Matt Keeter has been building geometry kernels based on implicit surfaces since his Master's thesis, and Halfspace is the latest artifact from that obsession: a cross-platform IDE for solid modeling with distance fields, powered by his Fidget kernel. The project shipped in mid-2025 with a web demo you can try right now, and it's the rare tool announcement that earns its length by teaching you something about the domain it operates in. The core insight Halfspace foregrounds is that distance fields — not just the binary inside/outside classification of constructive solid geometry — matter for getting correct normals, clean meshes, and well-behaved algorithms. Keeter demonstrates this with a sawtooth wave example: two fields with identical signs everywhere but different gradient behavior, where one produces broken normals and ugly shading on a cabin's shingles. Putting the field values front-and-center in the IDE makes these problems diagnosable. This is genuine design thinking, not feature-list engineering. The technical stack is ambitious for a solo project: Rust end-to-end, egui for the GUI, wgpu for both rendering and GPU compute, Rhai for scripting, wasm-bindgen for web targets, Rayon for parallelism. The fact that the same codebase ships native and web — with GPU-accelerated rasterization on both — is a real achievement, even if Keeter admits he has complaints about every layer. The biggest win was moving rasterization and post-processing entirely to the GPU, eliminating the WebAssembly performance bottleneck that made the web version non-interactive. What makes this interesting beyond the tech is the philosophy. Keeter draws an analogy between working with raw implicit surfaces and writing assembly language, then identifies two responses: build high-level abstractions (standard libraries of shapes and transforms), or make the low-level work less painful. Halfspace does both. He cites Kartik Agaram's Mu project — which wraps tracing and time-travel debugging around x86 assembly — as a kindred spirit. The IDE lets you split complex models into parameterized sub-pieces and visualize them individually, which is the implicit-surface equivalent of a REPL. The honest framing is refreshing. Keeter explicitly warns against using Halfspace for anything load-bearing, notes he's willing to aggressively iterate and break APIs, and positions the project as a vehicle for driving improvements in Fidget rather than a product seeking users. He also makes a point of noting this isn't vibe-coded — he wrote it with his human brain, starting April 2025, building on Fidget work dating to 2022. The limitation is scope: this is an experimental showcase for one person's geometry kernel, not a production CAD tool. The standard library is small but growing, mobile support is incomplete, and the project exists primarily because Keeter wants to learn from it. That's fine — the value here is in the ideas about distance field visualization, the cross-platform Rust/WebGPU stack patterns, and the existence proof that a single developer can build a functional solid modeling IDE with GPU-accelerated rendering on both native and web targets.
