What Is the Vaani Developer Platform? The Beginner's Guide to Voice AI Agents
CodexFuse Team · 2026-11-05T07:18:33.151Z · 21 min read min read
What Is the Vaani Developer Platform?
The Vaani Developer Platform is a self-serve developer platform from Vaani AI Research for building, deploying, and managing Vaani voice AI agents. It launched on October 29, 2026, and it turns a plain-language description of the agent you want into a working voice agent that can take and make phone calls, answer on a website or in an app, and talk to customers on WhatsApp.
The important distinction is in the word platform. A voice API gives you a voice and leaves the rest to you.
The Vaani Developer Platform gives you the whole stack: its own speech models, the telephony, the conversation logic, the memory, and the tools, all wired together so you do not have to assemble them yourself. Before you build anything, it is worth checking the number you will hash or sign in your own integrations with a SHA-256 hash generator, and worth knowing that the servers behind a voice agent like this usually run on VPS hosting.
Vaani AI Research is a foundation speech model company based in Bengaluru and Singapore, founded in 2025 by Tushar Shinde, Nitesh Tripathi, and Nitish Mishra. The platform grew out of the company's own enterprise work: it reports more than 14 enterprise deployments before the self-serve launch, and its homepage says the platform was the number one Product of the Day on Product Hunt at launch. This guide is the Vaani Developer Platform explained for beginners: what it is, how the harness works, what it costs, and how it compares with the best-known alternative, Vapi.
Here is the whole build in five steps, using the platform's own Teacher AI assistant:
- Describe the agent you want in your own words: who it should call, what it should say, and what counts as success.
- Answer Teacher AI's clarifying questions about languages, objections, qualifying criteria, and handoffs.
- Watch the agent build in seconds, with a readiness score as the prompt, voice, language mix, and tools come together.
- Review the prompt diff, with recommended settings for latency, voice, and turn-taking, and accept only what you approve.
- Deploy the agent across phone calls, websites, apps, and WhatsApp, then keep refining it by chat.
What Is the Voice Harness?
The Vaani voice harness is the infrastructure layer at the core of the Vaani Developer Platform. It connects Vaani's speech models with everything a voice agent needs to run in production, so a developer deals with one platform instead of five vendors and a pile of glue code. The company's launch announcement describes the harness as the piece that turns a successful demo into an agent that holds up in real conversations.
Real conversations are the hard part. Callers switch languages mid-sentence, interrupt, go quiet, ask something nobody predicted, and expect the agent to remember what they said last week. The harness exists because each of those problems is a different system, and the platform bundles all six of them.
| Harness piece | What it does | Why it matters |
|---|---|---|
| Inference optimisation | Runs the speech models fast enough for live conversation | Vaani claims a best-in-class p90 latency of about 450 milliseconds |
| Memory and context management | Keeps what the caller said, across turns and across calls | Customers do not have to repeat themselves on the next call |
| Conversational orchestration | Decides when to speak, pause, and hand over | Fixes dead air, awkward pauses, and talking over the caller |
| Telephony | Connects the agent to real phone lines | Phone calls work without a separate carrier integration |
| Tool integrations | Lets the agent act: check an order, book a slot, update a record | An agent that can only talk is a recording, not a worker |
| Human-in-the-loop workflows | Transfers a conversation to a human team when needed | Hard cases reach a person instead of looping forever |
Two details on the vendor's homepage show what the harness is tuned for. First, the agents make small talk to fill dead air instead of leaving silence on the line. Second, the voices carry emotional tags and switch language and dialect automatically, which is the multilingual bet the whole company is built on.
How Does Teacher AI Build an Agent?
Teacher AI is the platform's build assistant, and it is the feature that makes the Vaani Developer Platform feel different from a raw API. You describe the agent in plain language, and Teacher AI asks the clarifying questions a good product manager would ask: which languages, which objections, what qualifies as a good outcome, and when should the call hand off to a human. Each answer tightens the agent before it ever makes a call.
The build itself happens live. The prompt, the voice, the language mix, and the tools come together in seconds, with a readiness score shown as it goes.
What you get is a deployment-ready agent, not a draft you still have to finish. That matters because the usual voice-agent build is the opposite: pick a transcriber, pick a model, pick a voice, wire up telephony, write the prompt, test the handoffs, and debug the latency, one vendor at a time.
