Vapi

What is Vapi?

Vapi is a voice AI agent development platform. Instead of being a finished customer-service chatbot or a general-purpose voice assistant, it gives developers control over the components of a conversational system: prompts, language models, speech-to-text providers, voices, phone numbers, tools, knowledge sources, call behavior, and integrations.

A typical Vapi project connects an assistant to a phone number or embeds it in a website or application. The assistant listens to speech, sends the conversation through the configured transcription and language-model providers, performs permitted actions through tools, and returns a spoken response. Teams can then inspect calls, logs, usage, and performance as they move from testing to production.

What Vapi can do

Build inbound and outbound voice agents

Vapi supports phone-based assistants for inbound reception, customer support, appointment booking, lead qualification, sales conversations, business intake, and other call workflows. Assistants can also place outbound calls where the telephony setup and use case allow it. Calls can be transferred, handed between assistants, or routed through different stages of a conversation.

Connect conversations to business actions

The main practical distinction between Vapi and a basic voice interface is its tool system. Assistants can use API requests, function tools, webhooks, client-side actions, built-in tools, and MCP tools to retrieve information or perform operations. This can include checking a calendar, updating a CRM, creating a support ticket, sending an SMS message, or triggering a workflow in an external system.

Vapi documents integrations and connection patterns involving services such as Google Calendar, Google Sheets, Slack, GoHighLevel, Zapier MCP, Make, n8n, Composio, and customer-owned provider accounts. The exact implementation depends on the tool and usually requires configuration or development work.

Use knowledge bases and structured outputs

Teams can upload supported documents and associate knowledge bases with assistants so agents can answer questions using domain-specific material. Supported input patterns include documents, PDFs, DOCX files, CSV, JSON, URLs, and API data, although the quality of an answer still depends on the source material and retrieval configuration.

Vapi can also return structured data in workflows where a voice conversation needs to produce a predictable result, such as a completed intake form, appointment details, or information for another application.

Deploy in websites and mobile applications

Vapi is not limited to telephone systems. Web, Flutter, React Native, iOS, Python, TypeScript, and CLI tooling are documented for different development and deployment workflows. This makes it possible to add voice interactions to a website or mobile product while using the same broader assistant, tool, and monitoring concepts as a phone agent.

How people realistically use Vapi

A developer normally creates an assistant in the dashboard or through the API, writes instructions, chooses model, transcription, and voice providers, and adds tools or knowledge files. The assistant is then connected to a Vapi phone number, an imported telephony setup, or an application SDK. During a call, the system transcribes the caller, generates a response, invokes configured actions when needed, and speaks the result.

Teams can begin with simulated calls and test conversations before exposing an agent to customers. Production work commonly involves reviewing logs, call recordings or transcripts where configured, costs, failures, and latency; adjusting prompts and tools; and adding safeguards for transfers, voicemail, authentication, and sensitive workflows.

This makes Vapi closer to an infrastructure and orchestration layer than to a ready-made phone system. Users who want a largely visual customer-conversation builder may prefer a product such as Voiceflow, while teams seeking workflow automation around AI systems may also evaluate n8n or Make. Vapi is most relevant when the voice agent itself is a product or a carefully integrated business process.

Providers, models, and voice configuration

Vapi supports model and provider choices from documented vendors including OpenAI, Anthropic, Google, and xAI. Speech recognition and voice options can come from providers such as Deepgram, AssemblyAI, Azure, Cartesia, ElevenLabs, Google, Gladia, OpenAI, and Speechmatics. Availability changes as provider catalogs change, and language coverage varies by the selected transcription and voice services.

Vapi also documents Vapi Voices V2 support for more than 40 languages, while speech-recognition support depends on the selected provider and model. This provider flexibility is useful for teams that need to balance latency, voice quality, language support, cost, or data-handling requirements, but it also makes configuration and troubleshooting more involved.

Pricing and access

Vapi uses usage-based billing rather than a simple fixed-seat subscription. The usage-only option has a $0 monthly package fee and includes $5 in introductory credits, but usage is still metered. Vapi lists a hosting fee of $0.05 per minute, in addition to costs for selected models, voices, transcribers, and telephony.

The entry option includes two organizations, one Vapi phone number, four concurrent calls, 14 days of raw data retention, and community and email support without an SLA. Optional success packages add paid support, capacity, governance, retention, and service commitments. The documented Core package costs $29 per month; Pro is priced at 10% of Vapi hosting fees with a $999 monthly minimum; and Premier pricing is custom. A HIPAA-eligible data-handling add-on is listed separately at $2,000 per month.

Because several parts of a call can be billed separately, total cost is difficult to estimate from the platform fee alone. Model choice, voice provider, transcription, telephony, call duration, concurrency, and integration requirements all affect the final bill.

Important limitations and privacy considerations

Vapi requires meaningful technical setup for production use. Building a reliable agent involves prompt design, provider selection, tool integration, error handling, testing, telephony configuration, and monitoring. The usage-only plan has limited concurrency and retention, and free Vapi phone numbers use US area codes and support US national inbound calling; international or outbound scenarios may require another number arrangement.

Provider capabilities are not uniform. Voice quality, language support, latency, pricing, and data policies can vary across the chosen model, transcriber, voice service, and telephony provider. Vapi also does not provide on-premise deployment according to its FAQ.

Vapi processes account, configuration, call, audio, transcript, usage, and related service data to operate the platform. Its public materials describe configurable retention and zero-data-retention options for qualifying packages. Usage-only retention is listed as 14 days, Core as 30 days, and Pro as 180 days, with Premier configured separately. Organizations handling sensitive conversations should review Vapi's terms, data-processing materials, selected provider policies, and package-specific ZDR settings rather than assuming one universal training or retention policy.

Who should use Vapi?

Vapi is a good fit for developers and product teams building custom voice agents that must connect to telephony, business APIs, knowledge bases, calendars, CRMs, or internal workflows. It is also relevant to agencies and startups that need to deploy multiple specialized assistants or embed voice into their own products.

It is less suitable for someone who wants a consumer voice assistant, a no-configuration chatbot, a standalone transcription service, or a simple drag-and-drop phone system. Users who need a prebuilt conversational product rather than an extensible development platform may find the setup burden and variable usage costs disproportionate.

Bottom line

Vapi's strength is the control it gives developers over the full voice-agent stack: models, speech recognition, voices, phone infrastructure, tools, knowledge, deployment surfaces, and monitoring. That makes it useful for production-oriented phone and embedded voice applications. The trade-off is complexity. Teams must design the workflow, manage provider dependencies, test the agent carefully, and account for several usage-based costs and data-handling policies.