From npm install to a Live Call: Why Telenow AI Is Built Developer-First

Ask a developer who has shipped a voice agent what the first three weeks looked like, and you'll get some version of the same answer: telephony configuration, a WebSocket server, a VAD loop, an interruption handler that half works, and a state machine nobody wanted to write.

Three weeks in, they've heard maybe four calls. The product hasn't started.

That's the sequence Telenow is designed to break. Not by hiding the stack behind a no-code wall — developers hate that, and rightly — but by handing you the primitives with the plumbing already solved.

Minute one: a real call, in the browser

Here's the entire client-side surface for a live voice conversation:

javascriptimport { TelenowCall } from '@telenow/client';

const call = new TelenowCall({ publicSlug: 'YOUR_AGENT_SLUG', // from the agent's Publish tab baseUrl: 'https://api.telenow.ai', onTranscript: (line) => console.log(${line.role}: ${line.text}), onState: (state) => console.log('call:', state), });

await call.start(); // prompts for mic, connects, the agent greets // call.stop(); // hang up

npm i @telenow/client — zero dependencies, runs in any browser.

That's mic in, voice out, live transcript. No media server. No signalling code. No ngrok tunnel.

And dialling a real phone from your backend is one request:

javascriptimport { Telenow } from '@telenow/server'; const telenow = new Telenow({ apiKey: process.env.TELENOW_API_KEY });

const call = await telenow.calls.create({ agentId: 'agent_123', to: '+919876543210', }); console.log('calling:', call.sessionId);

Four channels — talk in your browser, dial our number, have the agent call your phone, or trigger an outbound call from your own backend — and each one is a handful of lines. You can try all four in the playground with no signup at all.

The point isn't that the snippet is short. It's that you hear a call on day one instead of week three — and voice quality is a thing you improve by listening, not by reading transcripts.

The parts we solved so you don't rebuild them

Underneath that SDK call is the work that eats most voice AI projects:

Barge-in that actually works. A Rust real-time core with configurable interruption handling and sub-300ms STT partials. Callers can cut the agent off mid-sentence and the agent stops — and, critically, the conversation history is truncated to what the caller actually heard, not what was generated. Skip that detail and every subsequent turn is built on a false record of the call.

Telephony, absorbed. Buy, port or bring numbers. Run on six CPaaS providers or a native SIP trunk, with G.711, DTMF and IP/digest auth handled. Nobody on your team opens Wireshark.

Latency you can see. Per-call STT, LLM, TTS and server latency are measured and exposed. When a call feels sluggish, you get a number, not a hunch.

Bring your own brain

This is the part most platforms won't give you.

Telenow doesn't have a house model you're stuck with. Plug in OpenAI, Anthropic, Gemini, Groq, Bedrock or OpenRouter — or any OpenAI-compatible endpoint, including one you host yourself.

And if your agent already exists as a service — your own orchestration, your own retrieval, your own tool-calling logic — you can point Telenow at it as a full custom agentic backend. Telenow runs the real-time voice pipeline; your backend does the thinking. Your model, your prompts, your evals, your keys, billed directly by your provider.

Same story for the carrier. BYOC on your own Twilio, Plivo, Vonage, Exotel or Vobiz account, or a native SIP trunk. Your trunks, your negotiated rates — not somebody else's margin quietly baked into a bundled minute.

Zero lock-in isn't a slogan here; it's the architecture.

Tools, MCP, and reading your real systems mid-call

An agent that can only talk is a demo. An agent that can act is a product.

Telenow agents can call any HTTPS endpoint as a function-calling tool, or connect to any MCP server — GitHub, Linear, Jira, Notion, Asana, Sentry — plus 50+ native integrations across CRMs (HubSpot, Salesforce, Zoho, Intercom), calendars (Google Calendar, Calendly), Sheets and Airtable, payments (Razorpay, Stripe, PayPal), forms, support and automation (Zapier, Make, n8n).

Alongside that:

Knowledge bases (RAG) — upload PDF/DOCX/CSV or point at a URL; the agent retrieves and cites your content mid-conversation. Caller memory — repeat callers are recognised across calls, with a rolling summary and a facts ledger, with a TTL you control. Flows and code — branch visually when the logic is simple, drop into a sandboxed JS code node when it isn't.

The events layer

Voice agents are asynchronous systems, and they need to be treated like one.

Signed HMAC-SHA256 webhooks fire on call.ended, recording.ready, transcripts and tool calls. There's a WebSocket web-call API and a text Chat API for any surface you want the same agent to appear on. Everything provisionable through a full REST API — create agents, place calls, manage the org — and five official SDKs: Web, React, React Native, Node and Python.

Which means the agent is a deployable artifact in your pipeline, not a thing someone configured in a dashboard once and nobody can reproduce.

The parts developers don't want to build, but someone has to

Here's the honest truth about voice agent projects: the API is the easy half.

The half that kills timelines is everything around the call — and shipping it yourself means building a dialer, a compliance workflow, a campaign runner, an analytics pipeline and a QA process before you can put the thing in front of a real customer.

