LiveKit Agents 1.5 for Telephony in 2026: Native SIP, MCP Tools, Adaptive Interruption

LiveKit Agents shipped 1.0 in April 2025, then iterated quarterly through 1.5 by April 2026. Native SIP support means LiveKit can be your phone-system end-to-end without a Twilio bridge in the middle. Add MCP-compatible tool calls and adaptive interruption handling and it is the default open-source telephony framework for new AI voice projects.

Background

LiveKit started as a WebRTC SFU and pivoted hard into AI voice in 2024. The Agents framework lets any Python or Node.js program join a LiveKit room as a real-time participant. Agents 1.0 (April 2025) standardized the API; 1.x added SIP integration so the room can dial out to a phone number or accept inbound SIP calls; the MCP work in late 2025 made tool calls portable across providers; adaptive interruption (1.4-1.5) tunes barge-in sensitivity per voice and per noise level.

The SIP service is open-source (livekit/sip on GitHub). It accepts inbound trunks, places outbound calls via SIPParticipant, supports DTMF and SIP REFER. Combined with Agents, you write one Python file and have a phone number plus an AI agent.

Architecture

graph LR
    A[PSTN Trunk] --> B[livekit/sip Service]
    B -->|inbound dispatch rule| C[LiveKit Room]
    D[Agent Worker] -->|join room| C
    D --> E[STT plugin]
    E --> F[LLM with MCP tools]
    F --> G[TTS plugin]
    G --> D
    H[Outbound CreateSIPParticipant] --> B
    B --> A

from livekit import agents, rtc
from livekit.plugins import openai, deepgram, cartesia, silero

class IntakeAgent(agents.VoiceAgent):
    def __init__(self):
        super().__init__(
            instructions="You are a healthcare intake assistant.",
            stt=deepgram.STT(),
            llm=openai.LLM(model="gpt-4o-realtime-preview"),
            tts=cartesia.TTS(),
            vad=silero.VAD.load(),
            turn_detection="multilingual",
        )
    @agents.function_tool
    async def book_appointment(self, ctx, date: str, time: str):
        # MCP-compatible tool call
        return await self.tools.call("calendar.book", {"date": date, "time": time})

CallSphere implementation

CallSphere terminates every product on Twilio (Healthcare AI on FastAPI :8084 to OpenAI Realtime, Real Estate AI, Sales Calling AI with 5 concurrent outbound, Salon AI, IT Helpdesk AI, After-Hours AI Twilio simul call+SMS 120-second timeout). 37 agents, 90+ tools, 115+ DB tables, HIPAA + SOC 2, $149/$499/$1499 plans, 14-day trial, 22% affiliate. We selected Twilio over LiveKit SIP for production because Twilio's compliance attestations, number porting, and global reach are still ahead. We track LiveKit because the Telnyx + LiveKit launch in April 2026 changes the math for high-volume tenants. Our reference architecture has a LiveKit-on-Telnyx fallback path that uses our same agent prompts via the Agents 1.5 API.

Build steps

1. Deploy LiveKit server (or use LiveKit Cloud) and the livekit/sip service alongside.
2. Configure SIP inbound trunk: livekit-sip-trunk create --inbound --numbers +18555550100.
3. Configure dispatch rule: every inbound caller goes to a room named after their phone number.
4. pip install livekit-agents livekit-plugins-openai livekit-plugins-deepgram livekit-plugins-cartesia.
5. Write your VoiceAgent class with stt/llm/tts plugins; expose tools as @function_tool.
6. Run the worker: python agent.py dev for local, production for managed.
7. For outbound: SIP outbound trunk plus CreateSIPParticipant API; agent joins the room and dials.

Pitfalls

• Native SIP is GA but rare configs (T.38 fax, video over SIP) are still spotty.
• MCP tool support is per-LLM-provider; double-check OpenAI and Anthropic plugin versions.
• Adaptive interruption requires VAD calibration; the default thresholds work for English clean audio but need tuning for accents and noise.
• LiveKit Cloud and self-hosted have different billing models; agent-minutes vs egress GB.
• The April 2026 Telnyx launch is a separate runtime; agent code is portable but config is not.

FAQ

LiveKit native SIP or stick with Twilio bridge? For greenfield, native SIP is cleaner. For existing Twilio investments, the bridge keeps your numbers and compliance.

MCP support means what? Tool calls are portable: define once, run on OpenAI, Anthropic, Gemini, or any MCP-compatible host.

Adaptive interruption tuning? The vad parameters expose threshold and prefix-padding. For noisy environments raise threshold; for fast talkers shorten prefix.

Self-host or LiveKit Cloud? Cloud for fast deployment under 1k concurrent. Self-host on Kubernetes when you need custom GPU pools or strict region-locking.

Latency? Sub-200 ms voice-to-voice with the LiveKit on Telnyx beta; 600-900 ms typical with external trunks.

Sources

• LiveKit Agents documentation
• LiveKit SIP repository
• LiveKit Build and Deploy 2026 Playbook
• LiveKit on Telnyx launch
• LiveKit Agents on GitHub

Start a 14-day trial of our Twilio-native AI voice, see pricing for $149/$499/$1499, or book a demo to compare LiveKit Agents to our managed stack.