CraftBot: An AI Architect That Writes Buildings as Code, and What Trainers Can Learn From It

Most AI 3D tools give you a mesh or a picture. CraftBot, an open research project by Luka Piskorec of {protocell:labs}, does something different: it writes the building as code. The agent reads a brief and reference material, writes Python that constructs each member of the structure in Blender, checks the result, and revises until the design converges. That approach has some useful lessons for anyone building construction training material.

The researchers are open about the limit. In their own summary, CraftBot remains largely incapable of finding its own geometric errors, so a human still has to point at the clash.

What CraftBot is

CraftBot is described in its GitHub repository as an architect AI agent. It interprets design briefs, grounds itself in domain knowledge by ingesting documents, images and other references, and outputs Python code that defines a building procedurally. From that model it can produce standard representations such as floor plans, sections, elevations, bills of quantities and BIM models, according to the project.

It is a research project, connected to work at Aalto University in Finland, where a page describes it as a way to test whether language models can produce physical designs that are structurally sound and buildable. The repository holds the code, the references and the outputs of 14 experiments so far, and is released under the MIT licence.

How the loop works

1. Brief and references. A design brief, plus photographs, drawings and construction manuals.
2. Python constructors. The agent writes scripts that build each member of the structure, rather than drawing a shape.
3. Headless Blender. The scripts run in Blender without a screen, then the model is rendered from set views.
4. Checks. Viewport screenshots are reviewed, and the repository includes an overlap check that flags parts passing through each other, and a contact check for parts that touch nothing.
5. Revise. The agent changes the code and repeats until the brief is met.

Because the design is code, the rules live in one place. Change a dimension once and everything that depends on it moves with it.

The experiments that matter for construction

The 14 experiments group by the kind of reference they used. Several read like a training syllabus:

• Carports with king post and hammer beam trusses (experiments 1 and 2), built from visual references.
• A prefabricated timber house from a construction manual (4).
• A staircase (10), and a hip roof, a dormer window and roof sheathing (11 to 13), which are the parts that most often go wrong in AI-built houses.
• A ten-storey timber building (9) and a timber tower refurbished as a house (14).

CraftBot and Blender MCP compared

Neither replaces the other. If you only need one explainer, a live scene you steer is faster. If you want repeatable, checkable models, the code-first idea is worth borrowing.

What trainers can borrow

You do not need to run the whole agent to use the ideas. Useful patterns in the repository include:

• Skill files for construction topics. The repository includes written procedures for timber framing, roof framing and sheathing, structural logic, and working from reference documents, which are worth reading as checklists.
• Derive once, never type a number twice. Tie dimensions together so a change flows through the model.
• Verify with a check, not just your eyes. An overlap test catches errors a quick look can miss.
• A viewer with layers. Being able to hide cladding and see the frame is the same idea as an inspection pane for learners.
• Manuals as source material. The project extracts timber construction manuals into summaries the agent works from, and leaves the copyrighted PDFs out of the repository.

Cautions

This is research, not a design tool, and the output is not engineered or code-compliant. A model that looks right is not the same as a structure that is sound, and, as the researchers say, human review is still needed to find errors. Agent runs of this kind can use a lot of tokens. Treat it as a source of ideas for training material, and check any dimensions, spans and member sizes against the standards and the engineer’s documents that apply to your project.

Links