September 2026 Summaries

LiveKit Private Links enable hosted AI agents to securely access sensitive internal data and services through cloud-provider private networking, allowing enterprises in regulated industries to support more complex, high-value workflows without exposing resources to the public internet or maintaining VPNs, public IPs, or inbound firewall rules. The feature creates a managed encrypted WireGuard tunnel between an agent and a LiveKit gateway in the relevant cloud region, then routes traffic through AWS PrivateLink or Azure Private Link to services behind private infrastructure; access is limited to TLS-secured TCP connections and does not support insecure TCP or general UDP. Organizations can configure links through the LiveKit CLI or dashboard by creating and approving endpoint connections in AWS or Azure, then using a generated private `.link` hostname from an agent deployed in the same LiveKit region. Private Links currently support US East and EU Central LiveKit deployments, have region-specific AWS and Azure availability, cost $50 per link plus $0.10 per GB transferred, and are initially limited to three links per project per region.

LiveKit Agents and Pipecat are presented as the leading open-source frameworks for production voice and video AI agents, both supporting standard speech-to-text, language-model, and text-to-speech workflows while differing primarily in abstraction level and infrastructure assumptions. LiveKit Agents is tightly integrated with LiveKit’s open-source WebRTC and SIP stack, providing prewired sessions, turn handling, scaling, telephony, observability, multi-platform SDKs, and managed cloud options, making it suited to teams prioritizing rapid deployment, enterprise compliance, and global production operations. Pipecat is a Python-only, transport-agnostic framework that exposes an explicit, customizable pipeline, giving developers greater flexibility to add, reorder, fork, or build non-conversational audio and video processing flows, but requiring them to manage more media, transport, pipeline, and provider complexity. The comparison notes that both projects are actively maintained and have meaningful adoption, while alternatives such as TEN Framework, Dograh, Vision Agents, and Vocode serve narrower roles or face licensing, platform-dependency, or maintenance limitations. The recommended choice depends on whether teams need a managed, opinionated agent platform with integrated infrastructure, where LiveKit is generally favored, or a highly configurable pipeline architecture for research, unusual media workflows, direct carrier audio handling, or parallel processing, where Pipecat may be more appropriate.

LiveKit has launched a Startup Program aimed at helping eligible early-stage companies develop and deploy voice, video, and physical AI agents without managing multiple model-provider accounts, API keys, billing systems, or integrations. The program offers approximately $23,000 in value, including $7,000 in inference credits across partners such as Deepgram, Inworld, Rime, Fish Audio, AssemblyAI, Gradium, and LiveKit-hosted Gemma models, one year of the $500-per-month LiveKit Cloud Scale plan, and an additional $10,000 in Cloud credits. Through LiveKit Inference, participants can test speech-to-text, language, and text-to-speech models by changing configuration settings while credits are automatically applied, and services revert to standard rates if a credit pool is exhausted. The Scale plan supports production deployment with agent session capacity, concurrency, region pinning, access controls, metrics exports, custom voices, and discounted inference, while LiveKit emphasizes SOC 2 Type II compliance and default zero data retention for prompts, audio, and outputs. Applicants generally must have fewer than 50 employees, less than $5 million raised, a live or imminent product centered on voice, video, or physical AI, and an identifiable founding team.

LiveKit voice-agent warnings stating that VAD inference is slower than realtime indicate that the voice activity detection model running within the agent process is falling behind incoming audio, rather than a problem with LiveKit Inference STT, TTS, or LLM services. Because VAD detects when users begin and stop speaking, persistent delay can cause late responses, missed interruptions, clipped greetings, choppy egress recordings, and rising memory use; the reported delay measures how far behind live audio processing has become, with Python reporting seconds and Node.js milliseconds. Brief startup warnings may be harmless, particularly with the Silero plugin, but recurring warnings should be addressed by updating LiveKit Agents and turn detection, using the bundled inference VAD where appropriate, checking for blocked event loops caused by synchronous calls or CPU-heavy tasks, and ensuring adequate CPU and correctly configured self-hosted worker capacity. VAD metrics can help distinguish sustained CPU starvation from intermittent event-loop blocking, while increasing VAD threads, speech buffers, or idle processes is not recommended because these changes do not improve realtime processing and may worsen performance.

LiveKit’s non-production deployments allow voice-agent teams to run the same container image, project, and shared secrets as production under separate named environments, reducing the need to maintain duplicate agents while enabling testing of models, prompts, and integrations on real infrastructure. These deployments can be created and managed through the CLI using a `--deployment` flag, sleep when idle to avoid compute costs, and can be targeted for calls through tokens or dispatch APIs, although they require recent Python or Node SDK versions. Unlike production, non-production deployments cold-start on demand, immediately terminate active sessions during redeployment, and do not support rollback, making them appropriate for iterative testing rather than interrupt-sensitive traffic. Deployment-specific behavior or credentials can be selected at runtime through the `LIVEKIT_AGENT_DEPLOYMENT` environment variable, while CI workflows can route branches to staging or production and optionally require approval before production releases. The feature is available on Ship and higher plans, with limits of two non-production deployments per agent on Ship and five on Scale; logs can be used for inspection, while deployment-specific observability metrics are planned for the future.

LiveKit has launched Connectors on LiveKit Cloud, enabling AI agents to handle calls directly through Twilio Programmable Voice and WhatsApp Business Calling without requiring SIP trunks, gateways, or redesigned telephony systems. Each call enters a LiveKit room as a connector participant, allowing agents to use existing capabilities such as turn detection, noise cancellation, observability, DTMF handling, and session analytics while retaining platform-specific context like Twilio call IDs or WhatsApp caller metadata. WhatsApp calls use native WebRTC without transcoding, while Twilio Media Streams audio is bridged from WebSockets, and processing follows regional settings. The service is designed for organizations that want to preserve Twilio workflows, phone numbers, call routing, and compliance arrangements, or serve WhatsApp-first markets through a unified agent platform. Connectors support agent dispatch at connection time, avoid connecting callers to silence, use customers’ own platform accounts and numbers, and do not store credentials or record media. LiveKit positions SIP as its preferred option for new Twilio deployments because of broader routing and latency advantages, while Connectors offer a simpler integration path for established Twilio and WhatsApp systems; pricing matches SIP usage, and additional platform integrations are planned.