The safety net is the prompt diff. Every change shows exactly what changed, with recommended settings for latency, voice, and turn-taking, and nothing is saved until you approve it.
One prompt tweak quietly breaking a live agent is the classic voice-AI failure, and the diff review is designed to stop it. You can also keep refining an existing agent by chat: ask Teacher AI to change the tone, add a language, or handle a new objection, and it updates the same way, with a diff to review.
What Can a Vaani Agent Remember?
A Vaani agent remembers past calls, and that memory is a first-class part of the platform rather than an afterthought. The homepage shows the pattern with a concrete example: a customer called about order number 48213 on October 12, asked about a delayed order and a UPI refund on October 15, said they prefer Hindi on October 18, and asked for WhatsApp instead of SMS on October 20. On the next call, the agent already knows the refund was approved, the customer speaks Hindi, and updates go to WhatsApp.
The mechanism is what Vaani calls cumulated context. You choose how many past calls to use and which ones, and only what the current call needs is passed in.
The example shows a filter for order 48213 from the last four calls, producing three facts: the refund was approved, the customer speaks Hindi, and updates go on WhatsApp. Nothing else from the call history is carried over.
Memory is also a pricing line, not just a feature. On the pay-as-you-go plan, agents with memory are included as an early launch offer; the crossed-out list price is 299 rupees per agent.
That tells you two things: memory is a paid capability in the platform's normal pricing, and the launch offer is the cheapest time to try it. The launch offer wording also means the price can change, so check the pricing page before you budget around it.
Where Can You Deploy a Vaani Agent?
A Vaani agent deploys across four channels: phone calls, websites, apps, and WhatsApp. The same agent, built once, runs on all four. That is the point of a harness: the conversation logic is written once, and the channel is just a different way for the caller to reach it.
| Channel | What the caller experiences | What it is best for |
|---|---|---|
| Phone calls | A real phone conversation, with call screening and voicemail detection | Support lines, outbound sales, appointment reminders |
| Websites | A voice agent on your site, with no telephony charge | Answering questions and qualifying visitors in the browser |
| Apps | Voice inside your own product | Hands-free help and voice-first features |
| The agent talks to customers where they already are | Markets where WhatsApp is the default messaging app |
The phone channel gets the most attention on the pricing page, and for good reason: it is the only channel with a telephony charge. WebRTC browser calls have no telephony charge, though speech recognition, reasoning, voice synthesis, and the platform fee still apply. Text chat is a separate offering with its own conversation or token-based pricing.
The platform also handles the awkward edge of phone calls: screening. A screening prompt is a step in the call where Vaani answers and stays on the line while the customer decides whether to pick up, keeping the conversation ready for the handoff. Call screening and voicemail detection are both unlimited on the pay-as-you-go plan, which suggests the company expects screening to be a normal part of outbound calling rather than a premium extra.
Why Is Multilingual the Whole Point?
Multilingual voice is the reason Vaani AI Research exists, and the launch announcement says so in the CEO's own words. Tushar Shinde, the company's co-founder and CEO, said that seventy-five percent of the world does not primarily speak English, and most voice AI still treats that as an edge case. The platform was built inside live enterprise deployments, where every call shows what breaks, and it is now open to anyone who wants to build Vaani multilingual voice agents in the language their customers actually speak.
The scale of the bet shows up in the training data. The company reports a proprietary Indic speech corpus of 4.3 lakh hours. A lakh is 100,000, so that is 430,000 hours of speech.
Divide by 24 and you get about 17,917 days of continuous audio, which is roughly 49 years of someone talking without stopping. That corpus is the reason the platform can claim realistic voices with emotional tags and automatic switching between languages and dialects, including switching mid-sentence, which is how many multilingual speakers actually talk.
The US-audience angle is not the geography; it is the pattern. The United States is itself multilingual, and so is every global customer base.
A support agent that only works well in English is a support agent that fails a large share of your callers. Vaani's pitch is that the language handling is built into the speech models themselves, not bolted on as a translation step, so the agent keeps the caller's accent, dialect, and switching habits instead of flattening them.
How Much Does the Vaani Developer Platform Cost?
The Vaani Developer Platform has two plans: pay as you go, and Enterprise. Pay as you go is usage-based billing from a prepaid wallet: you top up, choose your voice, models, and carrier, and pay for what you use. Enterprise is a custom commercial setup with negotiated terms, and it adds whitelabel and reseller options, custom carrier arrangements, and sub-account rates with credit allocation.
