Skip to content

Roadmap

v0.3.0: local models

Run the work and the voice on your own hardware, so a call and what it starts can stay on your machine.

  • A local model as a third agent, next to Claude and Codex, chosen per task
  • A local voice in place of the Realtime API: one speech-to-speech model, or speech recognition, a language model, and speech synthesis joined together
  • The PIN, the keypad, the call-backs, and interruptions working on the local voice
  • keryx setup and keryx doctor set up and check the local models

v0.4.0: talking to it at a Mac

Say a wake word and talk to the assistant through the Mac's microphone and speaker, with no phone and no Twilio. The work is on the feat/local-wakeword branch.

  • Test the wake word and the local channel on real hardware
  • keryx setup and keryx doctor cover the local channel
  • Merge feat/local-wakeword into main (macOS only at first)

Not scheduled: GPT-Live

GPT-Live listens while it speaks and keeps talking while a backend works, which suits Keryx. I'm looking into moving onto it without losing features. It isn't a model-name change:

  • Tools go through GPT-Live's delegated backend, with the PIN, the keypad, and the call-backs intact, and quick answers still quick
  • Keryx can tell when a reply ends, so it can hang up after a goodbye and stop when you interrupt, and the voice keeps the exact wording it's given
  • Giving the PIN still changes the voice's instructions mid-call, though GPT-Live only accepts additions after the call starts