Vapi

Vapi is a developer-first platform for building sophisticated voice AI agents that handle real-time telephone and web-based conversations. Unlike no-code voice builders, Vapi provides a programmable infrastructure where developers assemble their own stack from modular components - choosing their preferred speech-to-text (Deepgram, AssemblyAI), language models (OpenAI, Anthropic, custom endpoints), text-to-speech (ElevenLabs, PlayHT), and telephony providers (Twilio, Vonage, SIP).

The core abstraction is the 'assistant' - a JSON configuration that defines conversation behavior, system prompts, interruption handling, and tool access. This declarative approach means creating voice agents without writing audio processing code, but still maintaining granular control over every component. Advanced features like 'squads' allow multiple specialized assistants to collaborate within a single call, passing context between specialists for complex multi-step workflows.

Vapi excels at real-time conversation orchestration through sophisticated endpointing (detecting when users finish speaking), barge-in handling (allowing natural interruptions), backchanneling (inserting acknowledgments during processing delays), and latency optimization. The platform streams LLM responses to TTS while simultaneously processing new user input, achieving typical response times of 500-800ms.

Function calling during live conversations is Vapi's standout capability. When the LLM decides to invoke a tool (check inventory, update CRM, process payment), Vapi makes server-side webhook calls to your APIs, waits for responses, and seamlessly continues the conversation. This enables voice agents that perform real business actions, not just chitchat.

Telephony support includes inbound/outbound calling via SIP trunking, Twilio, and Vonage, primarily serving US and Canadian numbers. WebRTC integration enables browser-based voice for web applications. The platform supports batch outbound campaigns for sales, surveys, and customer outreach.

Recent 2026 updates include custom LLM backend support (including self-hosted models via OpenAI-compatible APIs), hallucination testing suites for pre-deployment quality assurance, global edge deployment for reduced latency, parallel tool execution during conversations, expanded language support for 20+ languages, and enterprise HIPAA compliance options.

Pricing follows a platform-plus-passthrough model: $0.05/minute orchestration fee plus underlying provider costs (STT, LLM, TTS, telephony). Typical all-in costs range from $0.07-$0.33/minute depending on component selection. Higher-quality voices and reasoning-heavy models push toward the upper range, while optimized setups can achieve lower costs. The free tier includes $10 in testing credits.

Vapi suits developer teams building voice-centric products who need deep customization and don't mind managing technical complexity. It's less suitable for non-technical teams seeking guided no-code experiences or organizations needing multi-channel (voice + chat + email) automation in a single platform.