Vapi

Vapi is a developer-facing voice AI agent orchestration platform that assembles speech recognition, large models, speech synthesis and phone lines into a voice agent you can put into production. The value is not in model weights but in orchestration, latency and reliability engineering.

Assistants and Squads Assistants use a single prompt plus tools to cover customer support, lead qualification and booking; Squads orchestrate multiple specialist assistants with context-preserving handoffs, which suits multi-role flows such as triage-to-booking and order returns.

Phone and web channels Inbound and outbound calls, number management and bring-your-own SIP trunks (BYO SIP) are supported, and the Web SDK embeds real-time voice into web pages and apps.

Pluggable three-stage pipeline For STT you can pick Deepgram, AssemblyAI or Soniox, for the LLM OpenAI, Anthropic or Gemini, and for TTS ElevenLabs or Cartesia; you can also plug in custom STT and a self-hosted OpenAI-compatible LLM, and each of the three stages has its own independent fallback.

Tools and observability Built-in transfer, hangup, SMS, DTMF and API Request tools, plus MCP and knowledge base support. Evals runs automated evaluations wired into CI/CD, Simulations has AI testers hold real conversations, Monitoring alerts on thresholds, and logs expose the cost of an individual call.

Configuration as code assistants, squads and tools live as files in git, go through PR review and are promoted across environments, backed by a CLI, REST API and Composer.

Where the AI core ends Vapi does not build its own foundation STT, LLM or TTS models; the company describes itself as an "orchestration layer over the transcriber, the model, and the voice". What Vapi does build is the real-time orchestration and low-latency path, three-tier fallback, the Model Intelligence benchmark, Evals and the SIP abstraction layer. The upstream models behind Vapi Voices are not publicly disclosed. On the enterprise side Vapi publishes SOC 2, HIPAA and PCI DSS compliance along with ZDR and SSO.