The platform fee on pay as you go is 1 rupee per minute. Everything else is a component you choose, and the pricing page lets you build your own stack with an estimator. The worked example on the page uses Deepgram nova-3 for speech recognition, Google gemini-3.5-flash for reasoning, Cartesia sonic-3.5 for voice synthesis, and Vobiz for telephony, and it prices 1,000 calls at 3 minutes each at 16,975.20 rupees per month, or 5.66 rupees per minute.
| Component (estimator example) | Rate | Share of the 5.66 rupees per minute |
|---|---|---|
| Voice synthesis, Cartesia sonic-3.5 | 3.15 rupees per minute | About 56 percent, the biggest single cost |
| Platform fee | 1.00 rupee per minute | About 18 percent |
| AI reasoning, Google gemini-3.5-flash | 0.66 rupees per minute | About 12 percent |
| Telephony, Vobiz | 0.45 rupees per minute | About 8 percent |
| Speech recognition, Deepgram nova-3 | 0.40 rupees per minute | About 7 percent |
The arithmetic is worth doing yourself, because the voice is where the money goes. At 5.66 rupees per minute, a 3-minute call costs about 16.98 rupees, and the voice synthesis alone is more than half of it.
Swap the voice for a cheaper one and the whole call gets cheaper; swap the reasoning model and the saving is small. That is also why the platform lets you bring your own provider keys: for any supported component connected with your own key, that component is not charged through Vaani, and you pay the provider directly under your own agreement.
| Pay as you go | Enterprise |
|---|---|
| Platform fee 1 rupee per minute | Custom platform fee, negotiated |
| Unlimited agents, knowledge bases, and daily and hourly calling | Unlimited agents, knowledge bases, and daily and hourly calling |
| 2 concurrent calls | Custom concurrency |
| Phone number 600 rupees per month | Custom number rates |
| Extra concurrency 2,000 rupees per 5 lines per month | Custom concurrency rates |
| 99.50 percent uptime SLA | 99.50 percent uptime SLA |
| Data retention 30 days | Zero data retention |
| Support SLA 3 days | Priority support SLA 6 hours, dedicated account manager |
| Warm transfers free as an early launch offer | Warm transfers included |
| HIPAA compliance not included | HIPAA compliance included as a launch offer |
Two limits in that table deserve a second look. First, 2 concurrent calls on pay as you go means at most two calls at once, so at 3 minutes per call the plan tops out at about 40 calls per hour, and extra lines cost 2,000 rupees per 5 lines per month, or 400 rupees per line per month.
Second, the billing pulse matters: Vaani supports per-second, 30-second, and per-minute billing depending on your offering, and the estimator does not model pulse rounding, so a 61-second call can cost double a 60-second one on per-minute billing. Confirm the pulse on your account before comparing rates.
Vaani vs Vapi: How Do They Compare?
Vapi is the best-known alternative in this space. Vapi is a developer-first voice AI platform for building phone agents: you choose your own transcriber, language model, and voice provider, and Vapi orchestrates them in real time.
It reports more than 1 billion calls handled, 2.5 million agents launched, and over 1 million developers, and it charges a 5-cent-per-minute base fee plus every provider at cost. Its visual builder was retired in August 2026, so building on Vapi now means APIs and code.
| Vaani Developer Platform | Vapi | |
|---|---|---|
| What you build with | Teacher AI: describe the agent in plain language | APIs and SDKs: pick each component yourself |
| Speech models | Vaani's own foundation speech models | Your choice: Deepgram, AssemblyAI, OpenAI, and others |
| Multilingual focus | Built in, with a 430,000-hour Indic speech corpus | Language support depends on the providers you pick |
| Platform fee | 1 rupee per minute, plus components at their rates | 5 US cents per minute, plus providers at cost |
| Bring your own keys | Yes, for supported components | Yes, the whole model is bring-your-own |
| Latency claim | About 450 milliseconds at p90 | Under 500 milliseconds infrastructure latency |
| Scale reported | More than 14 enterprise deployments | More than 1 billion calls handled |
| Best for | Teams that want a working agent fast, in many languages | Engineering teams that want total control of the stack |
The honest way to read that table is as a trade-off, not a ranking. Vapi hands you the parts and the freedom to tune every layer, which is why engineering teams that treat their voice stack as a product choose it. Vaani hands you a finished agent built from a sentence, with the speech models and the multilingual handling already inside.
