Vapi

Vapi is a developer platform for building, testing, and deploying AI-powered voice agents — conversational AI systems that communicate through spoken language over phone calls and web-based audio. It provides the complete infrastructure stack for voice AI: telephony integration, speech-to-text, LLM orchestration, text-to-speech, and real-time audio streaming, abstracted behind APIs that let developers focus on conversation logic rather than voice infrastructure.

The core abstraction in Vapi is the "assistant" — a configuration that combines an LLM (OpenAI, Anthropic, open-source models via providers), a voice (ElevenLabs, PlayHT, Deepgram, Azure voices), a transcription engine, and a set of tools the agent can invoke during calls. Assistants are defined via JSON configuration or the dashboard, specifying the system prompt, voice settings, interruption handling behavior, silence detection thresholds, and response timing parameters. This declarative approach means you can create and modify voice agents without writing audio processing code.

Vapi handles the real-time complexities that make voice AI challenging: turn-taking (detecting when the user has finished speaking), interruption handling (allowing users to cut in mid-sentence), background noise filtering, endpointing (deciding when to process partial speech), and latency optimization (streaming LLM responses to TTS for faster perceived response times). These features are configurable per assistant, letting developers tune the conversation dynamics for different use cases.

For agent tool use, Vapi supports function calling during voice conversations. When the LLM decides to invoke a tool (check a calendar, look up a database, transfer a call), Vapi makes a server-side webhook to your API, waits for the response, and feeds it back to the LLM for continued conversation. This enables voice agents that can actually do things — book appointments, process orders, transfer calls, access CRM data — not just chat.

Telephony integration supports inbound and outbound calling via SIP trunking, Twilio, and Vonage. Web-based voice uses WebRTC for browser integration. The API supports batch outbound calling campaigns for sales, appointment reminders, and surveys. Call recordings, transcripts, and analytics are available through the dashboard and API.

Pricing is per-minute based on the components used (LLM, voice, telephony). The free tier includes a small credit for testing. Key considerations include the inherent latency in voice AI pipelines (typically 1-3 seconds for response generation), cost per minute that can exceed traditional IVR systems, and the complexity of debugging real-time voice interactions compared to text-based agents.