TOOLCAD: Exploring Tool-Using Large Language Models in Text-to-CAD Generation with Reinforcement Learning

TOOLCAD addresses a high-value niche: the precision-heavy domain of CAD modeling where standard LLM generation fails due to lack of spatial coherence and syntax errors in CAD scripts. Its defensive moat is currently based on the complexity of the RL-based tool-optimization loop, which is significantly more sophisticated than simple prompt engineering. However, with 0 stars and 4 forks, it is currently a low-traction research project. The primary threat comes from two directions: 1) Frontier labs (OpenAI/Google) improving the innate 'agentic reasoning' and tool-use capabilities of base models, which could make the custom RL layer redundant, and 2) Industrial CAD giants (Autodesk, Dassault) who possess the proprietary training data (STEP/IGES files) and could integrate similar agentic workflows directly into their software. Compared to commercial-grade attempts like KittyCAD (Zoo), TOOLCAD is currently a academic reference implementation with limited ecosystem gravity. Its survival depends on moving from a paper-led project to a robust library that can handle industrial-scale CAD complexity.