FOUNDER BRIEF · NOVEMBER 15, 2026
Read the updated competitor research and launch plan · November 15, 2026
Give the web
a better voice.
Recommendation: build Vozubi as voice-first website support for small businesses, starting with WordPress/WooCommerce. The differentiator should be easy onboarding, reliable business actions, and a clear customer receipt plus private owner activity log. Voice alone is already available from competitors.
This page is a planning brief, not a claim that these backend features are implemented. The landing page is a static HTML/CSS/JavaScript prototype. Its call and chat interactions are scripted; it stores no customer records, captures no audio, and sends no signup requests.
Can GPT-Live power it? Yes.
OpenAI’s current GPT-Live documentation describes a conversational voice layer that can listen while speaking and delegate work to a backend. That maps well to Vozubi: the voice handles the conversation while a separate backend retrieves policies, looks up orders, and executes approved actions. GPT-Live overview.
Use WebRTC for the browser voice connection. Your server creates the session, keeps permanent API credentials private, and returns the connection data to the browser. OpenAI documents a Live session endpoint with a WebRTC SDP offer/answer flow. Confirm model access and quota in the project before implementation. Live session API.
The application still owns permissions, confirmations, business records, and task state. A voice model’s statement is never proof that a refund succeeded. Use the payment or commerce provider’s confirmed result. Delegation and tools.
The proposed architecture
Delegated task authenticated business backend policy checks connected system
Confirmed action event record customer confirmation + private owner activity
| Layer | Recommended starting point | Purpose |
|---|---|---|
| Landing page | Current static HTML/CSS/JS | Fast, portable, and easy to revise before brand validation. |
| Product UI | React + TypeScript; Next.js for owner dashboard | Reusable call controls, chat, onboarding, settings, and activity views. |
| Embeddable widget | Small TypeScript loader + isolated iframe | Keep client CSS from breaking the widget; lazy-load voice on user intent. Explicit microphone permission and iframe permissions are necessary. |
| Voice | GPT-Live + browser WebRTC | Natural conversation; Web Audio analyser for genuine audio-level animation. |
| Text and tasks | Responses API or independently hosted agent; Node.js/TypeScript service | Share conversation context across channels; call authorized tools. Benchmark task accuracy and latency before selecting a backend model. |
| Data | Postgres, tenant-scoped access, pgvector; object storage | Businesses, members, conversations, knowledge sources, action events, and optional uploads. |
| Background jobs | Durable queue + workers | Knowledge indexing, retries, webhook reconciliation, summaries, and notifications. |
| Operations | Managed auth, usage metering, billing, structured logs and error monitoring | Access controls, per-tenant quotas, spend visibility, incident handling. |
These are architecture recommendations, not installed dependencies. Use a Node service that supports persistent sideband connections if the chosen Live integration needs them; avoid assuming a short-lived request handler can own an entire call. The landing page is separate from the future production backend.
One event record, two views
The customer’s right-side panel should contain useful links, order details they are authorized to see, next steps, and confirmed action receipts. The business view adds policy version, approvals, execution status, provider receipt ID, failure details, and the staff or agent responsible. Do not expose internal notes or model reasoning to customers.
Persist a structured action event with tenant ID, conversation ID, customer ID, action type, amount/currency where relevant, approval, idempotency key, provider reference, timestamps, and status. Suggested lifecycle: requested awaiting approval executing confirmed / failed / unknown. Reconcile an unknown provider result before retrying; never risk duplicate refunds.
Customer and owner streams should be separately authorized projections of the same server-side events. Do not send private business fields to the browser and merely hide them in the UI.
