Embeddable voice AI SDK for web pages
Speak to fill forms, click buttons, navigate, and ask questions. Works with any LLM provider.

• Voice & text input — Browser Speech API with automatic text fallback
• Form filling — Detects fields, fills values, triggers React/Vue/Angular change detection
• Smart page scanning — Auto-discovers forms, headings, navigation, interactive elements
• Multi-LLM support — Gemini, OpenAI, Anthropic, Ollama (any OpenAI-compatible API)
• Themeable UI — Presets, sizes, full color control
• Conversation workflows — Guided multi-step flows with validation
• Accessibility — ARIA live regions, keyboard shortcuts, screen reader tools
• Zero dependencies — Self-contained SDK with Shadow DOM isolation

Browser (SDK)                    Server (proxy)
─────────────                    ──────────────
SpeechRecognition → text ──WS──→ Receives text
Execute DOM actions ←──WS──────← LLM tool calls
SpeechSynthesis (TTS)            LLM API (holds key)
Page context scanning            Session/history mgmt
Shadow DOM UI                    Context caching

The server is a thin proxy that holds your API key — it never reaches the browser.

cd server && npm install
GEMINI_API_KEY=your-key npm run dev

# OpenAI / GPT
OPENAI_API_KEY=your-key LLM_PROVIDER=openai npm run dev

# Anthropic / Claude
ANTHROPIC_API_KEY=your-key LLM_PROVIDER=anthropic npm run dev

# Ollama (local, no key needed)
LLM_PROVIDER=ollama npm run dev

Script tag (IIFE):

<script src="https://your-server.com/sdk/voice-sdk.iife.js"></script>
<script>
  const sdk = new VoxGlide.VoiceSDK({
    serverUrl: 'wss://your-server.com',
  });
</script>

import { VoiceSDK } from 'voxglide';

const sdk = new VoiceSDK({
  serverUrl: 'wss://your-server.com',
  autoContext: true,
  tts: true,
});

That's it. The SDK auto-discovers forms and interactive elements on the page.

const sdk = new VoiceSDK({
  serverUrl: 'wss://your-server.com',  // Required
  autoContext: true,                     // Auto-scan DOM for context
  context: 'This is a checkout page',   // Developer-supplied context
  language: 'en-US',                     // Speech recognition language
  tts: true,                             // Enable browser text-to-speech
  ui: { theme: 'ocean', size: 'md' },   // UI theming
  debug: false,                          // Verbose logging
  autoReconnect: true,                   // Reconnect after navigation
});

See docs/configuration.md for full configuration reference.

Pages can expose tools via window.nbt_functions — the SDK auto-discovers them:

<script>
  window.nbt_functions = {
    lookupOrder: {
      description: 'Look up an order by ID',
      parameters: {
        orderId: { type: 'string', description: 'The order ID', required: true },
      },
      handler: async (args) => {
        return await fetch(`/api/orders/${args.orderId}`).then(r => r.json());
      },
    },
  };
</script>

You can also register tools via SDK config or at runtime. See the custom tools docs.

The examples/ directory contains demo pages:

• basic.html — Minimal integration
• form-filling.html — Form auto-fill demo
• custom-actions.html — Custom tool registration

We welcome contributions! See CONTRIBUTING.md for setup instructions, code style, and PR guidelines.

git clone https://github.com/billiax/voxglide.git
cd voxglide && npm install
npm run check    # typecheck + lint + test

MIT