Building your own voice agent out of STT, an LLM and TTS? Here's what that costs you.

VARTA is the runtime underneath that pipeline, not another layer on top of it: a versioned HTTP API plus runnable examples, the same core that runs humanise.ai as a live, multi-tenant product today.

Cost: a bounded prompt, not a growing one

A pipeline you hand-roll typically resends a growing prompt every turn and re-synthesises every reply, even the ones a caller has said a hundred times before. VARTA keeps a bounded per-turn prompt instead of appending the whole conversation, and replays cached audio for repeated phrasing rather than calling the TTS provider again -- the same cached-playback path humanise.ai runs calls through today. No figure is quoted here; the cost calculator is where that shape gets modelled against your own call volume, and its own disclaimers are what govern any number it shows you.

Determinism: the L1-L7 trace decides, not the model

A hand-rolled agent typically asks an LLM to decide the next step on every turn, which means the same input can take a different path depending on what the model felt like that time. VARTA routes each turn through a deterministic rule engine -- seven explicit layers, L1 through L7 -- and only escalates to the model for the turns that genuinely need reasoning; the model is one layer, not the controller. Every turn returns which layer decided and what it cost, which is what makes a call flow testable at all. See how the engine works for the full architecture.

Time to ship: the hard parts are already built

Everything past the "hello world" call is where a hand-rolled pipeline stalls for months: barge-in that doesn't clip the caller, endpointing that doesn't cut people off mid-sentence, letting a caller correct or revisit something they already said mid-call instead of restarting the flow, standing up STT that copes with real multilingual mishearing rather than a clean transcript, and getting a call onto the phone network at all. VARTA ships all of it: a drift handler with named REPEAT and REVISIT paths for exactly that mid-call correction, a phonetic-confusion taxonomy for multilingual STT noise, and SIP/telephony trunking alongside the LiveKit-based voice worker for browser and phone calls alike.

Build vs buy

If the answer to build-vs-buy is "buy, but we still need to run it ourselves," VARTA ships as a self-hosted licence -- four Docker services (workflow engine, the LiveKit-based voice worker, the designer console, PostgreSQL) you deploy on your own infrastructure, so data residency is a decision you make rather than a claim you take on faith. Because it's your infrastructure, it's also the path for an integrator who wants to offer this under their own brand rather than VARTA's. See the enterprise deployment page for the full security and deployment posture. VARTA Cloud -- a shared, hosted endpoint you don't have to run yourself -- isn't open yet; the waitlist is the way in when it is.

The API reference, authentication, idempotency and error catalogue all live in the docs. For what teams build with it end to end, see use cases.