NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
PyPI
PyPI, #1945 most downloaded on PyPI
A powerful framework for building realtime voice AI agents
Latest stable
1.8.11
Last release 17 days ago
05 Nov 2026
Ships on a steady schedule
a new release about every 1 weeks
Nearly every release is documented
notes for 55 of the last 60 stable releases
4 versions withdrawn
withdrawn after publishing
3 years old
230 releases · first in 2023
Drag across the ruler to compare, or tap a release to open it.
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
You can now use the Responses API with your LiveKit Inference credentials, including structured output and web search. REST is the default; opt into W
You can now use the Responses API with your LiveKit Inference credentials, including structured output and web search. REST is the default; opt into WebSocket transport to reuse connections across turns.
from livekit.agents import inference
llm = inference.responses.LLM("openai/gpt-4.1", use_websocket=True)
kalpa-tts-1 streaming TTS in English and Hindi.stt extra.transcribe-stream-1 realtime STT with word timings and recognition context.gpt-realtime-translate through the beta translation interface, with gpt-realtime-whisper input transcription.gemini-3.5-live-translate-preview translation through Gemini API and Vertex AI.reasoning=None to retain the provider default.This release also improves realtime history and reconnection, task handoffs, STT timing, transcript synchronization, VAD updates on Python 3.14, worker cleanup, and multichannel TTS fallback.
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.8.10...livekit-agents@1.8.11
Your coding agent can read these notes before it upgrades. Set up the MCP server →
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
You can use constant-bitrate recording to avoid misleading duration estimates in players such as QuickLook, or match Egress’s 128 kb/s encoding. RecorderIO now lets you set the bitrate and encoding mode; defaults are unchanged.
RecorderIO(agent_session=session, bit_rate=128_000, constant_bit_rate=True)
voxtral-mini-transcribe-3, voxtral-small-transcribe-3, voxtral-mini-transcribe-realtime-3, and voxtral-mini-tts-3, plus 36 TTS voice presets.inworld-tts-2-flash through the plugin and LiveKit Inference.assemblyai and assemblyai/assemblyai-tts-1.api_mode is now ignored; migrate Chat Completions usage to Conversations. For recognize(), explicitly select a batch model; replace removed known IDs voxtral-mini-latest and voxtral-mini-2507. Without VAD or server-side turn detection, standalone realtime streams need stream.flush().mistralai>=3.5.0,<4.0.0; defaults move to Voxtral 3 STT/TTS and mistral-small-latest. Silero is no longer loaded automatically.inworld-tts-2.STT flushing, language overrides, token accounting, translation startup, avatar errors, and noise cancellation are more reliable.
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.8.9...livekit-agents@1.8.10
Inworld: max_connections and idle_connection_timeout are deprecated and ignored; TTS now uses one WebSocket per instance.
You can evaluate named decisions with LiveKit credentials instead of a provider key. inference.DecisionModel batches evaluations and reports refusals per decision without discarding the other answers. Access requires a project feature flag.
model = inference.DecisionModel("openai/gpt-6-luna-decisions")
response = await model.evaluate(
input=chat_ctx,
decisions={"handoff": llm.Probability("The caller requests a human.")},
)
You can attach prompt metadata directly to the provider’s llm_request span from your custom llm_node, without a custom OpenTelemetry processor. Attributes stay scoped to that node’s calls.
async def llm_node(self, chat_ctx, tools, model_settings):
gen_ai.set_llm_request_attributes(
{"langfuse.observation.prompt.name": "support-agent"}
)
async for chunk in Agent.default.llm_node(
self, chat_ctx, tools, model_settings
):
yield chunk
You can record background-tool progress without prompting a spoken response. Final background-tool results can now trigger follow-up tool calls.
