Bodhi Realtime Agent FrameworkBuild Real-Time Voice Agents

TypeScript framework for voice AI applications with Gemini Live, OpenAI Realtime, custom TTS, subagents, and telephony bridges

Get Started

API Reference

GitHub

Real-Time Voice

Bidirectional audio streaming via a provider-agnostic LLMTransport interface. Supports Google Gemini Live and OpenAI Realtime APIs with server-side turn detection, pluggable STT providers, and normalized usage events.

Multi-Agent

Define multiple agents with distinct personas and tool sets. Transfer between them mid-conversation with automatic context preservation.

Function Tools

Inline (blocking) and background (non-blocking) tool execution. Zod schema validation, timeout, cancellation, and abort signals. Artifact pipeline for cross-tool data flow.

Persistent Agents

Relay subagent pattern for bridging to stateful external agents like Claude Code and OpenClaw. Session routing, concurrent task isolation, and write-lock serialization.

Memory

Automatic extraction and persistence of durable user facts across sessions. LLM-powered distillation with configurable triggers.

Observability

Type-safe EventBus and lifecycle hooks for logging, metrics, and debugging. Zero overhead when unattached.

Multimodal

Voice, text input, file upload, image generation, and artifact attachments on a single WebSocket connection. Mixed-mode interaction out of the box.

Custom Speech

Swap native model audio for streaming TTS providers such as Cartesia or ElevenLabs, with sentence buffering, barge-in cancellation, word timing, and automatic resampling.

Telephony

Bridge browser sessions to Twilio phone calls, accept inbound phone callers, and transfer between AI audio and external human audio while preserving session context.