The outcome

Run a narrow voice workflow that users can understand, interrupt, correct, and exit without unauthorized consequences.

LiveKit is a Open-source realtime framework and cloud platform for voice, video, text, telephony, and multimodal AI agents. Building low-latency voice assistants, multimodal room participants, phone agents, realtime translation, tool-using conversational workflows, and human handoffs. This guide narrows that broad capability into one repeatable outcome, with checkpoints that keep the source material and your judgment in the loop.

Before you begin

Set the boundary before the tool starts.

Choose one real task, identify who will use the result, and decide what evidence or test will make the result acceptable. Gather only the source material needed for that task. If the work contains confidential, personal, regulated, or client-owned information, confirm that the platform and account are approved before sharing it.

Troiana principle

AI should make the work easier to inspect. If the workflow removes the source, the owner, or the review step, redesign the workflow.

Step by step

A workflow you can repeat.

  1. 01

    Define caller eligibility, purpose, disclosure, consent, languages, supported intents, identity verification, recording, retention, accessibility, tool permissions, and human escalation.

  2. 02

    Prototype with synthetic data and a simple pipeline, state that the caller is speaking with AI, keep credentials server-side, and make tools read-only or sandboxed.

  3. 03

    Tune turn detection and interruption without hiding latency, and require spoken confirmation plus backend authorization before any write, purchase, disclosure, or transfer.

  4. 04

    Test accents, noise, silence, crosstalk, barge-in, prompt injection, hallucination, unsafe requests, disconnection, tool timeout, repeated actions, and failed handoff.

  5. 05

    Canary with monitored low-risk traffic, provide mute, stop, correction, and human options, preserve approved audit metadata, and review complaints and failures before expansion.

Working standard

What good use looks like.

  • Disclose AI at the start.
  • Confirm every consequential action.
  • Keep human handoff always reachable.

Realtime agents can hear, transcribe, infer, speak, call tools, and initiate telephony actions before a reviewer can intervene. Hidden observation, recording, voice cloning, payment or identity collection, and cross-provider routing carry serious consent and privacy risk. LiveKit Inference is zero-data-retention by default, but plugins, observability, recordings, telephony, and external models have separate policies. Require clear disclosure, least-privilege tools, explicit confirmation, human handoff, and an immediate kill switch.

Official references

Check the current product documentation.

Features, plan limits, availability, and data controls change. These official pages are the starting points used for this collection.