v0.7.0 — Voice Picker & Deployment Improvements
Released April 9, 2026 A release focused on making the builder more interactive and deployments more reliable. TTS nodes now let you preview voices directly on the canvas, LiveKit Inference is available as a managed deployment option, API keys are now split between builder and agent, and the Cloud Run worker received a round of stability fixes.✨ What’s New
🎙️ Dynamic Voice Picker with Preview
TTS nodes now include an interactive voice picker — browse available voices for your selected provider and click the play button to hear a preview before committing. The picker loads voices dynamically based on the provider and selected model, so you always see exactly what’s available.☁️ LiveKit Inference Deployment
When deploying to LiveKit Cloud, you can now choose LiveKit Inference as the compute backend. This offloads model execution to LiveKit’s managed infrastructure instead of your own Python worker, reducing operational overhead for supported providers.🔑 Separate Builder & Agent API Keys
API keys are now split into two independent sections:- Builder keys — used by the AI assistant (flow generation and editing) inside the canvas
- Agent keys — used by the deployed agent at runtime for LLM, TTS, and STT providers
🔌 SIP Transfer: Disconnect After Transfer
The SIP Transfer node has a new disconnect after transfer toggle. When enabled, the agent disconnects from the call immediately after the transfer is initiated, rather than staying connected until the transfer completes.⚡ Improvements
Simplified Interruption Mode
The 3-mode interruption selector (Off / VAD / Adaptive) has been simplified back to a plain Off / On toggle. The toggle maps to VAD-based interruption when enabled. Realtime agents no longer expose an interruption setting (it is controlled by the provider).Cloud Run Worker Stability
Several fixes were applied to the agent Python worker running on Cloud Run:- SSE keepalive — an SSE ping is sent periodically to prevent Cloud Run from scaling to zero mid-session
- HuggingFace offline mode — the worker now forces HF offline mode so model files are loaded from the container image cache instead of attempting network downloads
- Resource allocation — memory increased to 4Gi and CPU to 2 cores to support the turn detector model reliably
- gVisor compatibility — model cache is now copied rather than symlinked, which is required for Cloud Run’s gVisor sandbox
- Startup reliability — spawn and fetch timeouts increased to 60s; room metadata is retried on warm dispatches