Skip to content
Agents tracked: 258 Downloads (7d): 219M up 6.1% GitHub stars: 5.5M VS Code installs: 148M Releases (7d): 327 Agent status: 1 with issues Updated Oct 7, 2026
AgentGid

LiveKit Agents changelog: what's new each month

Every stable LiveKit Agents release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 3 months covered; the current month updates daily.

October 2026 · so far

2 releases: 1.8.4 → 1.8.5

October brought expanded AI model support and improvements to multi-turn conversations and workflow reliability. The release added new TTS and speech recognition capabilities while fixing edge cases in inference handling and cross-platform compatibility.

Highlights

  • Azure Voice Live Realtime API support for real-time speech interaction.
  • Sticky LLM fallback routing to improve model selection reliability.
  • STT event metadata recording on user turns for better conversation tracking.
  • Support for Gemini 3.8-live and GPT-5.x model variants.
  • Unicode log preservation on Windows for improved debugging.

New

  • TTS continuations protocol support via SmallestAI.
  • Azure Voice Live Realtime API integration.
  • STT event metadata recording on user turns.
  • Sticky LLM fallback routing for automatic model fallback.
  • Inline mid-conversation instructions for Qwen/Qwen3.8-27B on Baseten.

Improved

  • Gemini 3.8-live support with API mismatch warnings.
  • Qwen/Qwen3.8-27B handling with inline mid-conversation instructions on Baseten.
  • Call ID forwarding with tool responses on Vertex AI.
  • Non-ASCII text readability in MCP results and task group tool descriptions.
  • Unicode log preservation on Windows CP1252 systems.
  • Twilio CallToken forwarding during warm transfers in workflows.

Fixed

  • STT transcript draining before closing to prevent data loss.
  • Twilio CallToken forwarding during warm transfers.
  • API mismatch handling for Gemini models.
  • Call ID inclusion with Vertex AI tool responses.
  • Unicode handling in logs on Windows systems.
  • Non-ASCII character preservation in MCP results.

September 2026

4 releases: 1.8.0 → 1.8.3

September brought refinements to telemetry and tracing, performance improvements to agent startup and lifecycle tracking, and fixes across multiple integrations including OpenAI, Deepgram, and voice handling.

Action needed
  • Telemetry.chatctxtootelevents() removed; use telemetry.utils.redactionenabled() instead.
  • Conversation content no longer emitted as span events (genai.system.message, etc.).
  • Custom LLM nodes no longer credited with configured model and provider in telemetry.

Highlights

  • Run loop blocking detection now alerts when agent code blocks the event loop.
  • Faster agent startup through prefork server library preloading.
  • Improved full-lifecycle agent tracing covers the entire agent lifecycle.
  • Telemetry changes normalize provider names and reorganize span structure for better observability.
  • Multiple fixes to TTS, STT, and LLM integrations across providers.

New

  • Run loop blocking detection for identifying code that blocks the agent event loop.
  • Prefork server library preloading to speed up agent startup.
  • Speechify-Caller-Version attribution header support.
  • Participant linking information logged for RoomIO voice connections.
  • Full agent lifecycle tracing coverage.

Improved

  • Agent startup performance via library preloading in prefork server.
  • Agent lifecycle tracing now covers the complete lifecycle.
  • Telemetry reorganized with realtime inference moved to dedicated child span.
  • Genai provider name normalization to registry spelling.
  • Content attributes stripped in-process when project mandates redaction.
  • Request and response attributes expanded with stream and output type information.

Fixed

  • OpenAI TTS prewarm no longer replaced while in flight.
  • Unused adapter metrics streams now properly drained.
  • Unused audio forwarding buffer removed.
  • Mid-conversation instruction positioning fixed for Gemini gateway and Mistral.
  • DuplexModel startup refused when under text simulation.
  • LiveAvatar now honors agent.stateupdated for speaking and barge-in.

August 2026

5 releases: 1.6.8 → 1.7.1

August brought refinements across speech recognition and text-to-speech integrations, with better handling of interruptions, error recovery, and provider-specific configuration options. The month focused on stability improvements and expanded support for real-time voice processing.

Action needed
  • Console and dev modes in CLI are deprecated in favor of lk agent command.

Highlights

  • Added Deepgram Flux TTS and Bland TTS plugin support for more voice synthesis options.
  • Improved interrupt handling to respect speech settings and prevent crashes during tool execution.
  • Enhanced real-time STT configuration exposure across ElevenLabs, X.AI, Cartesia, and Smallest AI providers.
  • Fixed fallback adapter reporting and isolated VAD inference for better error recovery.
  • Deprecated console and dev modes in CLI in favor of the unified lk agent command.

New

  • Deepgram Flux TTS API (v2/speak) integration support.
  • Bland TTS plugin.
  • User transcription timeout configuration.
  • Speechify streaming with timestamps via /v1/audio/stream/with-timestamps.
  • Smallest AI STT now supports endpointing, keyword boosting, format, and sentence timestamps.
  • Agent and user state tracking while tools are running.
  • X.AI STT exposes vadthreshold, smartturn, smartturntimeout, and keyterm parameters.
  • Cartesia STT exposes ink-2 turn-detection thresholds and keyterm parameters.

Improved

  • Auto-disable realtime server-side turn detection in voice module.
  • ElevenLabs realtime STT now supports previoustext parameter.
  • Phonic plugin exposes more realtime configuration options.
  • Each TTS fallback recovery path now gets its own task slot.
  • Non-retryable STT errors are now properly respected.
  • Fallback adapters now report the active instance's model and provider.
  • OpenAI realtime keeps derived capabilities in sync with update options.
  • Google realtime now keeps unspoken model text out of the transcript.

Fixed

  • Cartesia API keys are now redacted from websocket handshake errors.
  • Session no longer crashes on null usage token counts.
  • Interrupt walk now stops at queued speech that disallows interruptions.
  • No longer raises errors when interrupting cancelled or finished speech.
  • Duplicate agent speech end events eliminated.
  • ElevenLabs keyterms now sent on realtime STT WebSocket.

Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. LiveKit Agents: pricing, features and alternatives · All changelogs