Retell AI is the third name in this space: it positions itself as the faster-to-launch option, with a no-code builder and sub-500-millisecond latency out of the box, at a 7-cent-per-minute base rate. The three platforms share the same shape, a transcriber plus a model plus a voice, and they differ in how much assembly they ask you to do.
Vaani vs ElevenLabs: Platform or Voice API?
ElevenLabs is the published sibling of this topic, and the comparison is worth exactly one careful paragraph. ElevenLabs is a speech synthesis API: it turns text into a realistic voice, and it does that job very well.
Our guide on how to use ElevenLabs shows how to make your first voiceover with it, including its v3 audio tags and its free-credit line. Vaani is a different kind of product: a full platform that includes speech, telephony, orchestration, memory, tools, and human handoff.
The two are complements, not rivals. A team could use ElevenLabs as the voice inside a Vapi stack, and a team could use Vaani as the whole stack. The question to ask is not which one is better, but which layer you want to own: if you want the voice and nothing else, an API is the right tool, and if you want an agent that answers the phone, takes the order, and remembers the customer, a platform is the right tool.
Who Should Use the Vaani Developer Platform?
The platform fits three groups. First, startups and product teams that need a voice agent this week, not next quarter, and do not have a voice engineer on staff. Teacher AI is built for them: describe the agent, answer the questions, review the diff, and deploy.
Second, enterprises with specialised requirements, which Vaani serves with a forward-deployed engineering team for integrations and workflow optimisation. Third, teams whose customers do not primarily speak English, because that is the problem the company's speech models were built to solve.
The use cases on the vendor's homepage are concrete rather than generic. One case study shows an agent doing automated first contact for a compliance service: it qualifies founders who need compliance help, in their own language, and hands the qualified conversation to a compliance expert with context, instead of handing over raw registration records. The same shape fits support lines, appointment booking, order tracking, lead qualification, and payment reminders, anywhere the caller would otherwise wait in a queue or repeat themselves to three different people.
The customer logos on the homepage include Uber, Accenture, Randstad, SBI Life, Bajaj Finserv, Krafton, and Mphasis, alongside Indian enterprises and startups. Logos on a homepage are not proof of a deployment, and the press release's own number is more careful: more than 14 enterprise deployments before the self-serve launch. Treat the logos as a signal that large companies have looked at the platform, and the deployment count as the number the company itself is willing to stand behind.
What Are the Limits and Open Questions?
The platform is seven days old at the time of writing, and that youth is the biggest limit. Everything in this guide comes from the vendor's own pages and its launch announcement, because no independent review, tutorial, or test existed when this article was written. The latency figure, the readiness score, the memory behaviour, and the call quality are all vendor-published claims, not independently confirmed results.
The pay-as-you-go limits are concrete and worth knowing before you start. Two concurrent calls is a hard ceiling for a busy support line, and extra lines cost 400 rupees each per month.
Data is retained for 30 days on pay as you go, against zero data retention on Enterprise, which matters if your industry has its own retention rules. The uptime SLA is 99.50 percent on both plans, which allows about 3 hours and 36 minutes of downtime in a 30-day month, so this is not a five-nines telephony promise.
Pricing is published in rupees, and the estimator's exchange-rate assumption turns it into dollars only for the illustration. A US team should budget in rupees and watch the conversion, and should confirm the billing pulse on its own account, since per-minute rounding can double the cost of short calls.
The early launch offers, free warm transfers, free self-learning voice harness, and free agents with memory, are explicitly launch offers, so they can end. The backers named in coverage are Venture Catalysts and Google AI Accelerator, with no funding amount disclosed, so do not quote a valuation or a round size for this company.
How Do You Build Your First Vaani Agent?
Start on the vaani.dev platform, which lives at app.vaanivoice.ai. Sign up, top up the prepaid wallet, and describe your first agent in one sentence: who it should call, what it should say, and what counts as success. Pick a narrow job for the first agent, such as answering order-status questions or qualifying inbound leads, because a narrow agent is easier to test and easier to trust.
