Salut is a pre-release local-network control plane for distributed llama.cpp inference. It discovers peers, records mutual trust, supervises pinned native processes, and exposes an OpenAI-compatible API. GGUF is the only model format.
Salut peers find each other automatically through mDNS. No peer address list is required for normal discovery.
Each node has an Ed25519 identity. Both sides must accept pairing before signed peer controls or computation transport can be used.
Salut supervises pinned llama.cpp executables and exact GGUF artifacts. Raw RPC remains on loopback and crosses the LAN through exact-certificate mutual TLS.
A matching loaded deployment can serve chat completions through the local OpenAI-compatible API.
Automatic loading selects local and paired Apple-silicon devices after checking their available memory. Windows CUDA execution remains in development.
Inference stays on the selected local devices. Salut does not send prompts or generated text to a hosted inference service.
Salut has not been released. The source checkout includes supervised llama.cpp processes, exact-artifact leases, paired HTTPS controls, encrypted RPC forwarding, and public chat dispatch for a matching loaded deployment. The Models pane and load API can open an existing local GGUF with automatic memory admission and placement across supported Metal and Windows CUDA devices. Windows workers run from source with verified runtimes and supervised native processes. Model downloads, mixed-platform production evidence, and signed packages remain development work.
The product path does not use SSH. SSH may be used to operate development machines while testing the application.
Use the source checkout for development. There is no release download yet.
Development Setup Release Status