How clients install Vozubi
| Client website | Install path | Considerations |
|---|---|---|
| WordPress | A lightweight plugin that inserts the hosted loader | Store a public workspace identifier, not model secrets. Add admin controls for branding and permitted pages; test caching, themes, and script optimizers. |
| WooCommerce | Plugin plus authorized store integration | Verified customer/order ownership, scoped API permissions, webhooks, and configurable return/refund rules. Installation alone must not authorize refunds. |
| Custom websites | Async script embed; optional React wrapper | Domain registration, CSP configuration, signed customer identity issued by the client’s backend, and SPA navigation support. |
| Shopify | Theme app extension plus app authorization | Use platform scopes and authenticated sessions for store actions; confirm applicable distribution and installation requirements. |
| Webflow, Wix, Squarespace | Supported custom-code embed | Some plans/locations restrict scripts. Validate installation, iframe microphone policy, mobile layout, and cross-origin behavior per builder. |
The widget should work anonymously for public FAQs. Require stronger identity before disclosing order details or changing anything. HTTPS, user-initiated microphone access, keyboard navigation, captions/text fallback, iOS/Safari checks, reconnect behavior, and slow-network handling are launch requirements. Test the exact embed rather than only a standalone demo.
Competitive landscape
These are selected comparisons from official product documentation, reviewed November 15, 2026. The opportunity assessments are our strategic interpretation, not claims that competitors cannot offer a feature.
| Competitor | Verified offering | Implication for Vozubi |
|---|---|---|
| Intercom / Fin Voice | AI phone support with deployment controls and human-support routing. | Compete on a focused small-business website experience and simpler onboarding, rather than breadth of enterprise helpdesk functionality. |
| Vapi | Embeddable widget supporting voice and text, configurable appearance, and conversation context. | A voice widget is not unique. Vozubi needs opinionated business workflows, support-specific UX, and trustworthy action history. |
| Retell AI | Browser calls through a Web SDK. | Could be an infrastructure alternative as well as a competitor. Compare engineering effort, call quality, control, and total cost. |
| Tidio / Lyro | Conversational AI support with a documented knowledge and channel configuration workflow. | Use SMB support and onboarding as the comparison benchmark. Validate willingness to pay for voice before assuming it improves conversion. |
Proposed positioning: “Voice and chat support for your website—with every important action accounted for.” Start with one vertical and a few excellent workflows. Avoid promising universal integrations, guaranteed resolution percentages, or unlimited voice.
Launch in four stages
- Validate the experience: refine this landing page, run customer interviews, test voice preference, and choose one business category.
- Read-only pilot: real voice and chat, knowledge retrieval, order lookup with verification, owner transcripts, and human handoff. Measure answer accuracy, latency, interruptions, and cost per conversation.
- Controlled actions: one integration and one reversible workflow first. Start refund requests with human approval; add autonomous low-value refunds only after policy and reliability evaluations.
- Distribution and scale: WordPress plugin, self-service onboarding, billing, richer integrations, operational dashboards, and broader browser coverage.
Before serving paying clients: add tenant isolation tests, adversarial knowledge/tool tests, replayable evaluations, role controls, scoped secrets, rate limits, concurrency limits, idle disconnects, spend caps, retention/deletion controls, and alerting. Define escalation when the model, store, or network fails. Review recording consent and applicable privacy requirements with qualified counsel for the markets you serve.
Pricing needs real call data
Budget for voice input/output, backend reasoning, retrieval, storage, infrastructure, support, and any telephony or intermediary provider fees. Live web calls do not inherently need a telephone number. Measure short, medium, and long calls with realistic tool activity before setting plans. Recheck official model pricing when doing that exercise; this brief does not assert a fixed per-minute rate.
My recommendation is a base subscription with included usage, transparent overages, and an enforceable spending cap. The reference image’s $29/$59 prices and “unlimited” claims should not be copied without unit-economics validation.
Brand direction: lime + sky
The chosen identity is Fresh Frequency: lime #D8F8A8 and sky #8ED8FF, with ink #171A18 and soft gray #F3F5F4. Explore the original headset mascot and its expression system.
Design reference review
All nine supplied images informed this version: (1) oversized hero, capsule navigation, photo and floating cards; (2) central conversational visual; (3) broad paired benefit rows; (4) gradient text highlights and action cards; (5) orbit integration diagram; (6) product interface showcase; (7) large split narrative panel; (8) paired pricing cards; (9) spacious closing CTA and rounded surface. Vozubi substitutes an audio orb, customer details, and owner activity for the video-first references.
Reference testimonials and performance numbers were not reused as Vozubi claims. The hero uses a locally stored AI-generated fictional blonde support persona. Mascot renders are also local assets. DM Sans loads from Google Fonts.
Explore the prototype