Vapi

What is Vapi?

Vapi is an infrastructure platform for Voice AI voice assistants, designed for developers. It combines real-time speech recognition, large language models, speech synthesis, phone functionality, tool invocation, and low-latency orchestration to help teams create phone or in-app assistants that can engage in natural conversations.

The platform features a modular design; developers can combine services such as OpenAI, Anthropic, Deepgram, and ElevenLabs, as well as integrate their own models and API keys. Vapi is responsible for real-time orchestration, reliability, phone connections, testing, monitoring, and scaling.

Main functions of Vapi

• Voice assistant:Configure prompts, models, voice, transcription, and dialogue behaviors.
• Incoming and outgoing calls:Purchase numbers, import numbers, or connect to phone services to enable two-way calls.
• Workflows and Squads:Design multi-step processes and transfer conversations among multiple dedicated assistants.
• Tool invocation:Connects scheduling, CRM, databases, payment systems, and custom business APIs.
• Server URL and Webhook:Receive call events, tool requests, status, and analysis data in real time.
• Campaigns：Batch manage outbound call lists, scheduling, and call activities.
• Testing and monitoring:View call logs, transcripts, delays, errors, and Webhook deliveries.
• Web and mobile SDKs:Integrate voice assistants into web pages, as well as applications built with React Native, Flutter, and other frameworks.
• Multi-tenant and white-label:Create and manage voice capabilities for SaaS customers through APIs.

Voice AI technology pipeline

Vapi prices

The total cost of Vapi consists of the platform hosting fee, the fees charged by the model providers, and the telephone charges. Looking only at the platform fee of $0.05 per minute does not give an accurate picture of the overall cost of a single phone call.

Scale adds support for SOC 2, PCI, SSO, RBAC, data residency, SLAs, dedicated support, and priority access. Customers with high usage levels can negotiate a discount per minute; the actual contract should specify the platform fees, committed usage levels, overage charges, and the supplier’s costs.

The cost components of a single phone call

Tutorial on creating a Vapi voice assistant

1. First, define a clearly defined telephone goal, the tasks that can be handled, and the conditions under which calls should be transferred to a human operator.
2. Create an Assistant, select STT, LLM, TTS, and define the identity, tone, and business rules.
3. Configure the first sentence, interruptions, silence periods, end time, voice mail, and maximum call duration.
4. Add reservations, CRM queries, or custom functions, and use JSON Schema to constrain parameters.
5. In test calls, accent, noise, repetition, rejections, interface failures, and out-of-bound issues are covered.
6. After checking the logs, latency, and costs, assign a formal number and gradually increase the data volume.

Tutorials on Tool Calls and Webhooks

1. Enable HTTPS for the Server URL and verify the identity of requests coming from Vapi.
2. Only expose the APIs necessary for the assistant to complete its tasks; do not grant direct database administrator privileges.
3. Server-side validation is performed on fields such as name, phone number, date, amount, and identity.
4. Set timeouts, idempotent keys, retry mechanisms, and clear failure messages for the tool.
5. Record the request ID, call ID, result, and any manual overrides, to prevent the logs from containing sensitive data in its entirety.
6. After forwarding Webhooks locally using the official CLI and conducting tests, deploy the production endpoint.

Outbound Campaign Tutorial

1. Verify that the list has legitimate authorization and clean it by region, time zone, and contact rejection status.
2. Set up caller ID, dialing time, concurrency, maximum retries, and voicemail policies.
3. The assistant must clearly indicate its identity and purpose; it is not allowed to pretend to be a real person or to conceal its marketing nature.
4. First, run a small batch of samples to check the success rates for connection establishment, task completion, complaints, and transfers.
5. Set the minute limit and model budget; it will automatically pause in case of abnormal connections or increased costs.
6. Update the CRM and block lists in real time to ensure that rejected users are not called again.

SDK, CLI, and the GitHub ecosystem

• Official Web SDK and SDK for React Native;
• Server-side SDKs such as Node.js, Python, Go, Ruby, Java, and C#;
• Examples of Flutter, Expo, Next.js, and various other frameworks;
• The official CLI allows for managing assistants, numbers, calls, workflows, campaigns, and Webhooks;
• The official docs repository and multilingual server-side examples are under an open license;
• The core Vapi hosting platform, call orchestration, and production infrastructure are not open-source software.