await ctx.update("checking the fare rules", silent=True)
bolt.openai/gpt-6-luna-decisions.assemblyai-tts-1: streaming TTS with word timestamps and six languages.persona-2.0-live: speech-timed avatar expressions.drama-3: Inference TTS with free-form expressive steering.simba-3.5: streaming synthesis with timestamps.gemini-3.8-flash: added to supported chat models and selected by default.model; add it if you relied on URL-path detection.voice and voice_id now raises ValueError; choose one.max_connections and idle_connection_timeout are deprecated and ignored; TTS now uses one WebSocket per instance.manual_flush.auto for Opus 5.5, Sonnet 5.5, Fable 5.1 and Mythos 5.1, so tool use is not guaranteed.inworld-tts-2 uses server-side sentence splitting by default; client-side splitting defaults to 250 characters.phonic>=0.32.36, and the framework requires livekit-protocol>=1.1.28.This release also improves reliability across realtime history, tool handoffs, task shutdown, TTS retries, avatar interruptions, worker prewarming, and session reporting.
lk.agent.session messages with one-shot send_bytes instead of stream_bytes by @1egoman in https://github.com/livekit/agents/pull/7751lk agent simulate text by @u9g in https://github.com/livekit/agents/pull/7982Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.8.8...livekit-agents@1.8.9
LiveKit Inference: retired GPT Realtime 1.5/2 and Grok Voice Think Fast 2.0 IDs remain accepted with deprecation warnings. Realtime inference_class is…
You can now run GPT-Live with your LiveKit credentials instead of a separate OpenAI key. inference.DuplexModel handles its protocol, and you can adopt it directly with a model string:
session = AgentSession(llm="openai/gpt-live-1")
google/gemini-3.8-live, google/gemini-3.8-live-extended-thinking, and google/gemini-live.jev-1.13.0 and OpenRouter’s Decisions API for background probability, choice, and score evaluations without delaying replies.inference.RealtimeModel no longer subclasses the plugin’s openai.realtime.RealtimeModel; update isinstance checks. Incidental wildcard exports such as realtime_model.logger are no longer available.cedar replaces the unsupported aster voice; update voice selections.inference_class is deprecated and no longer sent.XML-aware TTS chunking, realtime startup retries, turn-audio flushing, and interruption handling are more reliable.
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.8.7...livekit-agents@1.8.8
You can now measure realtime replies from the last detected user speech to first assistant playback when you configure local VAD. Read the latency fro
You can now measure realtime replies from the last detected user speech to first assistant playback when you configure local VAD. Read the latency from assistant message metrics:
@session.on("conversation_item_added")
def on_message(event):
if event.item.type == "message" and event.item.role == "assistant":
print(event.item.metrics.get("e2e_latency"))
lk.fallback.label and lk.fallback.index move to adapter request spans; update trace queries for the flattened hierarchy.httpx2, not httpx.LogRecord(event_name=...); custom Logger subclasses must implement enabled().openai>=3.19,<4, mcp>=2.2,<3, websockets>=17.1,<17.2, and OpenTelemetry >=1.45,<1.46; Anthropic requires >=1.8,<2, and turn detection requires onnxruntime>=1.30,<1.31.Recording shutdown, VAD resizing, interruption teardown, chat summaries, structured outputs, multilingual sentence splitting, and tool discovery are more reliable.
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.8.6...livekit-agents@1.8.7
fix(sarvam): close the streams the STT collects by @dhruvladia-sarvam in https://github.com/livekit/agents/pull/7058
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.8.5...livekit-agents@1.8.6
fix: support gemini-3.8-live and warn on api mismatches by @chenghao-mou in https://github.com/livekit/agents/pull/7587
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.8.4...livekit-agents@1.8.5
feat(smallestai): add TTS continuations protocol support by @harshitajain165 in https://github.com/livekit/agents/pull/7111
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.8.3...livekit-agents@1.8.4
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.8.2...livekit-agents@1.8.3
Run loop blocking detection. The agent now detects code that blocks the asyncio event loop and surfaces the offending call in Agent Insights, so you c
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.8.1...livekit-agents@1.8.2
Speech models that are able to speak and listen simultaneously are now supported via the new DuplexModel class, with OpenAI's GPT-Live model as the fi
Speech models that are able to speak and listen simultaneously are now supported via the new DuplexModel class, with OpenAI's GPT-Live model as the first to be implemented.
session = AgentSession(
llm=GPTLiveModel(
voice="marin",
# backend Responses model that handles reasoning and tools
responses_options={
"model": "gpt-5.6-luna",
"instructions": "Use tools when current information is required.",
},
),
)
Read more about duplex models in our docs.
