AVAAZ

AI for Voice Agents

Give your business a voice.

Paste a website. Talk the gaps. Share a live voice agent — in one sitting.

Website → Voice → Agent → Customers talk

Avaaz is the voice layer for any business, creator, or campus. It turns a site, a conversation, or a few notes into a public AI voice agent anyone can call. Visitors get answers out loud — not another chatbot wall of text.

Avaaz — AI for Voice Agents — Give your business a voice.


Create	Google sign-in. No code. No widget install.
Go live	One `/talk` link. Follow. Embed.
Stay on	Knowledge first. Web search if needed. Mic stays open.

The business angle

Most companies already paid for a website. Almost none of them can talk.

Hours, products, pricing, and FAQs sit in pages that customers will not read — especially on a phone, after hours, or while walking in. The default answer is a chat widget that types walls of text, or a human who is closed, busy, or expensive.

Avaaz’s bet: if the knowledge exists, it should be a voice in under 10 minutes — not a six-week bot project.

Problem	What it costs	What Avaaz does
Unanswered “are you open / how much / do you have X?”	Missed visits and abandoned carts	A public voice that already knows the site
Support queues for the same 20 questions	$6–$15 per live-agent contact (typical BPO range)	Deflect FAQs to a shareable agent
After-hours silence	Lost demand when the shop is closed	24/7 talk page, no extra headcount
Chatbots nobody finishes	High bounce on long text replies	Speak, hear, keep the mic open
Custom “AI agent” builds	Weeks + an engineer	Paste URL → talk gaps → publish

Who pays: the business that owns the knowledge (shop, SaaS, campus, creator). Who talks for free: their customers. Distribution is a link or an embed — not an app store.

flowchart LR
  Site[Existing website] --> Avaaz[Avaaz agent]
  Owner[Owner talks the gaps] --> Avaaz
  Avaaz --> Link[One public link]
  Link --> C1[Customer on mobile]
  Link --> C2[After hours]
  Link --> C3[Embedded on the site]

Impact

Figures below are market context (widely cited industry ranges), used to size the problem Avaaz attacks — not live Avaaz production metrics.

Signal	Number	Why it matters for Avaaz
Calls small businesses never pick up	~60%+ unanswered in many local-business studies	A voice agent does not go to voicemail
Customers who expect an “immediate” reply	~80–90% (service-benchmark surveys)	Voice is faster than finding a FAQ page
Speaking vs typing	~3–4× faster than typing the same request	Lower effort than a chat box
Live support cost per contact	~$6–$15 typical; AI deflection cents to ~$1	FAQ voice pays for itself on a handful of contacts
Conversational AI market	tens of billions USD this decade (analyst estimates)	Voice is the interface customers already use on phones
Time to a traditional custom bot	weeks	Avaaz target: one sitting
Time to an Avaaz agent	~5–15 minutes after Google sign-in	Website crawl + a few spoken answers
Engineering required	0	Owner pastes a URL and talks

What changes for a business that ships Avaaz

Before	After
Site is a brochure	Site becomes something people can ask
Staff repeat hours, SKUs, policies	Agent recites them; humans take exceptions
Closed sign = zero answers	Talk page still listens
“Install our widget” friction	One URL, optional embed

A single deflected “what are your hours / do you deliver / what’s included?” loop, a few times a day, is already more value than a $12 Studio month. Scale is extra agents, extra brands, extra languages later — the wedge is always-on voice for knowledge that already exists.

Business model

Short version: free to create, paid to scale. The product is the agent. The customer is the owner.

flowchart TD
  Free[Starter · Free] --> Studio[Studio · $12 / mo]
  Studio --> House[House · custom]
  Free --> Use[Create + public talk link]
  Studio --> More[Unlimited agents · sharper presence]
  House --> Ent[Volume, crawl priority, guided setup]

Plan	Price	What you get	Who it is for
Starter	Free	Up to 3 agents, website + voice setup, public talk page	Trying Avaaz, one shop, a campus demo
Studio	$12 / month	Unlimited agents, custom logo & FAQs, follower insights	Teams and multi-brand owners
House	Talk to us	Priority crawl, team seats, guided setup	Brands that need volume and support

How money shows up

Motion	Detail
Land	Free Starter — Google, paste site, go live. No card to try.
Expand	Studio when a fourth agent, a second brand, or polish is needed
Enterprise	House — crawl volume, seats, white-glove
Unit cost	LLM tokens via OpenRouter; live mic/TTS can stay in the browser to keep COGS low
Not in v1	Ads on talk pages, taking a cut of customer purchases

Pricing on /pricing is display-first for the MVP. The model is SaaS subscription on agents owned, not per visitor question.

