Skip to content
Comparison

Voice agent APIs compared (2026): who each is for

Which voice agent API fits comes down to two questions: who sells your phone numbers, and who runs the agent. Bland AI, Retell AI and Vapi: hosted platforms that sell phone numbers. Wixzel Voice: a bring-your-own-trunk API. LiveKit Agents and Pipecat: open-source frameworks you run.

Last updated

By Aqeel Shamsudheen, founder of Wixzel Voice, which is one of the products below, so read its entry as the interested party it is. First published . Products are grouped by kind and listed alphabetically within each group, not ranked, and each entry shows the date it last changed.

Which voice agent APIs are there, at a glance?

ProductKindTelephonyBilling
Bland AIHosted platformProvisions numbers; SIP and enterprise carrier options.Per-minute pricing on connected calls, with enterprise agreements.
Retell AIHosted platformProvisions numbers; also connects SIP trunks and existing carriers.Per-minute pricing built from platform, model, voice and telephony components; enterprise tiers.
VapiHosted platformProvisions phone numbers and also accepts SIP trunks; web and browser calls too.A per-minute platform fee plus the underlying provider and telephony costs, paid to Vapi or to the providers directly when you bring keys.
Wixzel VoiceHosted API, your own SIP trunkSells no numbers or minutes. Phone calls go over your own SIP trunk with no markup; web and app sessions need no trunk.Prepaid, billed per second of actual usage: $0.0466 to $0.1851 per connected minute by engine, plus your carrier's rate.
LiveKit AgentsOpen-source frameworkLiveKit SIP accepts inbound and outbound trunks you configure; the cloud offering also sells connectivity.The framework is open source and free. You pay each provider directly, plus infrastructure: your own, or LiveKit Cloud for the WebRTC and SIP layer.
PipecatOpen-source frameworkThrough a transport you wire in: a carrier's media streams over WebSocket, or WebRTC with dial-in and dial-out through Daily. The carrier account and its rates are yours.The framework is open source and free. You pay each provider directly, plus the infrastructure you run it on or Pipecat Cloud's hosting.

Competitors’ rates are deliberately not quoted: they change by tier and by month, so check each pricing page. Wixzel Voice’s figures are generated from the price book its meter bills against.

How do you choose a voice agent API?

The questions below separate these products more than any feature list does. Each answer covers every group, so it holds on its own.

Do you already have a SIP trunk?

Wixzel Voice requires one for phone calls and puts no markup on your carrier's rates; it has setup guides for Twilio, Telnyx, Plivo, Vonage, Bandwidth, Exotel, Vobiz and any other SIP carrier. Bland AI, Retell AI and Vapi accept a SIP trunk as well as selling numbers. LiveKit Agents connects trunks you configure through LiveKit SIP, and Pipecat reaches the phone network through a transport you wire in, so the carrier account stays yours.

Carrier setup guides

Who should sell you the phone numbers?

If you want to buy a number and place a first call in one place, Bland AI, Retell AI and Vapi provision numbers. Wixzel Voice sells no numbers or minutes and does not port numbers: you buy the number from your carrier and register it on your trunk. LiveKit Cloud sells connectivity alongside the open-source framework; with Pipecat, the carrier account is yours.

How do you want to be billed?

Bland AI, Retell AI and Vapi price per minute, each in its own way and by tier; check their pricing pages for current rates. Wixzel Voice is prepaid and billed per second of actual usage, $0.0466 to $0.1851 per connected minute by engine plus your carrier's rate, with no subscription. LiveKit Agents and Pipecat are free to run: you pay each provider and the infrastructure directly.

Wixzel Voice pricing

Do your callers speak Hindi or another Indian language?

Wixzel Voice's sarvam engine runs Sarvam's Saaras speech recognition, the Sarvam-105B model and the Bulbul voice end to end, at $0.0623 per connected minute. The hosted platforms can reach Indian-language models where their catalogues expose them, and the frameworks can use any provider you wire in; check each catalogue.

Indian-language phone agents

Should an AI agent be able to place the calls?

Wixzel Voice ships an official MCP server, hosted at https://mcp.voice.wixzel.com/mcp and published as wixzel-voice-mcp on npm, so Claude and other MCP clients can build agents and place calls, asking before any tool that dials or spends money. This page has not verified MCP support in the other products; check their documentation.

The MCP server

Do you need open source, or to run it on your own servers?

LiveKit Agents is Apache-licensed and Pipecat is permissively licensed; both are free to run, change and redistribute. Wixzel Voice is not open source, but a one-time $499 self-host license runs it on your own servers with the source included, for your own business. Bland AI offers dedicated deployments for enterprise; Retell AI and Vapi are hosted platforms.

Self-hosted Wixzel Voice

Hosted platforms that sell phone numbers

Buy a number, design the agent in a dashboard or through an API, and the platform runs everything else. Established, venture-backed products with broad feature sets, and the quickest route to a first call if you have no carrier yet.

Bland AI

Best for teams that want an integrated voice stack from one vendor, conversational pathways for designing calls, and enterprise options such as dedicated infrastructure.

Bland AI is a platform for AI phone calls that runs its own end-to-end voice stack, with conversational pathways, phone number provisioning, batch calling and enterprise deployment options.

Where Bland AI is stronger than Wixzel Voice

  • An integrated stack tuned as one system, which can mean lower latency than a composed pipeline.
  • Conversational pathways for designing branching calls without code.
  • Enterprise deployment options and a sales and support organisation.
  • Provisioned numbers and a larger set of integrations.

