Description
Booking a table, chasing a support line, returning a call: it eats time and needs little skill, yet most ways to hand it to an AI send your call audio to the cloud. ai-phone does the opposite. It is a SIP phone daemon on your own Mac, with speech recognition and synthesis running entirely locally, and the conversation handled by a local model, your Claude Code or Codex subscription, or any OpenAI-compatible endpoint.
Using it is direct: give one command a task such as asking whether they have a table for two at eight tonight, and it returns the full transcript when the call ends. Every outgoing call opens with a disclosure that the other side is talking to an AI. Another agent can also take over turn by turn from the command line, speaking and listening at its own pace.
Incoming and outgoing: choose receive-only, call-only or both during setup, using a SIP address, your own SIP trunk, or a Telnyx or Twilio number.
Local voice: Parakeet for recognition, Phonon or Kokoro for synthesis (American and British voices, adjustable speed), all through MLX on Apple Silicon.
Model by memory: setup detects your RAM and recommends a 4B model at 16GB, 9B at 32GB, or a 2B model for 8GB machines, then downloads the weights.
Agent relay mode: call, say, listen and hangup each return one JSON object, so an agent like Claude Code can run a call one turn at a time.
Free local test: pick a SIP address named agent and dial 127.0.0.1 with pjsua to try a call at no cost.
Limits: macOS on Apple Silicon only, English only, one call at a time. A remote language model sees the call text, and carrier charges apply. Licensed AGPL-3.0.
Using it is direct: give one command a task such as asking whether they have a table for two at eight tonight, and it returns the full transcript when the call ends. Every outgoing call opens with a disclosure that the other side is talking to an AI. Another agent can also take over turn by turn from the command line, speaking and listening at its own pace.
Features
Incoming and outgoing: choose receive-only, call-only or both during setup, using a SIP address, your own SIP trunk, or a Telnyx or Twilio number.
Local voice: Parakeet for recognition, Phonon or Kokoro for synthesis (American and British voices, adjustable speed), all through MLX on Apple Silicon.
Model by memory: setup detects your RAM and recommends a 4B model at 16GB, 9B at 32GB, or a 2B model for 8GB machines, then downloads the weights.
Agent relay mode: call, say, listen and hangup each return one JSON object, so an agent like Claude Code can run a call one turn at a time.
Free local test: pick a SIP address named agent and dial 127.0.0.1 with pjsua to try a call at no cost.
Limits: macOS on Apple Silicon only, English only, one call at a time. A remote language model sees the call text, and carrier charges apply. Licensed AGPL-3.0.
