Installation

Package

npm install @oshara/voice-sdk livekit-client
# or
pnpm add @oshara/voice-sdk livekit-client
# or
yarn add @oshara/voice-sdk livekit-client

livekit-client and @livekit/krisp-noise-filter are runtime dependencies; in most apps your bundler dedupes a single copy. react / react-dom are optional peers — only needed for the /react entry point.

Entry points

The package exposes three importable entry points:

import { createVoiceAgent } from "@oshara/voice-sdk";       // headless core
import { mountVoiceUI }     from "@oshara/voice-sdk/ui";    // prebuilt widget UI
import { useVoiceAgent }    from "@oshara/voice-sdk/react"; // React bindings

If you use the prebuilt UI (/ui or <VoiceWidget>), also load the stylesheet once:

import "@oshara/voice-sdk/styles.css";

No build step? Use the script tag

If you can’t (or don’t want to) run a bundler, skip npm entirely and drop in the embeddable widget — it’s built on this SDK:

<script
  src="https://api.oshara.ai/widget.js"
  data-agent="support-bot"
  data-api-url="https://api.oshara.ai">
</script>

See Creating an Agent → Script tag.

Node / server-side

The core is safe to import and construct in Node:

• localStorage access (audio prefs, consent, visitor id) is wrapped in try/catch.
• navigator.mediaDevices / AudioContext are only touched when you call audio/device methods.
• The Krisp and DeepFilter noise filters are dynamically imported only when a call attaches them — import "@oshara/voice-sdk" never pulls browser-only Worker/WASM code server-side.

On Node < 18 there is no global fetch; pass one in the config:

import fetch from "node-fetch";
createVoiceAgent({ agentSlug: "support-bot", apiKey: process.env.OSHARA_KEY, fetch });