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.8.0...livekit-agents@1.8.1
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
Spans, metrics, and logs now follow the OTel GenAI semantic conventions, so Langfuse, Datadog Agent Observability, and other GenAI-aware backends read LiveKit traces natively. PII filtering also moved in-process.
Migration Notes
| Before | After |
|---|---|
event gen_ai.system.message |
attribute gen_ai.system_instructions |
events gen_ai.user.message, gen_ai.assistant.message, gen_ai.tool.message |
attribute gen_ai.input.messages |
event gen_ai.choice (role/content/tool_calls) |
attributes gen_ai.output.messages + gen_ai.response.finish_reasons |
realtime gen_ai.* on the agent_turn span |
on the new realtime_inference child span |
realtime gen_ai.operation.name = "chat" |
= "generate_content" |
gen_ai.provider.name = "api.openai.com", "AWS Bedrock", "google" |
"openai", "aws.bedrock", "gcp.gen_ai" |
custom llm_node always tagged with the configured model/provider |
tagged only when that LLM served the call; otherwise the node span records the convention itself |
| PII stripped at LiveKit Cloud's collector | stripped in-process; under project redaction content never reaches any exporter, Cloud included |
telemetry._chat_ctx_to_otel_events(chat_ctx) |
telemetry.gen_ai.to_system_instructions() / to_input_messages() / to_output_messages() |
telemetry.utils._redaction_enabled() |
telemetry.utils.redaction_enabled(span_attributes=None) |
Breaking
gen_ai.system.message,
gen_ai.user.message, gen_ai.assistant.message, gen_ai.tool.message and
gen_ai.choice are gone, replaced by the attributes in the table above (JSON,
part-based format). Dashboards and processors reading the events see nothing.agent_turn onto a new realtime_inference child span,
so queries reading realtime token usage from agent_turn return nothing. Its
gen_ai.operation.name also changed from chat to generate_content.gen_ai.provider.name is normalized to the registry spelling. Providers outside the
registry are unchanged. Update any group-by on the old values.llm_node is no longer credited with the configured model and provider
unless that LLM actually served the request; a node calling a third-party engine is
left unattributed and records the convention on its own span instead.gen_ai.usage.input_cached_tokens is emitted again on the pipeline path (reversing
#6852) and reasoning tokens go out under both gen_ai.usage.reasoning.output_tokens
and gen_ai.usage.reasoning_tokens. A backend summing both spellings double-counts.telemetry._chat_ctx_to_otel_events() removed; telemetry.utils._redaction_enabled()
is now redaction_enabled(span_attributes=None).Added
gen_ai.request.stream, gen_ai.output.type,
gen_ai.conversation.id, gen_ai.response.{id,model,finish_reasons,time_to_first_chunk}.gen_ai.usage.cache_write.input_tokens, .reasoning.output_tokens
and the per-modality .{text,audio,image}.* counts.execute_tool spans carry gen_ai.tool.{name,type,call.id,call.arguments,call.result,description};
turns carry gen_ai.agent.name, sessions gen_ai.workflow.name.error.type on failures, set even when redaction is on.gen_ai.client.token.usage, gen_ai.client.operation.duration,
gen_ai.client.operation.time_to_first_chunk, gen_ai.invoke_agent.duration,
gen_ai.execute_tool.duration, alongside the existing lk.agents.*.OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT=0 (or
telemetry.gen_ai.set_capture_content(False)) omits message payloads, tool
definitions and tool arguments/results. Content capture is on by default.set_tracer_provider(..., allow_pii=False) / LIVEKIT_TELEMETRY_ALLOW_PII=0 strips
conversational content in-process before third-party exporters. PII is allowed by
default.Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.7.1...livekit-agents@1.8.0
fix(voice): agent and user state while a tool runs by @longcw in https://github.com/livekit/agents/pull/6937
livekit to 1.1.15 by @1egoman in https://github.com/livekit/agents/pull/6970Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.7.0...livekit-agents@1.7.1
Introducing PII Redaction Support
This release adds PII redaction support for Agent Observability. It semantically redacts detected entities from chat history and audio recordings. It also filters sensitive data from logs and traces while retaining useful diagnostic context.