Telenow ships that as part of the platform:

Outbound campaigns — bulk CSV, 1–50 concurrency, retries, answering-machine detection, calling windows and DNC suppression, with live progress Numbers & compliance — search, buy, port or bring numbers, with in-app DLT and regulatory flows (yes, including India DLT) Warm transfer & softphone — AI answers, qualifies, resolves, then hands off to a human on the same number, with auto-resume. Your agents can dial out manually too. No silo between bot and human. Post-call analysis — auto summaries, sentiment, disposition, action items, up to 30 custom fields, QA-rubric scoring Live ops & wallboard — real-time KPIs, active calls, team presence, listen-in on any live call Spend controls — prepaid credits with a hard cut-off at zero, plus soft and hard spend caps and per-org concurrency limits

That last one matters more than it sounds. Most voice platforms let you discover your bill after the fact. Telenow's pricing is component-level — LLM + STT + TTS + telephony + platform fee, itemized in your currency — with a live calculator that shows the exact per-minute cost of your chosen stack before you build on it.

You should be able to read your own invoice. That's a low bar the industry keeps failing.

One agent, every channel

Build once. The same agent — same knowledge, same tools, same memory — shows up on:

Phone (inbound and outbound PSTN, six carriers or your own SIP) Web voice (embeddable widget, or the WebSocket / web-call API) Chat (the REST Chat API, drop it into any text surface) WhatsApp (five providers, or a built-in paired number)

Not four integrations. One agent, four surfaces.

And 30+ languages, including the ones most platforms skip

Multilingual voices across 30+ languages — including Hindi and 10+ other Indian languages via Sarvam and Smallest.ai, plus Indian-English accents, alongside global coverage through ElevenLabs, Polly, LMNT and Rime. Voice cloning via ElevenLabs.

If your callers speak Marathi or Tamil, that shouldn't be a roadmap item on someone else's backlog.

The developer loop, end to end

Build — pick a model, voice and language in the builder, or bring your own LLM or agentic backend. Every save is versioned; restore or A/B-compare any time. Connect — knowledge base, integrations, MCP servers, function-calling tools to your own APIs. Publish — buy or bring a number, or ship to web, chat and WhatsApp. Listen — monitor live calls, review post-call analytics, and iterate.

Free starter credits, no card to explore, and a live demo org that auto-expires in 24 hours if you just want to poke at it.

The thesis

Voice agents don't get good because someone architected the pipeline elegantly. They get good because somebody listened to two hundred bad calls and fixed them, one at a time.

Everything above exists to make that loop turn faster — and to make sure that when you find the fix, you're allowed to ship it, on your model, on your carrier, on your brand.

Build the product. We'll run the plumbing.

Start building → · Read the docs → · Try a live agent →

FAQs → rich results (AEO)

Click "+ Add" for each pair.

Q1. How quickly can a developer build a voice AI agent on Telenow? A1. Minutes, not weeks. A live browser voice call takes a few lines of the @telenow/client SDK, and dialling a real phone from your backend is a single server-side request. A self-serve agent on web or chat can go live in minutes; phone numbers that require regulatory approval take a little longer, and that compliance flow is guided inside the app. There are free starter credits and a no-signup playground for testing before you commit.

Q2. Can I use my own language model with Telenow? A2. Yes — Telenow has no house model you're locked into. You can use OpenAI, Anthropic, Gemini, Groq, Bedrock or OpenRouter, any OpenAI-compatible endpoint including one you host yourself, or point Telenow at a full custom agentic backend so your own service does the reasoning while Telenow runs the real-time voice pipeline. Bring-your-own model keys are billed directly by your provider.

Q3. What SDKs and APIs does Telenow offer? A3. A full REST API for provisioning agents, placing calls and managing your organization; signed HMAC-SHA256 webhooks for events like call.ended, recording.ready, transcripts and tool calls; a WebSocket web-call API and a text Chat API; and five official SDKs covering Web, React, React Native, Node and Python. Agents also support function-calling tools to any HTTPS endpoint and connections to any MCP server.

Q4. Can Telenow run on my existing telephony provider? A4. Yes. Telenow supports bring-your-own-carrier across six CPaaS providers — Twilio, Plivo, Vonage, Exotel and Vobiz — or a native SIP trunk with G.711, DTMF and IP/digest authentication. You keep your own trunks and your own negotiated rates rather than paying a vendor's margin inside a bundled per-minute price.

Q5. How does Telenow handle interruptions and latency? A5. Telenow runs a Rust real-time core with configurable barge-in, so callers can interrupt the agent mid-sentence and it stops immediately, with the conversation history truncated to what the caller actually heard. Speech-to-text partials arrive in under 300ms, and per-call STT, LLM, TTS and server latency are measured and visible to you, so slow calls produce a number to fix rather than a hunch to chase.

Q6. What does Telenow include beyond the voice API? A6. The operational layer most teams otherwise build themselves: outbound campaigns with concurrency, retries, answering-machine detection, calling windows and DNC suppression; number purchase, porting and in-app regulatory compliance including India DLT; warm transfer and a softphone so humans and AI share one line; RAG knowledge bases; caller memory across calls; post-call analysis with summaries, sentiment and QA scoring; a live wallboard; and spend caps with prepaid credits and a hard cut-off.