What it does

Avaaz is a voice-first knowledge agent. Four jobs, one product:

	Capability	What happens
1	Listen	Businesses: Firecrawl reads the website. Everyone else: you talk.
2	Understand	OpenAI (via OpenRouter) extracts name, hours, products, FAQs, personality.
3	Speak	Visitors talk on a public page. Answers from knowledge, then the web if needed.
4	Share	One `/talk/[slug]` link. Follow. Copy. Embed.

Not a generic chatbot with a microphone. The public page is built to listen, answer out loud, and keep the mic open.

Why it started

Knowledge is stuck in a website nobody reads, a PDF nobody opens, or someone’s head.

Typical “AI for business” asks you to fill a CMS, train a bot, then drop a widget. Slow to set up. Awkward on a phone.

Avaaz started as a hackathon MVP with one bet:

If the knowledge already exists, talking should be enough to turn it into an agent.

Start from a URL, not a blank dashboard
Fill only the gaps by voice
Ship a public voice page in one sitting
Keep the UI light, calm, and shareable on one screen

Designed and directed by Vansh. Built with OpenAI Codex as the implementation partner (see How Codex was used).

Who can use it

Audience	How they use Avaaz	Example
Businesses	Paste the company site → confirm gaps by voice → share with customers	Payments company, restaurant, jewellery shop
Creators	Talk through who they are → a public voice that introduces them	Portfolio / personal brand
Teachers & campuses	Point at a site or speak FAQs → students ask out loud	Global Campus X (`/talk/globalcampusx`)
Anyone else	Choose Other, speak the idea	Event, club, side project
Visitors	Open a talk link — no account to ask questions	Google required only to follow

Creating an agent requires Google sign-in. Talking to a public agent does not.

How easy it is

Four steps. Typing is optional after the URL.

flowchart LR
  A["1. Sign in with Google"] --> B["2. Pick Business / Personal / Other"]
  B --> C["3. Paste site or talk"]
  C --> D["4. Pick a voice and publish"]

Step	What you do	What Avaaz does
Sign in	One Google click	Ties agents and follows to you
Type	Business / Personal / Other	Routes the create flow
Website (business)	Paste a URL	Crawls the page, drafts knowledge
Voice gaps	Answer only what’s missing	Merges speech into the agent
Voice	Nova, Coral, Aria, or Atlas	Distinct female / male voices
Share	Copy the link	Anyone talks at `/talk/[slug]`

Time to live: one sitting. No SDK. No prompt engineering from the owner.

Product flow

flowchart TD
  Start([Avaaz /]) --> CTA[Create my agent]
  CTA --> Auth{Signed in?}
  Auth -->|No| Google[Google OAuth]
  Google --> Auth
  Auth -->|Yes| Type[Choose type]

  Type --> Biz[Business]
  Type --> Per[Personal]
  Type --> Oth[Other]

  Biz --> URL[Paste website]
  URL --> Crawl[Firecrawl reads the site]
  Crawl --> Extract[OpenAI extracts structured knowledge]
  Extract --> Gaps[Voice: only missing fields]
  Per --> Talk[Voice onboarding]
  Oth --> Talk
  Gaps --> Know[Knowledge + voice picker]
  Talk --> Know
  Know --> Save[(Supabase agents)]
  Save --> Dash[/agent/slug]
  Dash --> Public[/talk/slug]
  Public --> Listen[Visitor speaks]
  Listen --> Answer[Agent answers from knowledge]
  Answer --> Search{Still unknown?}
  Search -->|Yes| Web[OpenAI + web search]
  Web --> Speak[Browser TTS]
  Search -->|No| Speak

System architecture

flowchart TB
  subgraph Client["Browser"]
    UI[Next.js App Router]
    Mic[Web Speech Recognition]
    TTS[speechSynthesis]
    AuthUI[Google via Supabase Auth]
  end

  subgraph Next["Next.js server"]
    API["/api/*"]
    Prompt[Agent prompts]
    Extract[Knowledge extract]
  end

  subgraph AI["OpenAI via OpenRouter"]
    GPT["openai/gpt-4o-mini"]
    Whisper["whisper-large-v3 optional"]
    Speech["gpt-4o-mini-tts optional"]
  end

  subgraph Data["Data and crawl"]
    SB[(Supabase Postgres + Storage)]
    FC[Firecrawl]
    Logo[Logo.dev]
  end

  UI --> API
  AuthUI --> SB
  Mic --> UI
  UI --> TTS
  API --> Prompt
  API --> Extract
  Extract --> GPT
  Prompt --> GPT
  GPT -->|web plugin if needed| GPT
  API --> FC
  API --> SB
  UI --> Logo

Keys stay on the server. OPENROUTER_API_KEY, Firecrawl, and the Supabase service role never ship to the client.