To support this feature, we renamed sensitive trace attributes and log fields emitted by the framework. If your third-party observability queries depend on these names, update them when you upgrade.
Expressive mode allows voice agents to speak with natural prosody and emotion. Speech delivery is determined by emotion tags generated from the context of the conversation. Enable expressive mode in your AgentSession:
from livekit.agents import AgentSession, inference
session = AgentSession(
stt=inference.STT(model="deepgram/nova-3", language="multi"),
llm=inference.LLM(model="google/gemma-4-31b-it"),
tts=inference.TTS(model="fishaudio/s2.1-pro", voice="b347db033a6549378b48d00acb0d06cd"),
expressive=True,
# ... vad, turn_detection
)See the documentation and the blog post for more information.
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.6.10...livekit-agents@1.7.0
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
fix(cartesia): redact API keys from websocket handshake errors by @LHMQ878 in #6740
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.6.9...livekit-agents@1.6.10
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.6.9...livekit-agents@1.6.10
Add support for Deepgram's Flux TTS API (v2/speak) integration by @dg-edcharbeneau in https://github.com/livekit/agents/pull/6511
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.6.8...livekit-agents@1.6.9
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.6.8...livekit-agents@1.6.9
> `console` and `dev` modes are deprecated in favor of lk agent (#6656). Existing workflows that shell out to python agent.py console or python agent.…
[!WARNING]
consoleanddevmodes are deprecated in favor oflk agent(#6656). Existing workflows that shell out topython agent.py consoleorpython agent.py devshould migrate to thelk agentCLI.
[!IMPORTANT] Inference Ink-2 speech onset fix (#6630) — if you rely on the STT model for interruption detection without a local VAD, upgrade for this fix. Previously the provider's speech-onset signal was delayed until first non-empty interim transcript, so STT interruptions could be delayed on Ink-2 via Inference. Setups with a local VAD in the pipeline are unaffected.
lk agent by @u9g in https://github.com/livekit/agents/pull/6656Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.6.7...livekit-agents@1.6.8
Warning
console and dev modes are deprecated in favor of lk agent (#6656). Existing workflows that shell out to python agent.py console or python agent.py dev should migrate to the lk agent CLI.
Important
Inference Ink-2 speech onset fix (#6630) — if you rely on the STT model for interruption detection without a local VAD, upgrade for this fix. Previously the provider's speech-onset signal was delayed until first non-empty interim transcript, so STT interruptions could be delayed on Ink-2 via Inference. Setups with a local VAD in the pipeline are unaffected.
lk agent by @u9g in #6656Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.6.7...livekit-agents@1.6.8
fix(core): reset user away timer on final STT transcript AGT-3149 by @chenghao-mou in https://github.com/livekit/agents/pull/6478
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.6.6...livekit-agents@1.6.7
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.6.6...livekit-agents@1.6.7
elevenlabs: deprecate STT model_id in favor of model by @u9g in https://github.com/livekit/agents/pull/6378
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.6.5...livekit-agents@1.6.6
fix(assemblyai): respect mode preset for turn-silence; remap deprecated u3-pro by @dlange-aai in https://github.com/livekit/agents/pull/6220
mode preset for turn-silence; remap deprecated u3-pro by @dlange-aai in https://github.com/livekit/agents/pull/6220remove() method to ChatContext for flexible item removal by @3eid in https://github.com/livekit/agents/pull/5131Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.6.4...livekit-agents@1.6.5
mode preset for turn-silence; remap deprecated u3-pro by @dlange-aai in #6220remove() method to ChatContext for flexible item removal by @3eid in #5131Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.6.4...livekit-agents@1.6.5
disable retry for eot errors by @chenghao-mou in https://github.com/livekit/agents/pull/6195
Turn Detector
Turn handling/committing
[!WARNING] users are advised to upgrade if you use agent handoff with STT and are on 1.5.14 ~ 1.6.3
Core
Plugins
language_code doc to remove error case by @inickt in https://github.com/livekit/agents/pull/6197Version update
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.6.3...livekit-agents@1.6.4
fix(slng): expose speed in update_options by @adityajha2005 in https://github.com/livekit/agents/pull/6175
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.6.2...livekit-agents@1.6.3
feat(assemblyai): add universal-3-5-pro (now default) and Voice Focus streaming params by @dlange-aai in https://github.com/livekit/agents/pull/6119
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.6.1...livekit-agents@1.6.2
Determining when the agent should respond is a delicate balance: responding too early can interrupt the user, while responding too late introduces unn
Determining when the agent should respond is a delicate balance: responding too early can interrupt the user, while responding too late introduces unnecessary latency and awkward silence. We solve this with our state-of-the-art turn detector, which leverages both audio and text semantics to identify the optimal moment to respond.
