←→ navigate · Space next
New hire orientation
vapi
01 / 07 · what is vapi
Foundations

What is Vapi?

We build a voice AI platform that lets developers and enterprises ship production-grade voice agents over the phone and the web.

01
Composable pipeline BYO STT, LLM, and TTS — 30+ providers, swap any layer without rewriting the agent.
02
Real-time at scale Sub-second latency end-to-end. Interruptions, turn-taking, and backpressure handled out of the box.
03
Telephony built-in Buy or import Twilio / Vonage / Telnyx numbers. Inbound, outbound, transfers, SIP, voicemail.
04
Production primitives Tools, evals, monitoring, hooks, compliance plans (HIPAA, SOC2, PCI), per-call cost tracking.
→  VAPI = Voice API. Pronounced like "happy."
02 / 07 · agents
Primitives

The four primitives.

Build any voice agent from these.
01 / ASSISTANT
Assistants
A single configurable voice agent — model + voice + transcriber + prompt + tools — that handles an entire call end-to-end.
  • Tune speaking plans, voicemail detection, analysis
  • Attach tools, webhooks, compliance, observability
Pick when one persona owns the call.
02 / SQUAD
Squads
A team of assistants on one call, with handoffs so each specialist owns its slice of the conversation.
  • State and tool history carry across handoffs
  • The customer never has to repeat themselves
Pick when specialization beats generalization.
03 / TOOLS
Tools
Function-call capabilities the assistant invokes mid-conversation to take actions or fetch data.
  • 20+ built-in: function, apiRequest, transfer, end, dtmf, sms, MCP
  • Plus Google Calendar / Sheets, Slack, GHL, Make
Pick when the agent needs to do something.
04 / WORKFLOW
Workflows
A visual graph of nodes and edges that scripts deterministic conversation paths with branching logic.
  • Conversation, Tool, Hangup nodes connected by edges
  • Per-node model/voice/transcriber overrides
Pick when the flow is structured and rules-driven.
Mental model  ·  Assistant = one agent.   Squad = many agents.   Tools = what an agent can do.   Workflow = a scripted graph instead of an LLM-driven assistant.
03 / 07 · agents · update
Heads up

The four primitives.

Build any voice agent from these.
01 / ASSISTANT
Assistants
A single configurable voice agent — model + voice + transcriber + prompt + tools — that handles an entire call end-to-end.
  • Tune speaking plans, voicemail detection, analysis
  • Attach tools, webhooks, compliance, observability
Pick when one persona owns the call.
02 / SQUAD
Squads
A team of assistants on one call, with handoffs so each specialist owns its slice of the conversation.
  • State and tool history carry across handoffs
  • The customer never has to repeat themselves
Pick when specialization beats generalization.
03 / TOOLS
Tools
Function-call capabilities the assistant invokes mid-conversation to take actions or fetch data.
  • 20+ built-in: function, apiRequest, transfer, end, dtmf, sms, MCP
  • Plus Google Calendar / Sheets, Slack, GHL, Make
Pick when the agent needs to do something.
04 / WORKFLOW
Workflows
A visual graph of nodes and edges that scripts deterministic conversation paths with branching logic.
  • Conversation, Tool, Hangup nodes connected by edges
  • Per-node model/voice/transcriber overrides
Pick when the flow is structured and rules-driven.
DEPRECATED
Heads up  ·  Workflows is no longer maintained. For production builds use Assistants or Squads with branching tools instead.
04 / 07 · platform
Beyond agents

The platform around the agent.

Ship, monitor, test, iterate.
MONITORING
Call Logs & Analytics
Full transcripts, recordings, per-turn cost & latency, plus alert Monitors that fire on structured-output thresholds.
SIMULATIONS
Synthetic Calls
Compose personalities + scenarios to exercise an assistant before it ships. Batch-run with concurrency for regression testing.
EVALS
Test Suites & Evals
Live-call regression Test Suites plus AI-judge Evals over mock transcripts. CI-style gating for every model or prompt change.
TELEPHONY
Phone Numbers
Buy Vapi-managed numbers or import from Twilio, Vonage, Telnyx. SIP, transfers, voicemail detection — all native.
CAMPAIGNS
Outbound at Scale
Bulk outbound dialing with schedules, dial plans, retries. Jobs enqueued to BullMQ and processed by call workers.
KNOWLEDGE
Files & Knowledge Bases
RAG via vector DB + uploaded documents injected into the assistant context at runtime. Plus structured-output extraction from transcripts.
05 / 07 · why vapi
Differentiators

What makes us different.

Three names come up most. Each makes a tradeoff we don't.
Sierra · Decagon
Too closed.

A black box built around one use case. You don't control the pipeline, you don't pick the providers, and you can't push it past the lane it ships in.

LiveKit
Too open.

Bare infrastructure. Every knob is yours — and so is the obligation to build and maintain the platform. Infrastructure is the hardest part of voice AI, and with LiveKit it's all on your team.

ElevenLabs
Too locked-in.

Similar agents platform, but tied to ElevenLabs' own models. The industry moves faster than any one company — pick them and you move at their pace.

Vapi is the middle path: the pipeline and the knobs, the infrastructure handled, and the providers wide open.

06 / 07 · surfaces
Dashboard / API

Dashboard  vs  API.

Two surfaces. One platform.

Dashboard The on-ramp

40+ feature surfaces. Most users ship production agents without ever opening a REST client.

  • ▸
    Assistant builder — model, voice, prompt, tools
  • ▸
    Squads & workflows — visual graph editor
  • ▸
    Phone numbers & campaigns — buy, import, dial
  • ▸
    Logs, monitors, evals, simulations — observe, test, iterate
  • ▸
    Files, knowledge bases, billing — full self-serve

API The escape hatch

Long-tail advanced configuration. Where the dashboard stops, the API keeps going.

  • ▸
    Hooks — 7 lifecycle events (interrupt, timeout, time elapsed) → transfer, function, say, tool actions
  • ▸
    Pronunciation libraries — ElevenLabs & Cartesia custom phonemes ("Jira" → "JEE-rah")
  • ▸
    Assistant request — server-side dynamic assistant per call via your server.url
  • ▸
    Fallback model chains, custom-LLM endpoints, server-message subscriptions
Build it in the dashboard.  Unlock it with the API.
07 / 07 · skills
Agentic build

Vapi Skills.

The agentic way to build Vapi agents.

Skills are scoped instruction packs that let Claude Code drive the Vapi API in plain English. Describe the agent you want; the skill picks the model, voice, and transcriber, writes the prompt, wires the tools, and creates the resource.

you
"Make me a triage bot for an HVAC company on Cartesia + GPT-5."
→
/create-assistant
creates an Assistant · picks model/voice/transcriber · writes prompt
you
"Wire up a sales-to-support handoff."
→
/create-squad
spins up a Squad · stitches transfer tools · routes context across members
you
"Add a Google Calendar booking tool with timezone handling."
→
/create-tool
scaffolds a Tool · function/API/transfer/end-call types · auth + schemas
you
"Launch 500 outbound calls Tuesday at 9am Pacific."
→
/create-call · /create-phone-number
configures the Campaign · imports/buys the number · schedules dial-out
Stop building agents.  Describe them.
You're in.

Welcome.

docs.vapi.ai vapi.ai https://github.com/VapiAI/skills #ask-eng