Which users are it suitable for

• Voice AI startup team:Quickly build your own intelligent phone agent product.
• SaaS developers:Embed voice conversations in web pages or mobile apps.
• Customer service and sales platform:Automation of inbound calls, outbound calls, scheduling, and qualification checks.
• Corporate engineering team:Combine proprietary models, operators, and business systems.
• Agent:Create white-label voice assistants for customers in various industries.
• Large institutions:High concurrency, data retention, SSO, and SLA are required.

Product advantages

• STT, LLM, TTS, and phone components can be combined freely, avoiding reliance on a single model provider;
• Developers can use their own API keys to directly manage the supplier’s account and costs;
• It offers low-latency interruption, tool invocation, multi-assistant handover, and high concurrency capabilities;
• APIs, SDKs, CLIs, Webhooks, and examples are provided in fairly comprehensive form;
• Supports web, mobile devices, phones, and embedded devices;
• Scale offers options for enterprise security, data residency, and on-premises deployment.

Usage restrictions and precautions

• The platform fee does not include the costs related to models and telephone calls; all these components must be taken into account when estimating the budget.
• Build only stores call history for 14 days; for long-term auditing, it is necessary to archive the data manually or use a custom solution.
• Voice latency and accuracy are affected by the network, service provider, model, accent, and noise.
• Errors in tool calls can lead to duplicate bookings, incorrect refunds, or data leaks;
• High-concurrency outbound calls require additional lines and compliance controls; it is not sufficient to merely increase the technical capacity for handling concurrent calls.
• Both HIPAA and zero data retention are expensive add-ons that are not included automatically in the basic version of the service.
• Phone recording, automatic dialing, marketing, caller ID, and AI-related functions are subject to local laws.

Security, Privacy, and Compliance

Vapi explicitly states that it meets the requirements of HIPAA, SOC 2, and GDPR; Scale also offers support for PCI, SSO, RBAC, and data residency. Specific certification details, contracts, and boundaries of responsibilities should be verified through the official security portal and sales agreements.

• The transmission of PII and PHI between prompts, logs, Webhooks, and external models should be minimized;
• Having a model key in place does not automatically meet compliance requirements; it is also necessary to evaluate the corresponding supplier.
• For recordings and transcriptions, retention periods, deletion procedures, and access auditing should be established.
• API keys should not be stored on the client side or in public repositories; key management should be handled on the server side.
• High-risk operations should require secondary confirmation or manual approval.

Open-source instructions

The Vapi core business platform is not an open-source project, but the official GitHub repository provides a CLI, documentation, SDKs, and various examples. The CLI and documentation are licensed under the MIT license; it is necessary to check the LICENSE file of each respective repository for details.

Open-source clients can only be used to call Vapi APIs; they cannot replace the hosted real-time voice orchestration services, telephone networks, and enterprise infrastructure. In the catalog, the column \"Is it open-source\" should indicate that the core products are not open-source, while some development tools are open-source.

Frequently Asked Questions

How much is Vapi per minute?

The hosting fee for the Build platform is $0.05 per minute, in addition to costs related to STT, LLM, TTS, and telephone service providers. The actual total cost per minute depends on the model used, the length of the prompt, the voice quality, and the region.

Is there a free quota in Vapi?

Build is billed on a pay-as-you-go basis; the official information indicates that a starting usage of over 60 minutes is included. The specific benefits, validity period, and account requirements can be found in the console.

How much concurrency is supported by Vapi?

Build supports 10 concurrent connections; an additional connection costs $10 per month. Scale can be customized, and according to the official FAQ the infrastructure is capable of handling more than 1,000 concurrent connections.

Does Vapi support its own models?

It is supported; the STT, LLM, and TTS components can be replaced, and vendor API keys can also be used. When such keys are available, Vapi does not charge fees for those models, but the vendor will still charge fees.

Is Vapi open source?

The core platform is not an open-source product. While there are several open-source projects available in the official CLI, SDKs, documentation, and examples, the entire production-grade voice infrastructure is still hosted by Vapi.