Read more about our turn detector in our blog.
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.6.0...livekit-agents@1.6.1
When a long-running tool is in progress, the user will be met with silence until it completes. Asynchronous tools can hand control back to the LLM bef
When a long-running tool is in progress, the user will be met with silence until it completes. Asynchronous tools can hand control back to the LLM before it finishes, streaming updates into the conversation as it progresses.
class TravelAgent(Agent):
@llm.function_tool(flags=llm.ToolFlag.CANCELLABLE, on_duplicate="confirm")
async def book_flight(self, ctx: RunContext, origin: str, destination: str, date: str) -> str:
"""Book a flight."""
# First update is delivered immediately and releases control back to the LLM,
# so the agent can say e.g. "Sure, searching flights to Tokyo — this'll take a minute."
await ctx.update(
f"Searching flights from {origin} to {destination} on {date}. "
"This will take a couple of minutes."
)
await asyncio.sleep(30) # simulate long-running work
# Later updates are coalesced into a deferred reply, delivered when the agent
# is idle: "Good news — best price is $289 on Delta, confirming now."
await ctx.update("Found 3 options. Best price $289 on Delta. Confirming now.")
await asyncio.sleep(40) # simulate more long-running work
# The final return is also delivered when the agent is idle.
return f"Booked! Confirmation FL-{random.randint(100000, 999999)}."
Use filler words to break long silences via ctx.with_filler(). You can set delay to the amount of quiet seconds per interval for when the filler words are played. You can also rotate between phrases, set interval to configure the cooldown between filler fires.
followups = [
"Almost there, just confirming.",
"Still working on it, won't be long.",
"Hang tight — almost done.",
]
async with ctx.with_filler(
lambda step: followups[step], delay=5, interval=10, max_steps=len(followups)
):
await asyncio.sleep(40) # simulate a long API call
Read more about async tools in our docs.
pytest --unit/--plugin openai/ … flags by @Bobronium in https://github.com/livekit/agents/pull/5945Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.5.17...livekit-agents@1.6.0
Nothing published for this version
Nothing published for this version
Nothing published for this version
Nothing published for this version
feat(gradium STT): add language option to gradium.STT by @FLoppix in https://github.com/livekit/agents/pull/5878
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.5.15...livekit-agents@1.5.17
Nothing published for this version
(openai realtime): add status_details to incomplete response logs by @tinalenguyen in https://github.com/livekit/agents/pull/5873
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.5.14...livekit-agents@1.5.15
Update download-files deprecation message by @bcherry in https://github.com/livekit/agents/pull/5781
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.5.12...livekit-agents@1.5.14
Nothing published for this version
deprecate mcp_servers param on Agent and AgentSession by @longcw in https://github.com/livekit/agents/pull/5667
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.5.10...livekit-agents@1.5.12
Nothing published for this version
fix: surface Deepgram TTS websocket errors by @nightcityblade in https://github.com/livekit/agents/pull/5728
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.5.9...livekit-agents@1.5.10
An outbound call can reach a person, voicemail, an IVR menu, or a number that can't accept messages. Answering machine detection (AMD) listens to the
An outbound call can reach a person, voicemail, an IVR menu, or a number that can't accept messages. Answering machine detection (AMD) listens to the start of the call, classifies it with an LLM, and returns a result so your agent can respond appropriately.
Read more about using AMD in our blog.
