Unique features
SG-Lib is not another geometry kernel and not a CAD GUI, but a code-driven design-to-manufacture library: a triangulated surface mesh (B-Rep, not CSG), contour-first construction, frame-based assembly (align/match), a built-in connection designer [H,N,S,Supp,Parts] with fits, standard parts and a bill of materials, additive-manufacturing correctness by construction (nofrd/Nyquist) and self-documenting scripts (XX). Boolean, projection and slicing run — since July 2026 — on the embedded Manifold library (fallback: SG-Lib’s own routines).
LLM auto-construction: SG-Lib can be driven via the ChatGPT API since 2024 and via Claude since July 2026. From a task description it generates MATLAB code that produces not only the geometry, but also the construction path, the bill of materials and the suppliers’ web addresses — geometry + path + parts list + sourcing in a single run.
Comparison with alternative approaches
| Category & examples | The alternative’s approach | Difference to SG-Lib | SG-Lib advantage |
|---|---|---|---|
| Script CAD with CSG — OpenSCAD, JSCAD, ImplicitCAD | Text script builds an exact CSG tree from primitives via boolean chains | CSG instead of triangulated mesh B-Rep; no frames, fits or BOM; no manufacturing knowledge | Contour-first instead of fragile boolean chains (triangle-sparse), relative frame assembly, DfM/fits built in |
| Parametric code CAD — CadQuery, build123d (OpenCASCADE) | Python API on an exact B-Rep/NURBS kernel; sketch → feature, locations/joints | Exact kernel instead of mesh; no AM tessellation model (Nyquist); no standard-part/connection catalog with BOM | Printable by construction (nofrd, mm) + machine-element kit [H,N,S,Supp,Parts] with parts list/product codes |
| Mesh / geometry libraries — libigl, CGAL, trimesh, Open3D, VTK | Toolboxes for mesh operations (some booleans), without construction semantics | Pure geometry building blocks without joining/assembly logic, frames or BOM | A complete construction library instead of a toolbox; uses such kernels as substrate |
| Robust boolean kernel — Manifold | Guaranteed watertight mesh boolean kernel; a pure kernel without a construction layer | No frame/DfM/BOM concept; solves “only” robustness and performance | Manifold is SG-Lib’s substrate since July 2026 (boolean/projection/slicing); SG-Lib adds frames, fits, printability, BOM |
| Exact CAD kernels & GUI — OpenCASCADE, SolveSpace, Rhino/Grasshopper | Exact solids/NURBS, interactive or constraint-based | Kernel/GUI-centric, partly proprietary; no code-driven DfM automation with a BOM | Scriptable/reproducible, self-documenting (XX); manufacturability and BOM emerge automatically |
| Open-source parametric CAD with a Python API — FreeCAD | Parametric feature modelling on OpenCASCADE inside a document model; full Python API, workbenches (Sketcher, FEM, TechDraw, BIM) | Exact kernel and a GUI document model instead of a triangulated mesh B-Rep; scripting automates GUI objects; no AM tessellation model, no standard-part/connection designer with a BOM, no LLM control | A pure code library without a document model; printability, fits and the parts list arise while constructing, not as a downstream export |
| AI-native CAD platform — Zoo (zoo.dev) with the KCL language | An AI agent plus its own constraint language KCL in a studio application (Mac/Windows/Linux); by its own description it produces “B-rep geometry, no meshes”; sketching, prompting and coding in one workspace | Its own language and its own platform instead of a library inside an established computing environment; exact B-Rep instead of an AM mesh; no frame/fit/standard-part kit with a bill of materials and suppliers | SG-Lib is a MATLAB/Python library that any agent can download and embed — the automatic construction then runs in the user’s own environment, for example through ChatGPT Codex or Claude Code, instead of inside someone else’s platform. It runs locally and is free for academic use; the LLM control yields not only geometry but also the construction path, the parts list and the suppliers’ URLs |
| Implicit / design-for-AM — nTop, libfive, SDF/Fidget | Implicit fields (SDF/F-Rep); strong at lattices/TPMS and topology optimization | Fields instead of mesh B-Rep; no frame/connection kit with standard parts | Mechanical connections, standard parts, assembly order + finished printable parts instead of pure lattices |
| Frame / mechanism toolkits — ROS/URDF, MuJoCo, Robotics Toolbox | Frames, kinematics and simulation of mechanisms/robots | Do not produce printable geometry with fits — only the “frame half” | Combines frame semantics with manufacturing-ready geometry: align/match become real, assemblable parts |
What makes it unique
No other known project unites in one code-driven environment: mesh B-Rep + contour-first + frame-based assembly + a built-in machine-element/connection designer with a bill of materials + AM-correctness by construction + self-documenting, shareable files (frames encoded in the STL) + LLM control (ChatGPT since 2024, Claude since July 2026) that produces geometry, construction path, parts list and supplier URLs in one run.
The closest contender is Zoo (zoo.dev). Zoo genuinely occupies the LLM pillar: an AI agent turns a prompt into editable CAD. The remaining pillars are missing there — Zoo delivers exact B-Rep rather than an AM-correct mesh, has no frame-based assembly, and no standard-part/fit kit with a bill of materials and suppliers. And Zoo is a platform with its own language (KCL), whereas SG-Lib is a library that an agent downloads and embeds into its own environment. The other projects each cover one of these pillars — SG-Lib the integration across all.
This comparison is the personal assessment of Tim Lueth. I may be mistaken about the features of the other projects — corrections are welcome.
Third-party project states web-checked (Manifold 3.5.4, FreeCAD 1.1.4, build123d 0.13.0, CadQuery 2.8.0, OpenSCAD 2021.01 = last stable, nTop, Zoo/KCL). Research date: 2026-10-27.