Voice conversation flow

Live talk: browser for mic and speech. Server only reasons.

sequenceDiagram
  participant V as Visitor
  participant B as Browser
  participant API as /api/voice/turn
  participant OR as OpenRouter / OpenAI
  participant DB as Supabase

  V->>B: Tap mic and speak
  B->>B: SpeechRecognition to text
  B->>API: POST history + user text + slug
  API->>DB: Load agent by slug
  API->>OR: Chat with agent knowledge gpt-4o-mini
  alt Answer looks like I don't know
    API->>OR: Retry with web search plugin
  end
  OR-->>API: Assistant text
  API-->>B: JSON assistantText
  B->>B: speechSynthesis with chosen voice
  B->>V: Spoken answer, mic stays open

Layer	Technology	Role
Speech-to-text (live)	`SpeechRecognition`	Speech → text in the browser
Reasoning	OpenAI `gpt-4o-mini` via OpenRouter	Onboarding + public answers
Fallback search	OpenRouter `plugins: web` or `:online`	When stored knowledge is not enough
Text-to-speech (live)	`speechSynthesis`	Nova / Coral / Aria (female), Atlas (male)
Optional STT / TTS	Whisper + OpenAI TTS via OpenRouter	Wired in `lib/openrouter/client.ts`

How OpenAI is used

Avaaz talks to OpenAI through OpenRouter with the official openai Node SDK at https://openrouter.ai/api/v1. One key, one client, model names like openai/gpt-4o-mini.

Env var	Default	Used for
`OPENROUTER_CHAT_MODEL`	`openai/gpt-4o-mini`	Extract, onboarding, public Q&A
`OPENROUTER_TTS_MODEL`	`openai/gpt-4o-mini-tts-2025-12-15`	Server TTS (optional)
`OPENROUTER_STT_MODEL`	`openai/whisper-large-v3`	Server transcription (optional)

flowchart LR
  App[Avaaz server] -->|OpenAI SDK| OR[OpenRouter]
  OR --> Chat["Chat Completions openai/gpt-4o-mini"]
  OR --> Audio["Audio APIs Whisper + TTS"]
  Chat --> Extract[JSON knowledge]
  Chat --> Reply[Voice replies]
  Chat --> Search[Web-backed reply]

Job	File	Model returns
Website → knowledge	`lib/openai/extract.ts`	JSON: name, hours, products, FAQs
Voice onboarding	`app/api/voice/turn/route.ts`	Next question, only missing fields
Public agent	same + `lib/agents/prompt.ts`	Short spoken answers
“I don’t know”	same, `search: true`	Web answer, prefer official site

The OpenAI key never appears in the browser.

How Codex was used

Avaaz was built as a directed pairing session with OpenAI Codex — not an unattended dump. Vansh set product taste (voice-first, light UI, talk-the-gaps). Codex implemented, refactored, and iterated against a running app.

flowchart TD
  Idea[Avaaz: talkable knowledge] --> Codex[OpenAI Codex]
  Codex --> Loop[Read codebase then change then verify]
  Loop --> Review[Human review: copy, UX, keys, taste]
  Review --> Loop
  Review --> Ship[Hackathon MVP]

Area	What Codex helped ship
App shell	Next.js App Router, Tailwind v4, navbar, footer, about, pricing
Create flow	Category → website crawl → voice gaps → knowledge → generating
Voice UX	Mic that stays open, stop-while-speaking, history, voice picker
APIs	`/api/voice/turn`, website extract, agents CRUD, follow
Data	`supabase/schema.sql`, featured agents, follow counts
Integrations	OpenRouter / OpenAI client, Firecrawl, Logo.dev, Google OAuth
Polish	Gallery, DriftWall, About, this README

Rules while using Codex

Product decisions stay with the human
Secrets stay in .env.local — never committed, never echoed
OpenRouter / service keys stay server-only
Iterate on a running app, not a disconnected prototype

Codex was the implementation partner. The intent — give your business a voice — is human.

Final tech stack

Layer	Choice	Role in Avaaz
App	Next.js 16 (App Router)	Pages, RSC, `/api` in one repo
UI	React 19 + TypeScript	Type-safe product UI
Style	Tailwind CSS v4	Light, calm marketing + talk page
Auth	Supabase Auth (Google)	Create + follow
Database	Supabase Postgres	Agents, files, conversations, follows
Files	Supabase Storage `agent-files`	Optional uploads
LLM	OpenAI `gpt-4o-mini` via OpenRouter	Extract, interview, answers, web fallback
Optional audio	Whisper + OpenAI TTS via OpenRouter	Server STT / TTS if credits allow
Live voice I/O	Web Speech API	Mic + playback with low COGS
Crawl	Firecrawl	Business website → text
Logos	Logo.dev	Gallery and talk-page marks
Icons	lucide-react	UI
Docs / PDFs	mammoth, unpdf	Knowledge uploads
Share	qrcode	Dashboard / share flows
Build partner	OpenAI Codex	Implementation speed

