GitHub - oxnw/ai-phone: A SIP phone daemon with local speech and language models on Apple Silicon.

Why it matters

A developer has published an open-source AI phone agent on GitHub that runs locally on a user's own computer rather than as a hosted service. The project, released under the AGPL license, handles speech recognition and synthesis entirely on-device using MLX, which restricts the host machine to Apple Silicon; callers on the other end can use any ordinary phone or SIP client. Setup guides the user through choosing incoming calls, outgoing calls, or both, then configures the model, phone line, and voice, recommending a local model based on detected memory. The tool currently supports English and one call at a time, and every outgoing call begins with an automated disclosure that the caller is an AI. A relay mode lets an external agent with shell access control each conversational turn. The developer reports that installation from a fresh clone was verified and local call tests have run, though live carrier and end-to-end agent acceptance checks remain pending. Local measurements of first-audible response missed the developer's stated target under heavy memory pressure.