delete a project. Importing from the curated and free libraries is the only way it can bring new content in.
• It can route your project's email but can never see or write your SMTP username or password, and it is permanently blocked from writing custom scripts or project-wide custom CSS.
• Custom CSS on a block or form field is allowed, but it is checked and refused outright — never silently cleaned — if it contains anything that could load or run code in a shopper's browser.
• A preview thumbnail can only be set to an image already in your own Simplio3D storage, and clearing one never deletes the image file, because project copies and templates share it.
• Shopper personal details stay hidden: saved configurations report only whether an email or name was captured, never the values.
• Pricing, SKUs, deletions, bulk edits, custom CSS and store links still never apply on their own — they wait for you, and no AI tool exists that can approve one.

August 13, 2026

Generate with AI: describe your product, get a working configurator

New

• Projects → Generate with AI: a short brief — "a powder-coated steel shelving unit in three widths, with Oak, Walnut and Black finishes" — becomes a working configurator, complete with 3D models, materials, option blocks, conditional rules, a base price and a form.
• Nothing is built until you say so. The plan appears as an editable checklist beside the chat, so you can rename blocks, drop a finish, change the project type or remove a step before pressing Build.
• Missing an asset? Upload a 3D model, texture or graphic straight from the conversation — it lands in your normal Assets library and is used in the build.
• Wire up your store in the same flow: pick a Shopify or WooCommerce product and set what the storefront button does, without opening the Integrations page.
• Materials already in your library are reused when the name matches, so a brief mentioning "Oak" does not create a second Oak.
• Modular projects are built too, minus the parts that are editor-only — the builder tells you up front that module variants and snap sides still need configuring in the editor.

Improved

• Materials the assistant creates can now carry texture maps from your library — base colour, normal, roughness, metallic and ambient occlusion — instead of only a flat colour.
• The build runs as a live step checklist you can watch. It pauses to ask whenever it is not sure, and an unfinished build can be answered, approved or picked up later from Dashboard → AI Workflows.
• AI apps connected through Dashboard → Integrations → AI Connections can now see whether a store is connected and list its products, and propose a product link or a checkout-mode change for you to approve.

Fixed

• "Generate with AI" had been switched off since June because the previous generator produced inconsistent results. The flow has been rebuilt and is available again on the Projects page.
• A brief that names a 3D model you do not have now stops before anything is created and tells you which model to upload, instead of producing a project with missing models.
• An interrupted build is no longer a dead end: what has been built is a normal, editable project at every step, and the build resumes from Dashboard → AI Workflows.

Security

• Store credentials are never readable by the AI. It can see that Shopify or WooCommerce is connected and which checkout mode is in use — never a key, a secret or a store address — and connecting or disconnecting a store stays a dashboard-only action that no AI can perform.
• The base price, the product link and the checkout mode each pause for your explicit approval before anything is written, and the checkout-mode approval warns that the change applies to every configurator checking out through that store.
• Product links and checkout modes can only be changed by workspace owners and admins — an editor seat can look, but not change.
• The builder never publishes, never deletes, and never uploads anything on your behalf.

August 12, 2026

AI Configuration now defaults to a current Claude model

Improved

• If you already have a retired model id saved, clear the Model field and save: the platform fills in a current default for you.

Fixed

• Fixed: Admin → Settings → Integrations → AI Configuration failed "Test connection" with "Model not found: claude-sonnet-4-20250514". The platform default is now claude-sonnet-5.
• Fixed: with a current model selected, every AI request then failed with "`temperature` is deprecated for this model" — the platform was sending a setting that current Claude models no longer accept. It is no longer sent, and if a provider retires another setting the request now retries without it instead of failing outright.
• Fixed: the model advertised in the help text and the one the assistant actually fell back to were different — both are now the same value, and a test fails the build if they ever drift apart again.
• The Model field now lists the current Claude ids underneath it, and says that a "Model not found" result usually means the id has been retired rather than that your key is wrong — the provider's own message blames the key first, which sends people off to regenerate a working key.

August 12, 2026

Track AI workflows from the dashboard, and stop re-answering catalogue questions

New

• New Dashboard → AI Workflows page: filter by what needs you, what is still running, and what finished; answer questions, apply or decline changes, cancel, and pick up a workflow you started in an earlier session.
• Recurring price lists no longer ask the same questions twice — once you confirm which column is the customer-facing price, the next import of that same spreadsheet reuses your choice, and shows you that it did.
• If the spreadsheet shape changes (a column added or removed), the assistant asks again rather than risking the wrong column.
• Catalogue headers are now recognized in German, French, Spanish, Italian, Dutch, Portuguese and Nordic languages, including accented spellings and units like "Breite (mm)" or "Preis EUR". A header it does not recognize is still asked about rather than guessed.
• Building options from a catalogue can now apply the materials to your 3D model in the same run: the values become a Select Material block, with parts targeted and existing library materials reused.

August 12, 2026

Agent workflows: the AI Assistant now completes whole goals

New

• Ask "Check this configurator" for a full project audit with a 0–100 health score, every finding split into safe fixes and decisions that need you.
• Say "Fix the safe problems" and the assistant proposes one reviewable cleanup listing every dead reference — nothing is deleted until you press Apply.
• Build material options straight from your 3D model — "Make the upholstery configurable with Oak, Walnut and Black" — with the assistant asking which parts to target or whether to reuse an existing material whenever it is not sure.
• Import pricing from a CSV, Excel, or JSON catalogue: the assistant shows the detected columns, asks which column is the customer-facing price when several qualify, and stops on duplicate SKUs or missing prices instead of guessing.
• Turn plain language into conditional rules — "Only show the footrest when Frame B is selected" — with conflicts against existing rules flagged before anything is added.
• Ask "Is this ready to publish?" for a readiness report covering validation, pricing, and sharing — publishing itself always stays in the editor's Share dialog, never with the AI.

August 12, 2026

Connect Claude, Cursor and other AI apps to Simplio3D (MCP)

New

• The Simplio3D MCP endpoint is live at app.simplio3d.ai/mcp, with an OAuth 2.1 sign-in flow: add it as a custom connector in Claude and pick the workspace and scopes to grant on a consent screen — no token to copy.
• A new Dashboard → Integrations → AI Connections page: create, rotate and revoke connection tokens, see when each connection was last used, and review pending AI changes.
• All 54 AI tools are available to connected apps, filtered per connection by 13 scopes (for example read-only project inspection, or pricing proposals) — a connection only sees what you granted.
• Pricing, SKU and deletion changes always require your approval in the dashboard before anything is modified — an external AI app can never bypass that confirmation.

Security

• Connection tokens are stored hashed, and revoking a connection takes effect immediately.
• Your workspace membership and role are re-verified on every request, so a removed or downgraded team seat stops working instantly.
• Starter plans and lapsed subscriptions are blocked server-side — a connection can never outlive the workspace’s billing.

August 11, 2026

The AI Assistant can now build your configurator (controlled writes)

New

• Ask for changes in plain language: "Create a material selector for the chair upholstery", "Add Walnut, Oak and Black finishes", "Show the footrest only when Frame B is selected", "Increase leather prices by 12%".
• Every risky change (pricing, SKUs, deletions, bulk edits) first appears as a proposal card listing every change with its before → after value — nothing is applied until you press Apply, and proposals expire after 15 minutes.
• Small, safe changes (rename a block, add a variant) apply immediately and every AI change is undoable: an automatic snapshot is taken first and the chat shows an Undo button.
• Failed changes roll back automatically — the AI can never leave a half-finished project.

Improved

• The assistant knows which block, pricing block, or form field you have selected in the editor, so "add a green version" targets the panel you have open.
• Chat messages show exactly which tools read (eye icon) or changed (pencil icon) your project.
• After any applied change the editor reloads automatically so what you see always matches what was saved.

Security

• Team roles are enforced on every AI change: Viewer seats can never modify a project through the assistant, and every applied change is validated, rate-limited, and recorded in the audit log.

August 11, 2026

AI agents get their own documentation surface

New

• The AI Assistant now inspects your real saved project data through 25 read tools: it can audit a configurator for broken references, read the actual part names inside a 3D model file, explain why an option is hidden, and calculate a configuration price.
• Authenticated REST access to the same tools (GET /ai/tools, POST /ai/tools/execute) for your own automations.
• This changelog: a public record of platform releases at /docs/changelog, grouped by month — alongside the developer-facing API and SDK changelogs.

Improved

• Every docs page header now cross-links all documentation surfaces (Tutorials ⇄ API ⇄ SDK ⇄ MCP ⇄ Changelog).

August 10, 2026

Headless rendering: bring your own UI

New

• Clean viewer embed: append ?ui=viewer to a published share link and build your own option buttons, summary, and cart around it. A copy-ready snippet lives in Project → Share → Embed.
• A step-by-step headless integration tutorial walks through building a fully custom UI on the share data.

Fixed

• Wearable AR (virtual try-on) silently failed inside embeds because generated iframe snippets omitted the camera permission — re-copy your embed snippet to pick up the fix.
• Opening a saved-configuration link or an AR handoff link restored the selections into the UI but left the 3D model on defaults; both now repaint the scene correctly.

Security

• The public AR-analytics endpoint now validates share tokens and rate-limits per project and IP, so nobody can inflate or fabricate a project's AR metrics.

August 5, 2026

3D model loading made resilient

Improved

• Model downloads now retry transient gateway errors automatically and always show a status line instead of a silent black screen.
• Re-exported models using mesh compression (Draco, Meshopt, KTX2) now load in the editor, Preview, Share view, thumbnails, and wearable AR alike.
• A model that fails to parse shows an actionable banner naming the model and the fix, while the rest of the scene keeps loading.

Fixed

• Pattern Designer color channels that silently did nothing (colors written as rgb(), named colors, or quoted attributes) now recolor the model on every surface, and channels that cannot paint anything are flagged in the editor instead of shown to shoppers.