Answer Teacher AI's clarifying questions carefully, because those answers become the agent's behaviour. Set the languages your callers actually use, name the objections the agent will hear, and define the handoff rule: the exact moment the agent should transfer to a human instead of guessing. Review the prompt diff before you accept it, and check the recommended latency, voice, and turn-taking settings against what your callers will tolerate.
Test with real phone calls, not just the web preview. Call the agent yourself, interrupt it mid-sentence, switch languages, ask something it was not trained for, and stay silent for ten seconds to see what it does with dead air.
Check the call logs, the memory it kept, and the variables it extracted, then refine by chat and review the next diff. If you want to automate the testing loop with code, our guide on how to use Claude Code shows how to run a coding agent that can script those test calls for you.
Sources
Vaani Developer Platform homepage, vaani.dev, fetched November 5, 2026. The source for the voice harness description, the Teacher AI build flow, the memory example, the Product Hunt number-one badge, and the p90 latency claim of about 450 milliseconds.
Vaani pricing page, vaani.dev/pricing, fetched November 5, 2026. The source for the pay-as-you-go and Enterprise plans, the 1-rupee-per-minute platform fee, the 600-rupee phone number, the 2,000-rupee concurrency lines, the 99.50 percent uptime SLA, the 30-day retention, and the estimator example of 16,975.20 rupees per month.
CIO and Leader, "Vaani AI Research Launches Self-Serve Platform to Build and Deploy Voice AI Agents," published October 29, 2026. The source for the launch date, the harness components, the enterprise background, and the CEO's statement that seventy-five percent of the world does not primarily speak English.
ANI press release via Business Standard, "Vaani AI Research will launch Self-Serve Platform to Build and Deploy Voice AI Agents on 29th Oct," published October 28, 2026. The source for the founding year of 2025, the founders' names, the Bengaluru and Singapore base, and the vaani.dev address.
YourStory, "Startup news and updates: daily roundup (October 29, 2026)," published October 29, 2026. The source for the launch summary, the inference optimisation and conversational controls, and the enterprise deployment background.
AI Startup Impact newsletter, fetched via search on November 5, 2026. The source for the figure of more than 14 enterprise deployments, the proprietary Indic speech corpus of 4.3 lakh hours, and the backers Venture Catalysts and Google AI Accelerator, with no funding amount disclosed.
Lindy, "What is Vapi AI? Review, Pricing and Real Call Costs in 2026," fetched November 5, 2026. The source for Vapi's developer-first positioning, the 5-cent-per-minute base fee plus providers at cost, the retirement of its visual builder in August 2026, and the vendor-reported scale of more than 1 billion calls and 2.5 million agents.
Frequently Asked Questions
Is the Vaani Developer Platform free?
No free tier is listed on the pricing page as of November 5, 2026. Pay as you go is usage-based billing from a prepaid wallet, with a platform fee of 1 rupee per minute plus the components you choose, and a phone number costs 600 rupees per month. Some features are free as early launch offers, including warm transfers and agents with memory, but those offers can end.
What languages does the Vaani Developer Platform support?
The platform is built for multilingual voice agents, with an Indic speech corpus of 4.3 lakh hours behind it, and its agents switch language and dialect automatically, including mid-sentence. The vendor's pages do not publish a full list of supported languages as of November 5, 2026, so check the documentation or ask the team before promising a specific language to your customers.
How is the Vaani Developer Platform different from Vapi?
Vaani builds the agent for you from a plain-language description, using its own speech models and a built-in multilingual focus, with a 1-rupee-per-minute platform fee. Vapi gives you the building blocks and lets you pick your own transcriber, model, and voice through APIs, at a 5-cent-per-minute base fee plus providers at cost. Vaani is the faster start; Vapi is the deeper control.
Can I use my own models and phone numbers with Vaani?
Yes, for supported components. You can bring your own provider keys, and that component is then billed by the provider directly instead of through Vaani. Enterprise plans add custom carrier arrangements, and the pricing page also offers whitelabel and reseller options for teams that want to offer Vaani under their own brand.
Do I need to host anything myself to run a Vaani agent?
No. The voice harness runs the whole stack, including the speech models, telephony, orchestration, memory, and tools, so there is nothing to install or host. Enterprise customers can ask for on-premises or VPC deployment, and a dedicated orchestration server is available on demand, but the self-serve platform is fully hosted.
