Tentacl
Distributed, local-first mix intelligence: per-channel feature extractors stream to an engine that suggests bounded moves toward a reference.
What it is
Mixing engineers carry reference tracks in their heads; Tentacl makes that explicit and local. Lightweight VST3/AU plugins sit on each channel and stream features — level, spectral balance, dynamics — to a C++ mix engine that compares your mix against a reference and returns bounded, conservative processing suggestions to a Tauri desktop app. Nothing leaves your machine.
Architecture
A monorepo: C++20 channel extractors and the comparison engine, Protobuf schemas on the wire, a Tauri + React control surface, and a PyTorch role-classifier exported to ONNX that labels each channel (kick, vocal, pad) so suggestions are role-aware. The hard constraints — real-time safety in the audio thread, bounded suggestions that can't wreck a mix — drove the whole design.