Talk to Voxy widget

The “Talk to Voxy” widget is a floating bottom-right CTA you can mount on a marketing site. A visitor clicks it, fills a short form, solves a captcha, and is connected to a live AI voice conversation in their browser — no phone, no install.

What ships today

• Floating brand-pill CTA on the landing page.
• 6-step state machine: closed → intro → captcha → connecting → live → ended | error.
• Mic capture via AudioWorklet at 16 kHz mono Int16 PCM.
• Playback at 24 kHz via sequenced AudioBufferSourceNode.
• Live transcript pane with role colouring and auto-scroll.
• Hangup button + visitor-initiated close.

Lifecycle

1. Visitor clicks the floating CTA.
2. Form: name (optional), email (required), phone (optional).
3. Captcha: 7+4 or similar math challenge; verified via HMAC.
4. POST /v1/public/web-agent/sessions/start returns a session id + audio-socket UUID.
5. Browser opens a WS to /ws/web-agent?session=<uuid> on the voice-bridge.
6. Voice-bridge resolves the session via the API, opens a Voxy AI voice connection, starts streaming audio.
7. On hangup the session ends; a post-session AI summary stamps sentiment + lead temperature + entities.

Hard caps

• max_session_seconds from the active web-agent config (default 600).
• 25 concurrent sessions platform-wide.
• Per-IP captcha rate limit (30/min via web-agent-public throttle bucket).
• Daily session cap per active config (default 500).

Admin surface

/admin/web-agent shows the active config + recent sessions. /admin/web-agent/sessions is an SSE-streamed live roster with a Stop button per row that force-closes a visitor's WS. Prompt editor, KB browser, and recordings player are the v2 polish.