跪拜 Guibai
← All articles
Frontend · GitHub · Architecture

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

By 徐小夕 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

ApexForge collapses the gap between a product idea and a tweakable 3D prototype into a single browser tab. For teams that lack dedicated 3D tooling, it removes the Blender learning curve and the cost of proprietary CAD licenses while keeping the output in standard, engineerable formats.

Summary

ApexForge is a newly open-sourced AI-driven 3D CAD and parametric modeling workbench built over a weekend. It accepts plain-language prompts — "retro portable radio," "futuristic aircraft" — and produces interactive Three.js models that render directly in the browser with lighting, shadows, and orbit controls. The editor mimics the layout of professional 3D software, exposing sliders for complexity, curve strength, and connection density alongside a texture-upload panel that applies images to the model in real time.

The toolchain is a full-stack TypeScript project: a React 18 + Vite frontend styled with TailwindCSS and shadcn/ui, a NestJS 11 backend with Prisma and SQLite for task and asset persistence, and an LLM adapter that routes prompts to DeepSeek, Kimi, or Qwen through an OpenAI-compatible protocol. A Harness service orchestrates multi-agent execution and logs the 3D skill chain for each generation.

Standard 3D formats — GLB, GLTF, OBJ, STL — can be imported directly into the workspace, and API keys are configured in the UI rather than in environment files. The project ships with multi-page HTML developer docs covering architecture, the AI generation pipeline, and the rendering stack.

Takeaways
A plain-language prompt generates a parametric Three.js model that renders immediately in the browser with lighting, shadows, and orbit controls.
The editor supports real-time texture uploads: an image applied to a model updates the render without a page reload.
API keys for DeepSeek, Kimi, and Qwen are configured through an in-UI panel, not by editing .env files.
Existing 3D assets in GLB, GLTF, OBJ, and STL formats can be imported into the workspace.
Adjustable properties include color, material, complexity, curve strength, and connection density.
The backend uses NestJS 11 with Prisma and SQLite to store templates, generation tasks, assets, versions, and quality scores.
A Harness service orchestrates multi-agent execution and documents the 3D skill chain for each generation.
Conclusions

Bundling an LLM adapter that normalizes multiple Chinese model providers behind an OpenAI-compatible endpoint is a pragmatic pattern for any tool that wants to avoid vendor lock-in without building custom integrations.

Putting API-key configuration in the UI rather than in environment files lowers the barrier for non-developers — product managers and designers can use the tool without touching a terminal or a config file.

The project treats the generated 3D output as a starting point for further engineering, not a dead-end render, by supporting standard import/export formats and exposing parametric controls.

Concepts & terms
Harness Service
A backend service that orchestrates multiple AI agents to execute a task and records the sequence of skills used, producing an auditable chain of reasoning for a generation.
LLM Adapter
A unified interface that translates requests from different large language model providers into a single, OpenAI-compatible /chat/completions call, so the application can swap models without changing its core logic.
Parametric 3D Model
A 3D model defined by adjustable parameters — such as dimensions, curve strength, or connection density — rather than a fixed mesh, allowing the shape to be modified after generation.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