Soundbox v0.3.0

Added

• Custom endpoints. Settings → Endpoints sends sound effects, music or
  speech to a server you already run (LocalAI, an OpenAI-compatible speech
  server) instead of downloading the model. The request body is a JSON
  template with variables such as {{prompt}}, {{lyrics}}, {{bpm}} and
  {{key}}. Each endpoint appears in the model pickers. WAV and PCM are read
  directly; other formats need ffmpeg. Endpoints outside the local network
  are marked, since prompts leave the machine. See docs/ENDPOINTS.md. (#22)
• Updates in the desktop app. On launch, an installed copy checks for a
  newer release and offers to install it and restart. Models and projects are
  kept. Copies installed before this version need one manual download to get
  it.

Fixed

• GPU name and VRAM on NVIDIA. /api/health reports the GPU's name and
  its total and free VRAM from nvidia-smi, for the first GPU in
  CUDA_VISIBLE_DEVICES or GPU 0, so the sidebar, Settings and onboarding show
  them instead of only "cuda". Other GPUs still report null. (#17)
• SOUNDBOX_ALLOWED_HOSTS behind a reverse proxy. The server's Host
  check now accepts the names in SOUNDBOX_ALLOWED_HOSTS, so soundbox on
  127.0.0.1 behind Caddy or nginx no longer answers 403 "Forbidden host" to
  every request. Before, the variable only reached CORS and /mcp. A bare
  host matches any port and host:port only that port, the same in every
  check. (#22)
• Dev mode from another machine. npm run dev passes
  SOUNDBOX_ALLOWED_HOSTS to Vite's server.allowedHosts, and a new
  SOUNDBOX_DEV_HOST sets the address Vite listens on. Both are unset by
  default, which keeps the dev UI on localhost. (#22)

Installers are not yet signed. Windows shows a SmartScreen prompt (More info → Run anyway); macOS asks you to allow the app in System Settings → Privacy & Security. docs/DESKTOP.md has the details.