---
title: "Construct 1.0: Agents Build 3D Models and Label What They Know"
slug: "construct-ai-agent-3d-models-provenance-labels"
author: "ibl.ai Engineering"
date: "2026-11-02 11:00:00"
category: "Premium"
topics: "construct, ai agents, 3d models, gltf, blender, open source ai, provenance, agent evaluation, independent critic, enterprise ai verification"
summary: "Construct 1.0 turns a plain-language request into a 3D model with named parts, under Apache-2.0. It doesn't scan objects or replace CAD. Its useful idea is that every part is labelled verified, supplied, inferred or decorative."
banner: ""
thumbnail: ""
linkedin: |
  An open-source project called Construct reached 1.0 on October 27, and the claims going around it are mostly wrong.

  It doesn't scan real objects. You describe a thing in plain words, an agent looks up facts and reference photos, then writes a text file describing the parts.

  It isn't MIT-licensed any more. It moved to Apache-2.0 after version 0.3.0.

  And it doesn't replace enterprise CAD. Its README talks to game developers, 3D artists, film makers, educators and museums.

  What it does do is more interesting than the hype.

  Every part in every model is labelled verified, supplied by you, inferred, or decorative. Press P and the model shows you which is which. On its Saturn, the rings are published data and the cloud bands were designed by the model, and you can see that.

  Most agent output, a report, a summary, a forecast, reads in one confident voice. Imagine asking your agents to mark each claim the same way.

  That idea will outlast the 3D.

  #iblai #AIAgents #OpenSource #AIGovernance #EnterpriseAI
---

## The Short Answer

**Construct 1.0 is an Apache-2.0 tool that lets an AI agent turn a plain-language request into a 3D model with named parts, for Blender and game engines. It doesn't scan objects or replace CAD. Its lesson is provenance: every part says how it is known. With ibl.ai you own all the code and the data your agents reason over.**

Construct is small and new. The [repository](https://github.com/rtmandreyev/construct), by Artyom Andreyev, was created on October 24, 2026, and [v1.0.0](https://github.com/rtmandreyev/construct/releases/tag/v1.0.0) followed on October 27. It had 102 stars when we checked.

## What does Construct 1.0 actually do?

Construct has an AI agent describe an object as data, then turns that description into a hologram page and a 3D model. You ask in plain words, and the agent looks up facts for a named subject. It then writes a scene: a text file listing facts and parts, built from simple shapes.

Construct checks the scene, builds one offline `construct.html`, and the agent reviews stills from each view.

The [1.0 release](https://github.com/rtmandreyev/construct/releases/tag/v1.0.0) adds export to glTF 2.0 and OBJ, with named parts and working joints, plus a Blender add-on. The release notes say the output was checked in Godot 4.7.2 and passes the Khronos glTF Validator.

It works with Claude Code, Hermes, any MCP client or any OpenAI-compatible endpoint. It needs no packages beyond Python 3.10 or newer, though review stills need a Chromium browser and video needs ffmpeg.

## Does Construct scan real-world objects into 3D models?

No. Construct doesn't scan anything, despite how it has been described. The input is a sentence. For a named subject, reference photos are "only for comparison" and "never copied into the construct or published".

Three other claims in circulation need correcting too:

<table>
<thead><tr><th>Claim</th><th>What the repository says</th></tr></thead>
<tbody>
<tr><td>It scans objects</td><td>It builds from a text request, and the agent writes the parts</td></tr>
<tr><td>It is MIT-licensed</td><td>Apache-2.0, with versions up to 0.3.0 under MIT. The Blender add-on is GPL-3.0-or-later, and four NetLogo examples are CC BY-NC-SA 3.0</td></tr>
<tr><td>It replaces $25,000-a-seat tools</td><td>We found no source for that figure. Examples are independent reconstructions, and the listed audience is game, 3D, film, education and museum work</td></tr>
<tr><td>Any agent can do it</td><td>Any can connect, but the three small models tested failed to save a scene in about half of runs</td></tr>
</tbody>
</table>

Limits are stated plainly too. The README says it has been tested on macOS only, and masterpiece mode can use up to two hours of frontier-model tokens. Those are honest disclosures, and they are part of why we would trust the project.

## Why do provenance labels matter for enterprise AI agents?

Provenance labels make an agent separate what it knows from what it made up, and enterprise output rarely does this. In Construct, every fact and part is marked verified, supplied by you, inferred or decorative, and pressing P colours the model by those labels.

The README's Saturn example shows the value. The rings come from published data, while the cloud bands were designed by the model. Both look equally convincing, and only the label tells them apart.

The same labels travel with the export, in the glTF file's `extras.construct` field. Construct's check also flags "a claim stronger than its evidence" for repair.

A summary, risk memo or forecast from an agent has this problem in a worse form, because it reads in one confident voice. A reader can't see which sentence came from a record and which was the model filling a gap.

## What other agent design patterns does Construct demonstrate?

Construct shows two more patterns worth borrowing, and both are about limiting what a model decides.

**Data, not code.** "A scene is data, never code": the agent says what the thing is, and Construct decides how it looks. That shrinks what can go wrong and makes the output checkable.

**An independent critic.** In masterpiece mode, a fresh model compares each version with the best so far without knowing which is which. Only a better version is kept. It stops after two rounds without a win, after 12 rounds, or after two hours.

The author's blind comparisons, on two subjects, showed no visible gain after about round 12. That is a small test, but publishing the stopping rule beats promising perfection.

## How can you apply provenance to agents on ibl.ai?

You can write the same discipline into your own agents. On ibl.ai, a [Skill](/docs/os/agent-settings/skills) is a plain Markdown file, so a domain expert can require an agent to tag each claim as sourced, supplied or inferred. This is a practice you configure, not a built-in label set.

[Evaluations](/docs/os/agent-settings/evals) can then score whether agents do it, with an LLM judge and human annotation. [Transcripts](/docs/os/analytics/transcripts) keep the record for review. For the same close reading applied to open-source licenses, see [our look at agent builders in K-12](/blog/open-source-agent-builders-k12-classroom-guardrails).

With ibl.ai you own all the code and the data, the platform is model-agnostic across any LLM, it has no per-seat pricing, and you can deploy anywhere, so those rules live in your perimeter.

ibl.ai is family-owned and operated from New York, NY.

*Sources: all Construct facts are from its README, documentation and release notes. Star count as of November 2, 2026.*

## Why does owning the AI stack matter?

**ibl.ai is the agentic AI platform where you own all the code and the data. You self-host the entire stack inside your own perimeter, run it model-agnostic across any LLM and switch anytime, and pay by usage with no per-seat pricing — so you can deploy anywhere: your cloud, on-premise, GovCloud, or fully air-gapped.**

- **You own all the code and the data.** Full source code under a perpetual license, running on your infrastructure. Not API access to someone else's platform — the stack itself is yours.
- **Model-agnostic.** Run any LLM — Claude, GPT, Gemini, Llama, Command, or your own fine-tune — and switch providers without rewriting the platform.
- **No per-seat pricing.** Usage-based billing against a budget cap you set. Cost tracks what your organization actually uses, not how many people you employ.
- **Deploy anywhere.** Your cloud, your VPC, on-premise, GovCloud, or a fully air-gapped network with no outbound connectivity.

1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.

ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.