ApexForge Is an Open-Source AI Workbench That Generates Editable 3D Models from Plain Language

I've been deeply working on AI Agents and large model applications recently, such as the JitKnow AI knowledge base, JitWord collaborative AI documents, and Pxcharts super tables. At the same time, I've been continuously sharing high-quality, practical AI open-source projects from GitHub that can truly be implemented and solve real problems.

Over the weekend, I took some time to write a 3D modeling tool — ApexForge.

The main purpose of building this open-source project is to help everyone generate interactive, commercial-grade Three.js 3D models using plain language. It also supports multiple model providers, online API Keys, model importing, texture uploading, and real-time property editing.

The image above shows a 3D model of a tape recorder generated by this open-source project I built. After several adjustments, I finally achieved decent rendering quality.

Without further ado, here's the link:

GitHub address: https://github.com/MrXujiang/ApexForge

Why Build ApexForge

ApexForge aims to solve a very specific problem: allowing product managers, designers, and developers to quickly get a previewable, adjustable, and further-engineerable 3D prototype using a single natural language sentence, without opening complex 3D software.

It also serves to provide a better theoretical and experiential foundation for my "AI Full-Stack Learning Handbook," helping everyone learn and implement AI applications more deeply.

It is not a simple model display page, but a complete AI 3D workflow:

• Natural Language Model Generation: Input descriptions like "retro portable radio," "futuristic aircraft," or "smart mirror" to generate corresponding parametric 3D models.
• Multi-Model Provider Access: Supports DeepSeek, Kimi, and Qwen, using the OpenAI-compatible calling protocol.
• Online API Key Configuration: Fill in the model Key directly on the left side of the page, without needing to modify the project's .env file.
• Three.js Real-time Preview: Complete 3D rendering, lighting, shadows, orbit controls, and model normalization within the browser.
• 3D Model Import: Supports common Three.js formats like GLB, GLTF, OBJ, and STL.
• Texture and Property Editing: Supports real-time adjustments for image textures, color, material, complexity, curve strength, and connection density.
• Comprehensive Developer Documentation: Built-in tech/doc multi-page HTML documentation covering architecture, frontend, backend, AI generation pipeline, and 3D rendering.

Since the recorded 3D video for the product preview is quite large, I'll use screenshots instead. You can refer to the generated effects below:

We can not only generate models but also apply textures to them. For example, if you want to add a custom sticker to a water cup you designed, you just need to upload the texture in the right-side editing panel:

After uploading the texture, it renders onto the 3D model in real-time. Below is the effect after I applied the texture:

Of course, everyone can also engage in more creative designs and texturing. The editor interface of the AI modeling platform I built references mainstream 3D modeling software, providing a large number of configurable properties. You can also extend it further to build a web-based 3D modeling software. Below is a display of the editor's complete interactive interface:

The core capabilities are summarized as follows:

This project adopts mainstream AI technologies such as Harness engineering, a multi-model routing perception system, prompt engineering, and Skills, which can help everyone quickly build AI cognition and practical experience.

Let's Talk About the Technical Architecture and Tech Stack I Adopted

The image above is the technical architecture I designed for ApexForge. You can refer to it; it includes the entire design chain of the AI modeling technology. The core tech stack is as follows:

Frontend:

• React 18 + TypeScript + Vite builds the Studio workbench.
• TailwindCSS + shadcn/ui style components implement a dark-themed, productized interface.
• lucide-react provides a unified icon system.
• Three.js is responsible for real-time 3D rendering, imported model parsing, and parametric geometry display.

Backend:

• NestJS 11 provides REST APIs and modular services.
• Prisma + SQLite saves templates, generation tasks, assets, versions, quality scores, and feedback.
• LLM Adapter unifies OpenAI-compatible /chat/completions calls for DeepSeek, Kimi, and Qwen.
• Harness Service generates multi-agent execution processes and 3D skill chain descriptions.

I also drew a flowchart for the entire project's AI model generation pipeline for your reference:

The project is now open-source on GitHub, and an online demo version is also available. If you're interested, you can refer to and experience it:

GitHub address: https://github.com/MrXujiang/ApexForge

Open-sourcing a project is not easy, so I hope everyone will give it a star~

I also welcome all tech experts to submit Issues and Pull Requests. Suggested contribution directions:

• Add new parametric model categories.
• Optimize Three.js materials, lighting, and post-processing.
• Enhance model import, export, and asset management.
• Integrate more OpenAI-compatible model providers.
• Improve the tech/doc developer documentation and examples.