voice-agent-kit · packages

Orchestrates STT, MCP, and TTS pipelines for voice-enabled AI agents using an event-driven `Pipeline` class and session management utilities. It provides latency enforcement, OpenTelemetry-instrumented lifecycle hooks, and Zod-validated configuration schemas.

Connects to Model Context Protocol (MCP) servers via JSON-RPC 2.0 to manage tool discovery, conversation history, and request retries. It provides an `MCPClient` class that returns sanitized, TTS-ready text responses and tool call metadata.

Provides a unified interface for streaming audio to Deepgram, AWS Transcribe, or Google Cloud Speech-to-Text via provider-specific classes. It includes built-in utilities for audio format conversion, sample rate resampling, and automatic reconnection with buffering.

Provides a `TwilioMediaStreamHandler` class that manages the bidirectional Twilio Media Streams WebSocket protocol, including audio buffering, barge-in detection, and event lifecycle tracking. It is designed to wrap an existing WebSocket connection to facilitate real-time audio streaming between Twilio and voice AI services.

Provides a unified interface for text-to-speech synthesis across Deepgram, AWS Polly, and Google Cloud, returning audio as an `AsyncIterable<AudioChunk>`. It includes utility classes for Twilio-compatible audio formatting, silence generation, and text chunking to support low-latency streaming voice applications.