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.5.8...livekit-agents@1.5.9
feat(interruption): barge-in cooldown window for corrections by @chenghao-mou in https://github.com/livekit/agents/pull/5269
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.5.7...livekit-agents@1.5.8
fix(openai): forward session.update on RealtimeModel.update_options by @longcw in https://github.com/livekit/agents/pull/5531
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.5.6...livekit-agents@1.5.7
Add Qwen 3 TTS support for Simplismart-livekit plugin by @simplipratik in https://github.com/livekit/agents/pull/5474
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.5.5...livekit-agents@1.5.6
(hedra): note deprecation in readme by @tinalenguyen in https://github.com/livekit/agents/pull/5475
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.5.4...livekit-agents@1.5.5
Refines default behavior for preemptive generation to better handle long or intermittent user speech, reducing unnecessary downstream inference and as
Refines default behavior for preemptive generation to better handle long or intermittent user speech, reducing unnecessary downstream inference and associated cost increases.
Also introduces PreemptiveGenerationOptions for developers who need fine-grained control over this behavior.
https://github.com/livekit/agents/blob/78a66bcf79c5cea82989401c408f1dff4b961a5b/livekit-agents/livekit/agents/voice/turn.py#L115
class PreemptiveGenerationOptions(TypedDict, total=False):
"""Configuration for preemptive generation."""
enabled: bool
"""Whether preemptive generation is enabled. Defaults to ``True``."""
preemptive_tts: bool
"""Whether to also run TTS preemptively before the turn is confirmed.
When ``False`` (default), only LLM runs preemptively; TTS starts once the
turn is confirmed and the speech is scheduled."""
max_speech_duration: float
"""Maximum user speech duration (s) for which preemptive generation
is attempted. Beyond this threshold, preemptive generation is skipped
since long utterances are more likely to change and users may expect
slower responses. Defaults to ``10.0``."""
max_retries: int
"""Maximum number of preemptive generation attempts per user turn.
The counter resets when the turn completes. Defaults to ``3``."""
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.5.3...livekit-agents@1.5.4
> livekit-agents 1.5 introduced many new features. You can check out the changelog [here](https://github.com/livekit/agents/releases/tag/livekit-agent
[!NOTE]
livekit-agents 1.5 introduced many new features. You can check out the changelog here.
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.5.2...livekit-agents@1.5.3
fix(utils): preserve type annotations in deprecate_params by @longcw in https://github.com/livekit/agents/pull/5200
[!NOTE]
livekit-agents 1.5 introduced many new features. You can check out the changelog here.
generate_reply timeout to 10 seconds by @qionghuang6 in https://github.com/livekit/agents/pull/5205min_words_to_interrupt to Phonic plugin options by @qionghuang6 in https://github.com/livekit/agents/pull/5304Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.5.1...livekit-agents@1.5.2
> livekit-agents 1.5 introduced many new features. You can check out the changelog [here](https://github.com/livekit/agents/releases/tag/livekit-agent
[!NOTE]
livekit-agents 1.5 introduced many new features. You can check out the changelog here.
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.5.0...livekit-agents@1.5.1
Old keyword arguments (min_endpointing_delay, allow_interruptions, etc.) still work but are deprecated and will emit warnings.
Preemptive generation starts LLM and TTS inference before the end of a user’s turn is detected, reducing overall latency.
To disable it:
session = AgentSession(preemptive_generation=False)
The headline feature of v1.5.0: an audio-based ML model that distinguishes genuine user interruptions from incidental sounds like backchannels ("mm-hmm"), coughs, sighs, or background noise. Enabled by default — no configuration needed.
Key stats:
When a false interruption is detected, the agent automatically resumes playback from where it left off — no re-generation needed.
