Many foundational neuroimaging tools face challenges with complex dependencies, version conflicts, non-commercial licenses, and platform incompatibilities. While containerized environments like NeuroDesk help run these established toolsets, combining them into custom workflows or deploying them directly to the web remains difficult. Our mission is to provide these core capabilities as modular building blocks that run anywhere, empowering researchers and developers to build flexible, high-performance pipelines for novel scientific challenges.

Beyond modularity, our tools are built around four core commitments:

Native Execution

Porting high-level scripts to native, compiled code delivers order-of-magnitude speedups for data-heavy processing.

Modern Hardware Optimization

Established codebases offer significant opportunities for performance gains on modern hardware. We accelerate critical execution paths by integrating SIMD vectorization, FMA instructions, OpenMP parallelism, and cache-conscious memory layouts.

Cross-Vendor GPU Acceleration

Our graphics and compute pipelines are built on open, cross-platform standards including WebGPU, Metal, and Vulkan. This ensures peak native performance across the full spectrum of modern hardware, including AMD, Apple, Intel, and NVIDIA.

Frictionless Community Collaboration

To make open science as accessible as possible, our tools are distributed under widely accepted permissive licenses (BSD and MIT). This eliminates legal hurdles for university compliance teams, enterprise partners, and independent developers alike, allowing the entire community to share, build upon, and integrate these components freely.

Our software aims for functional equivalence with the standard tools the scientific community relies on, but with a drastically smaller footprint, minimal dependencies, and clear, open licensing. Note that “equivalent” does not mean bit-identical: modern hardware optimizations (such as fused multiply-add operations) preserve higher precision during intermediate calculations than traditional sequential floating-point routines.