A high-performance on-device LLM inference engine written in Rust and Metal, built at Mirai to run models locally on Apple silicon with zero latency and full data privacy.

A Metal Shading Language language server (LSP) for VS Code, Cursor, Zed, and IntelliJ, offering real-time diagnostics via xcrun metal, auto-completion, hover documentation, and clang-format-based formatting.

Rust bindings for the Apple Neural Engine via the private AppleNeuralEngine.framework, providing a graph builder and compilation pipeline with zero-copy IOSurface-backed I/O.

Rust bindings for Apple's Metal API, built on the modern objc2 ecosystem with coverage of the latest Metal 4 API.

Modern Rust bindings for Apple's Metal Performance Shaders Graph framework, exposing a high-level, type-safe graph API for defining and running neural networks on the GPU.

MetalTools provides a convenient, Swifty way of working with Metal. This library contains a lot of read-to-use compute kernels and is heavily used in computer vision startups like Prisma and ZERO10.

A set of tools and extensions that allow sharing page-aligned memory allocation between different graphics APIs and provide view interop for them.

CoreVideoTools offers a more idiomatic Swift interface to CoreVideo functionality, making it easier and safer to work with CVPixelBuffers, IOSurfaces, and related CoreVideo concepts in Swift code.

Overhead-free ONNX graph inference engine with Metal Performance Shaders under the hood.

A Rust library providing utility functions and abstractions for working with WGPU.