To opt out and use VAD-only interruption:
session = AgentSession(
...
turn_handling=TurnHandlingOptions(
interruption={
"mode": "vad",
},
),
)
Blog post: https://livekit.com/blog/adaptive-interruption-handling
Endpointing delays now adapt to each conversation's natural rhythm. Instead of a fixed silence threshold, the agent uses an exponential moving average of pause durations to dynamically adjust when it considers the user's turn complete.
session = AgentSession(
...
turn_handling=TurnHandlingOptions(
endpointing={
"mode": "dynamic",
"min_delay": 0.3,
"max_delay": 3.0,
},
),
)
TurnHandlingOptions APIEndpointing and interruption settings are now consolidated into a single TurnHandlingOptions dict passed to AgentSession. Old keyword arguments (min_endpointing_delay, allow_interruptions, etc.) still work but are deprecated and will emit warnings.
session = AgentSession(
turn_handling={
"turn_detection": "vad",
"endpointing": {"min_delay": 0.5, "max_delay": 3.0},
"interruption": {"enabled": True, "mode": "adaptive"},
},
)
New SessionUsageUpdatedEvent provides structured, per-model usage data — token counts, character counts, and audio durations — broken down by provider and model:
@session.on("session_usage_updated")
def on_usage(ev: SessionUsageUpdatedEvent):
for usage in ev.usage.model_usage:
print(f"{usage.provider}/{usage.model}: {usage}")
Usage types: LLMModelUsage, TTSModelUsage, STTModelUsage, InterruptionModelUsage.
You can also access aggregated usage at any time via the session.usage property:
usage = session.usage
for model_usage in usage.model_usage:
print(model_usage)
Usage data is also included in SessionReport (via model_usage), so it's available in post-session telemetry and reporting out of the box.
ChatMessage.metricsEach ChatMessage now carries a metrics field (MetricsReport) with per-turn latency data:
transcription_delay — time to obtain transcript after end of speechend_of_turn_delay — time between end of speech and turn decisionon_user_turn_completed_delay — time in the developer callbackContext summarization now includes function calls and their outputs when building summaries, preserving tool-use context across the conversation window.
Set the agent log level via LIVEKIT_LOG_LEVEL environment variable or through ServerOptions, without touching your code.
| Deprecated | Replacement | Notes |
|---|---|---|
metrics_collected event |
session_usage_updated event + ChatMessage.metrics |
Usage/cost data moves to session_usage_updated; per-turn latency moves to ChatMessage.metrics. Old listeners still work with a deprecation warning. |
UsageCollector |
ModelUsageCollector |
New collector supports per-model/provider breakdown |
UsageSummary |
LLMModelUsage, TTSModelUsage, STTModelUsage |
Typed per-service usage classes |
RealtimeModelBeta |
RealtimeModel |
Beta API removed |
AgentFalseInterruptionEvent.message / .extra_instructions |
Automatic resume via adaptive interruption | Accessing these fields logs a deprecation warning |
AgentSession kwargs: min_endpointing_delay, max_endpointing_delay, allow_interruptions, discard_audio_if_uninterruptible, min_interruption_duration, min_interruption_words, turn_detection, false_interruption_timeout, resume_false_interruption |
turn_handling=TurnHandlingOptions(...) |
Old kwargs still work but emit deprecation warnings. Will be removed in v2.0. |
Agent / AgentTask kwargs: turn_detection, min_endpointing_delay, max_endpointing_delay, allow_interruptions |
turn_handling=TurnHandlingOptions(...) |
Same migration path as AgentSession. Will be removed in future versions. |
generate_reply to resolve with the current GenerationCreatedEvent by @qionghuang6 in https://github.com/livekit/agents/pull/5147Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.4.6...livekit-agents@1.5.0
Nothing published for this version
Nothing published for this version
(google realtime): replace deprecated mediaChunks by @tinalenguyen in https://github.com/livekit/agents/pull/5089
null in enum array for nullable enum schemas by @MSameerAbbas in https://github.com/livekit/agents/pull/5080required field in tool schema when function has no parameters by @longcw in https://github.com/livekit/agents/pull/5082Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.4.5...livekit-agents@1.4.6
fix(mcp): replace deprecated streamablehttp_client with streamable_http_client by @longcw in https://github.com/livekit/agents/pull/5048
generate_reply and update_chat_ctx support to Phonic Plugin by @qionghuang6 in https://github.com/livekit/agents/pull/5058Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.4.4...livekit-agents@1.4.5