flowchart TB
  subgraph Frontend
    NextJS[Next.js 16]
    React[React 19]
    TW[Tailwind v4]
  end
  subgraph Intelligence
    OR[OpenRouter]
    OAI[OpenAI gpt-4o-mini]
    OR --> OAI
  end
  subgraph Platform
    SB[Supabase Auth + Postgres + Storage]
    FC[Firecrawl]
    LD[Logo.dev]
  end
  NextJS --> OR
  NextJS --> SB
  NextJS --> FC
  NextJS --> LD

Setup

npm install
cp .env.example .env.local

Fill keys in .env.local, then run supabase/schema.sql in the Supabase SQL editor.

npm run dev

Open http://localhost:3000.

Environment variables

Variable	Required	Used for
`OPENROUTER_API_KEY`	Yes (voice + extract)	Server-only OpenAI calls
`NEXT_PUBLIC_SUPABASE_URL`	Yes	Auth + data
`NEXT_PUBLIC_SUPABASE_ANON_KEY`	Yes	Publishable key
`SUPABASE_SERVICE_ROLE_KEY`	Optional	Server uploads
`FIRECRAWL_API_KEY`	For website create	Crawl business sites
`NEXT_PUBLIC_SITE_URL`	Optional	OpenRouter attribution
`NEXT_PUBLIC_LOGO_DEV_TOKEN`	Optional	Company logos
`OPENROUTER_CHAT_MODEL`	Optional	Defaults to `openai/gpt-4o-mini`

Never put OPENROUTER_API_KEY or SUPABASE_SERVICE_ROLE_KEY in client code.

Google login: enable Google in Supabase Auth. Add http://localhost:3000/auth/callback (and production) to the redirect allow list.

Routes

Path	Who it’s for
`/`	Landing, how it works, gallery
`/about`	Overview, products, why Avaaz
`/pricing`	Starter / Studio / House
`/create`	Business / Personal / Other
`/create/website`	Paste URL (business)
`/create/voice`	Talk the gaps
`/create/knowledge`	Files + voice picker
`/create/generating`	Build the agent
`/agent/[slug]`	Owner dashboard
`/talk/[slug]`	Public voice (`?embed=1` for embed)
`/auth/callback`	Google OAuth return

Featured agents

Try talk pages without creating one:

Agent	Talk path	Site
Dodo Payments	`/talk/dodo-payments`	dodopayments.com
Post Bridge	`/talk/post-bridge`	post-bridge.com
Supabase	`/talk/supabase`	supabase.com
Global Campus X	`/talk/globalcampusx`	globalcampusx.com

Local walkthrough

Open /
Create my agent → Google
Choose Business
Paste a website (or skip to voice for Personal / Other)
Answer only the missing questions out loud
Pick Nova, Coral, Aria, or Atlas
Open the public agent and ask What do you do? / What are your hours?
Share the /talk/[slug] link

Security notes

OpenRouter and service-role keys are server-only
Public talk pages are meant to be shared; do not put secrets in agent knowledge
Schema RLS is open for the hackathon MVP — tighten before production

AVAAZ
AI for Voice Agents
Give your business a voice.

Made by Vansh · OpenAI Codex · OpenAI via OpenRouter · Next.js · Supabase

Name		Name	Last commit message	Last commit date
app		app
components		components
lib		lib
public		public
supabase		supabase
types		types
.env.example		.env.example
.gitignore		.gitignore
AGENTS.md		AGENTS.md
CLAUDE.md		CLAUDE.md
README.md		README.md
eslint.config.mjs		eslint.config.mjs
next.config.ts		next.config.ts
package-lock.json		package-lock.json
package.json		package.json
postcss.config.mjs		postcss.config.mjs
tsconfig.json		tsconfig.json

Navigation Menu

Repository files navigation

AVAAZ

Table of contents

The business angle

Impact

Business model

What it does

Why it started

Who can use it

How easy it is

Product flow

System architecture

Voice conversation flow

How OpenAI is used

How Codex was used

Final tech stack

Setup

Environment variables

Routes

Featured agents

Local walkthrough

Security notes

About

Releases

Packages

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

AVAAZ

Table of contents

The business angle

Impact

Business model

What it does

Why it started

Who can use it

How easy it is

Product flow

System architecture

Voice conversation flow

How OpenAI is used

How Codex was used

Final tech stack

Setup

Environment variables

Routes

Featured agents

Local walkthrough

Security notes

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages