Bldr AI Takes Browser CAD Into IFC 4.3 and AI Design
Bldr AI, the small open-source team behind the browser-based Share viewer, is making two moves at once this autumn. Its engine is learning to draw IFC 4.3 infrastructure files, and its developers have opened an epic to scope AI that builds models from a text prompt. Together they point to a plan that goes well beyond a simple viewer.
For anyone who works with BIM and CAD files, the timing matters. Bldrs.ai describes Share as high-performance, browser-based CAD visualization with no SaaS upload, which is a different bet from cloud-first rivals. Here is what is new, what is still unfinished, and where it could go.
What is Bldr AI, and why do engineers keep sharing links to it?
Share is a browser-based BIM and CAD viewer and collaboration platform that supports IFC, STEP, STL, OBJ and GLTF. The workflow is deliberately light. According to the project’s README, you can open any IFC model on GitHub by pasting its address into the search bar, or upload one from your desktop. From there you can navigate the model’s structure, use cut planes to look inside, and search elements and properties.
The sharing feature is what gives the tool its name. Permalinks point to specific model parts with exact camera views, so a colleague lands on the same beam or duct you were looking at. That removes a lot of screenshot-and-email back and forth.
The privacy design is the other differentiator for Bldr AI. Share is built in React and Three.js/WebGL, handles files client-side using the Origin Private File System, and uploads nothing to a SaaS backend. GitHub sign-in is optional and adds git-based model versioning, visual issue tracking and large file sharing. The team is also small: a directory listing puts it at 2–10 employees, spread across the USA, Australia and Switzerland.
IFC 4.3: the road-shaped hole Conway is now filling
Share relies on Conway, its home-built geometry engine. The project says Conway offers full IFC 2×3 and 4 coverage, with bindings for more than 900 entities across 7,000+ pages of specification. That is a serious base for a young engine.
For Bldr AI, IFC 4.3 is where the work is happening now. It extends the standard to roads, rail and other infrastructure, and the gap has been plain in the developers’ own notes: Conway can parse IFC 4.3 but cannot yet draw most of it, because the geometry pipeline is typed against IFC4. Rather than draw files wrongly, the engine refuses the 4.3 files it would otherwise render incorrectly.
The newest release opens a narrow path through that wall. A recent pull request lets IFC 4.3 files whose geometry is IFC4-shaped load through the IFC4 pipeline. The proof case is a test road model that now renders 66 products, made up of 39 pavements and 27 kerbs. That file had previously never rendered in Conway.
Around it, the engine has been shipping rapid fixes. One release (1.1604.715) is labelled “deflection certificate,” and the two before it covered b-spline face triangulation and tessellation subdivision. A separate fix corrected a bug where a driveway’s curved edge was wrongly bounded. These are the unglamorous geometry details that decide whether a browser viewer is trusted with real project files.
Still missing is the heavy infrastructure content. The roadmap lists alignment curves, spirals and linear placement as the next phase. The team has also decided to report IFC 4.3 “not supported yet” refusals to Sentry, grouped by schema, so real-world frequency can rank the 4.3 work. That means user demand, not guesswork, will set the order.
From viewer to teammate: the text-to-model experiment
The AI side of Bldr AI is newer and much earlier. The organization’s GitHub page says it wants to build an AI that participates like a teammate rather than a tool, working in a space shared by humans, AI and open standards.
The concrete step arrived on October 5, when a maintainer opened an epic for an “AI Create head”: text or prompt in, model out, followed by iterative edits. This is a plan, not a shipped feature, and it begins with a research question.
Three engines, one open question
The epic asks which engine produces the geometry, and in what representation. It lists three candidates. The first is a Conway write path that mutates and re-emits IFC/STEP semantics, which is native to the team’s formats but does not exist yet. The second is CADAM, an open-source text-to-CAD app, whose fit on license, runtime and data sovereignty is the subject of the research story. The third is direct mesh or parametric primitives, which need no engine but are limited in expressiveness.
Where Bldr AI still has to prove itself
The pace is real, but so are the limits. IFC 4.3 support covers only a slice of files today. Some collaboration features, such as commenting on model parts, were still listed as in development in the README. Bug reports also show typical growing pains, such as images in notes on private models failing to load with a 404 error. A team this small has to choose its battles.
Visibility is modest too. One directory ranked it 99th of 417 in AI project management as of January 2026, which says more about category noise than about the product itself. The open-source licensing may help here, since Share is released under the AGPL and the team actively invites contributors.
Quick answers about Bldr AI
Is Bldr AI free to use? The code is open source, and the project asks people who use it for private hosting to contribute to its Open Collective. A hosted-pricing page was not found, and one directory lists pricing as unknown, so check the official site before relying on it for a commercial workflow.
Does Share upload my model to a server? By its own description, no. Files are handled client-side, and uploads to a SaaS backend are not part of the design. Only the optional GitHub integration involves remote storage.
Can it open IFC 4.3 road models? Partly. Files whose geometry is IFC4-shaped now render, but full alignment support is still on the roadmap.
Is there a Bldr AI assistant I can use today? Not as a shipped feature. The text-to-model work is at the engine-selection research stage.
The bottom line
The project is a small team making a clear bet: that engineering models should open in a browser, stay in the user’s hands, and eventually be shaped by an AI collaborator. The IFC 4.3 progress shows steady engineering. The text-to-model epic shows ambition, with a long road ahead. Watch the Conway releases and that epic over the coming weeks to see which way the project goes.