Upgrading Cartesia TTS default to Sonic 3 by @chongzluong in https://github.com/livekit/agents/pull/4922
agent_worker.py by @kradkfl in https://github.com/livekit/agents/pull/4970Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.4.3...livekit-agents@1.4.4
chore: remove models to be deprecated on March 19 2026 by @chenghao-mou in https://github.com/livekit/agents/pull/4895
Full Changelog: https://github.com/livekit/agents/compare/browser-v0.1.4...livekit-agents@1.4.3
Stability-focused release with significant reliability improvements. Fixes multiple memory leaks in the process pool — job counter leaks on cancellati
Stability-focused release with significant reliability improvements. Fixes multiple memory leaks in the process pool — job counter leaks on cancellation, pending assignment leaks on timeout, socket leaks on startup failure, and orphaned executors on send failure. IPC pipeline reliability has been improved, and several edge-case hangs have been resolved (participant never joining, Ctrl+C propagation to child processes). STT/TTS fallback behavior is now more robust: STT fallback correctly skips the main stream during recovery, and TTS fallback no longer shares resamplers across streams. Other fixes include ChatContext.truncate no longer dropping developer messages, correct cgroups v2 CPU quota parsing, proper on_session_end callback ordering, and log uploads even when sessions fail to start. Workers now automatically reject jobs when draining or full, and the proc pool correctly spawns processes under high load.
RecordingOptions APIThe record parameter on AgentSession.start() now accepts granular options in addition to bool. All keys default to True when omitted.
# record everything (default)
await session.start(agent, record=True)
# record nothing
await session.start(agent, record=False)
# granular: record audio but disable traces, logs, and transcript
await session.start(agent, record={"audio": True, "traces": False, "logs": False, "transcript": False})
Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.4.0...livekit-agents@1.4.2
Nothing published for this version
This release adds Python 3.14 support and drops Python 3.9. The minimum supported version is now Python 3.10.
This release adds Python 3.14 support and drops Python 3.9. The minimum supported version is now Python 3.10.
Tools and toolsets now have stable unique IDs, making it possible to reference and filter tools programmatically. Changes to agent configuration (instructions, tools) are now tracked in conversation history via AgentConfigUpdate.
LLMStream.collect() APIA new LLMStream.collect() API makes it significantly easier to use LLMs outside of AgentSession. You can now call an LLM, collect the full response, and execute tool calls with a straightforward API — useful for background tasks, pre-processing, or any workflow where you need LLM capabilities without the full voice agent pipeline.
from livekit.agents import llm
response = await my_llm.chat(chat_ctx=ctx, tools=tools).collect()
for tc in response.tool_calls:
result = await llm.execute_function_call(tc, tool_ctx)
ctx.insert(result.fnc_call)
if result.fnc_call_out:
ctx.insert(result.fnc_call_out)
Realtime models now support commit_user_turn, enabling turn_detection="manual" mode. This gives you full control over when user turns are committed — useful for push-to-talk interfaces or scenarios where automatic VAD-based turn detection isn't ideal.
@ctx.room.local_participant.register_rpc_method("end_turn")
async def end_turn(data: rtc.RpcInvocationData):
session.input.set_audio_enabled(False)
session.commit_user_turn(
transcript_timeout=10.0,
stt_flush_duration=2.0,
)
When the agent server temporarily loses connection and reconnects, active jobs are now automatically migrated rather than being dropped. This significantly improves reliability during transient network issues.
Fixed a bug where late end-of-speech events could trigger duplicate false interruption timers, causing the agent to incorrectly stop speaking. The agent now properly deduplicates these events and tracks STT completion state more reliably.
xai.responses.LLM()azure.responses.LLM(), with support for deployments and Azure authid to tools by @theomonnom in https://github.com/livekit/agents/pull/4653AgentConfigUpdate & initial judges by @theomonnom in https://github.com/livekit/agents/pull/4547LLMStream.collect API & external easier tool executions by @theomonnom in https://github.com/livekit/agents/pull/4680Full Changelog: https://github.com/livekit/agents/compare/livekit-agents@1.3.12...livekit-agents@1.4.0
Nothing published for this version
Nothing published for this version
One column per quarter.