Vapi.
Developer voice-AI platform for phone agents — Twilio-meets-OpenAI in one stack.
What it is.
San Francisco voice-AI infrastructure company. Vapi wires telephony, speech-to-text, an LLM, and text-to-speech into a single API call, letting engineers ship a phone agent in an afternoon. Y Combinator W23. The platform is model-agnostic — buyers route to OpenAI, Anthropic, or Deepgram on the same account.
Where it fits.
Engineering teams that want to own the agent logic but not the telephony plumbing. Series A and B SaaS companies adding a voice channel without hiring a contact-center vendor. Outbound sales tooling startups built entirely on top of Vapi. Bring-your-own-Twilio is supported for buyers who already have carrier contracts.
- Model-agnostic — swap LLM and TTS without rewriting the agent
- Sub-second turn latency on Cartesia and Deepgram paths
- Transparent per-minute pricing under existing Twilio contracts
- Per-minute pricing compounds fast at contact-center scale
- Less polished no-code builder than Synthflow for non-engineers
Frequently asked.
Can I use my own Twilio account?
Yes. Bring-your-own-Twilio is supported and removes the telephony markup. Most production deployments above 50,000 minutes a month run on BYOT.
What is the typical end-to-end latency?
650 to 900 milliseconds on the Deepgram-Cartesia path. Higher when routed through OpenAI Realtime; lower when models and TTS share a region.
Does Vapi handle HIPAA workloads?
Yes on the Enterprise tier with BAA. The default tier is not HIPAA-eligible — healthcare buyers need the dedicated contract.
