AI Phone Agent

Why it's ranked here

It demonstrates the whole call path instead of stopping at a chatbot demo. SIP handling, browser calling, transcription, conversation state, speech output, transfers, monitoring, and call records are represented. The verdict is mixed because the primary flow is deliberately short and the real-time path still substitutes dummy transcription and a tone for working speech services.

What's good

The repository makes telecom boundaries visible. FreeSWITCH owns signaling and media while the TypeScript service owns dialogue and call state. The browser phone lowers the barrier to local testing. OpenRouter failures fall back to keyword replies, generated speech has a valid silent-audio fallback, shutdown covers all three servers, and per-call JSON records preserve operational context.

Tradeoffs

The turn-based flow records fixed five-second segments and ends after three turns, so conversation cannot feel continuous. Primary speech generation depends on macOS tools, while its fallback may play silence. The streaming path detects speech and moves audio bidirectionally, but transcription is hard-coded to a greeting and speech output is only a short tone. Local WebRTC also requires manual network configuration and has no STUN service.

How to use it well

Use it to learn or prototype the boundary between SIP infrastructure and language-model orchestration on a local macOS development machine. Start with the browser phone, inspect active calls and call records, then replace the streaming speech placeholders before testing realistic traffic. It suits engineers evaluating call flow and transfer mechanics. It does not supply production speech streaming, hosted telephony, or a deployment-ready receptionist service.

Observed

Primary language

TypeScript running on Node.js

Install surface

npm project installed with npm install; FreeSWITCH development environment uses Docker Compose

Call interfaces

Inbound SIP softphone and browser-based WebRTC phone

Service interfaces

FreeSWITCH ESL outbound socket, bidirectional WebSocket audio, and read-only HTTP JSON monitoring endpoints

Platform support

Primary text-to-speech path requires macOS say and afconvert; ffmpeg and sox conversion fallbacks are implemented

External service

OpenRouter provides language-model chat and recorded-audio transcription

Structural observation

The supplied repository tree lists no test directory or test script