The article from reshef.io explores a practical step toward AI-assisted CAD through the Onshape MCP plugin, a system that lets a language model like Claude directly operate a real CAD environment. Rather than chasing fully autonomous design, the project reframes the role of AI as a copilot that executes clearly defined design intent. When given precise, unambiguous instructions, the model can reliably construct geometry through a structured feature tree, shifting the bottleneck away from CAD execution toward interpretation of inputs.
At the core of the plugin is a toolkit of roughly 60 operations covering sketches, extrusions, fillets, assemblies, and parametric variables. Each operation returns structured feedback about what changed, what failed, and why, eliminating the silent errors common in traditional CAD workflows. This “truth-telling” approach ensures that every modeling step is inspectable and correctable in real time.
A second key feature is the integration of vision-based workflows. The system introduces a “vision decomposition” process where the model first analyzes a reference image step by step, generating a structured description of features before building. While this does not fully solve the challenge of interpreting engineering drawings, it exposes errors early and allows human intervention before costly mistakes propagate.
The experiments highlight a critical insight: the limitation of AI-driven CAD is not geometry generation or software capability but visual understanding. Models can execute design logic effectively once the specification is clear, but they struggle to infer precise geometry from images alone.
Ultimately, the plugin demonstrates a shift in CAD interaction. Instead of navigating complex interfaces, engineers describe designs in natural language, review the generated feature tree, and iterate conversationally. This hybrid workflow, combining human intent with machine execution, suggests a near-term future where AI augments design productivity without replacing engineering judgment.