You Can’t Inspect a PDF on a Slab: Blender MCP and Three.js for Construction Training

A two-storey house looks simple until you ask a model to put a roof on it. Then you remember why truss shops exist.

Construction training often falls into the gap between a 2D drawing and a learner who has never stood under a hip roof. Slides and code extracts do not close that gap. A 3D scene you can orbit, isolate and click can. This article looks at using Blender, MCP and Three.js to build that kind of explainer, and what it took on a real two-storey build.

This is not “AI builds the house.” It is AI as a fast draughtsperson that you still have to supervise closely.

What the stack is

Three pieces, and it helps to keep them separate:

The useful part is not generation, it is inspection. Once the model is steered into shape, ask the assistant for the scene information, then for a Three.js viewer that matches the names you already gave the parts.

Why it suits construction training

Learners tend to struggle in the same three ways. A named 3D model helps with each:

• Sequencing. Slab before frame before roof is obvious on site and easy to lose in a paragraph.
• Interfaces. Where a truss sits on the top plate, where the wet area stops, where an inspection hold point sits in space.
• Reading a drawing. Plan, elevation and section stay three separate documents instead of one object.

That only works if you name parts carefully. Name meshes to match your zone or ITP references (for example Building_Shell, Floor_1, Zone_Wet), export a GLB, then in Three.js find them by name. That gives you a clickable inspection view, not a render farm. Here is what that changes for common explainer jobs:

• Massing a two-storey: instead of one elevation, orbit and isolate Floor_1 and Floor_2.
• Roof and truss logic: instead of a detail sheet nobody opens, hide the cladding and leave the truss set.
• Inspection and ITPs: instead of a checklist in a Word file, click a wet-area zone and see its evidence.
• Sequencing: instead of a Gantt chart, a four-stage scene of setout, massing, frame and envelope.

What we learned on a real two-storey

We ran a two-storey house through the Pascal editor and into Blender, then worked the roof. It was not instant, and it used a lot of tokens. Left unsupervised, the model does not produce something a builder could build.

The roof and the stairs are where it breaks down. Photos alone do not give you scale, so someone still needs a tape measure. When the truss set finally landed, the honest caption was: “Imagine getting the truss guys in for this one.”

That is useful for teaching, not a flaw. Show learners the intermediate result, the hip roof that came out wrong, and the steer that fixed it. Quality planning is the habit of not accepting the first output.

How to run an explainer without building a game studio

Keep the pipeline simple:

1. Mass in Blender via MCP. One prompt per stage: slab, walls, floors, openings, roof last.
2. Name parts like an ITP. Slab_OnGround, Frame_L1, WetArea_Ensuite, Roof_TrussSet. If you cannot click it later, it was named wrong.
3. Steer the hard parts. Roofs, stairs, eaves and falls to waste are not one-shot jobs.
4. Export a GLB once the hierarchy is clean, not before.
5. Load it in Three.js with orbit controls, click-to-isolate, and a side panel for the hold point, RFI or lesson learned.
6. Record the orbit if your LMS needs a linear video. The 3D file is the source and the video is a copy.

Later on, you can look at Geometry Nodes or shader work for parametric facades and materials, but that is craft for after the first explainer.

What this is not

It will not produce structure that complies with the National Construction Code. It will not size a lintel, and it will not replace a truss fabricator’s software. Treat the scene as a spatial story, and keep calculations in the tools that own them. If you only need a simple massing diagram, a plain Three.js scene may be faster, so bring in Blender when the learner needs to see parts such as plates, trusses and junctions.

Five explainers worth building first

Each is a reusable GLB plus a thin viewer, so one naming convention can serve many units:

1. Slab-on-ground sequence: setout, edge beams, vapour barrier, pour. Click each layer.
2. Two-storey load path: not engineering, just “this sits on that.”
3. Wet-area hold points: membrane, falls, flange, flood test, with the ITP text in a side panel.
4. The roof that went wrong: a steered hip beside an unsupervised one. Probably the best lesson of the set.
5. Site inspection walk: an RFI pinned to a beam, and the learner finds the defect.

A follow-up article covers how to set up Blender MCP so you can try this yourself.

Links and further reading

Blender MCP is a third-party community project (MIT licensed) and is not made by Blender. Its package has recently been renamed MCP for Blender, and older setups keep working.

Bottom line

Blender MCP is useful in construction education because a practitioner who already knows the build can direct geometry in plain language, then hand a named scene to the browser. Three.js means the learner does not need Blender to inspect the result. The leverage is the loop: steer in the workshop, explain in the browser, and keep the names aligned with how the job is actually inspected. Do not ask it to be the truss fabricator. Ask it to put the truss in the room so a first-year apprentice can see why that trade exists.