Wixzel Voice vs Bland AI, in detailEntry last updated

Retell AI

Best for teams that want a conversation-flow builder, provisioned numbers, batch campaigns and analytics in one product, with an API underneath.

Retell AI is a platform for building and deploying AI phone agents, with a conversation-flow builder, phone number provisioning, SIP support, batch calling and analytics.

Where Retell AI is stronger than Wixzel Voice

  • A visual flow builder, so an agent can be designed and tested by someone who does not write code.
  • Analytics, call review tooling and post-call evaluation built in.
  • Provisioned phone numbers, and a longer-established web calling SDK.
  • Scale and a support organisation behind it.

Wixzel Voice vs Retell AI, in detailEntry last updated

Vapi

Best for teams that want a mature, feature-rich platform with SDKs, provisioned phone numbers, web calling and the option to bring their own provider keys.

Vapi is a developer platform for building voice AI agents, with SDKs, a web dashboard, phone number provisioning and a large catalogue of pluggable speech and model providers.

Where Vapi is stronger than Wixzel Voice

  • Breadth. Vapi supports far more providers, voices and models, and adds features faster than a one-person alpha can.
  • Web calling maturity. Both have browser SDKs, but Vapi's has been in production longer and its browser audio is higher fidelity; Wixzel Voice's web sessions carry phone-quality (8 kHz) audio today.
  • Phone numbers. Vapi will sell you a number; Wixzel Voice requires you to bring a trunk first.
  • A visual builder, testing tools and a community with answers already written.

Wixzel Voice vs Vapi, in detailEntry last updated

A bring-your-own-trunk API

A hosted API that runs the agent for you but sells no telephony: phone calls go over a SIP trunk you already have, at your carrier's rates. Wixzel Voice is the one product in this group, and it is the author's.

Wixzel Voice

Best for developers who bring their own SIP trunk, or want agents in their own web and mobile apps, and want one prepaid balance to cover every engine

Wixzel Voice is a voice AI API for building AI agents that place and answer real phone calls over your own SIP trunk, or talk to people in your web and mobile apps. One API key and one prepaid balance cover every voice engine, billed per second of actual usage. Engines are presets (classic, sarvam, gemini-live, deepgram-agent) or composed stage by stage from named provider/model pairs, and the sarvam engine covers Indian languages end to end. Every price on the site is generated from the price book the meter bills against, and every charge is itemised per usage row.

Pricing. $0.0466 (gemini-live) to $0.1851 (deepgram-agent) per connected minute at typical speech rates, including a $0.012 per minute orchestration fee. Prepaid, $5 minimum top-up, no subscription, no free tier or trial. Telephony is your own trunk and is not marked up. Or a one-time $499 license to run it on your own servers.

Wixzel Voice limitations, honestly

  • Alpha. Behavioural changes ship behind a dated Wixzel-Version header, but the product launched in September 2026 and is operated by one person.
  • Five concurrent calls per account by default, phone calls and web sessions together. Raised on request, not by plan.
  • Campaigns dial their leads back to back and do not queue behind that limit: a lead over it is marked failed, and starting a stopped campaign again re-dials every lead.
  • Human transfer is blind only: the agent bridges the caller to a number you listed without briefing the person first. It is not available on test calls or in web and app sessions.
  • SDKs for TypeScript and Dart only. Other languages use the plain HTTP API and the OpenAPI document.
  • No phone numbers or minutes sold, and no number porting. A SIP trunk is required before the first phone call; web and app sessions need none.
  • Web and app sessions use µ-law 8 kHz audio today, the same fidelity as a phone line.
  • No visual agent builder. Agents are created through the API.
  • Fewer providers than the platforms: Deepgram, ElevenLabs, OpenRouter, Sarvam and Google today, under platform-owned accounts rather than your own keys.

All Wixzel Voice factsEntry last updated

Open-source frameworks you run

Write the agent in code against a framework, run it on your own servers or the vendor's cloud, and hold an account with every speech, model and voice provider you wire in. The most control, and the most to operate.

LiveKit Agents

Best for teams who want to own the conversation loop in their own code, already run infrastructure or need to self-host, and are willing to operate a worker fleet and hold provider accounts to get that control.

LiveKit Agents is an open-source framework for building realtime voice agents on top of LiveKit's WebRTC infrastructure, with SIP support for phone calls and a hosted cloud option for the infrastructure underneath.

Where LiveKit Agents is stronger than Wixzel Voice

  • Control. The turn-taking, interruption handling and tool-calling loop is your code, so behaviour Wixzel Voice does not expose as a setting is a patch away rather than a feature request.
  • Open source. Wixzel Voice can run on your own servers with a paid self-host license and includes its source, but the licence is for your own business; LiveKit Agents is Apache-licensed, free, and yours to redistribute.
  • Scale and track record. LiveKit runs realtime media in production for large deployments; Wixzel Voice has never been load-tested above single-digit concurrency.
  • Ecosystem. Far more provider plugins, client SDKs and community answers, and no dependency on one operator continuing to run a service.
  • No lock-in. The framework is Apache-licensed and the provider accounts are yours, so leaving costs you a deployment rather than a migration.

Wixzel Voice vs LiveKit Agents, in detailEntry last updated

Pipecat

Best for Python teams who want to compose the voice pipeline themselves from services they choose, keep every provider account, run the agent process on their own infrastructure or Pipecat Cloud, and change any part of the turn-taking loop by editing code.

Pipecat is an open-source Python framework from Daily for building realtime voice and multimodal agents as a pipeline of processors, with pluggable transcription, model and voice services, WebRTC and WebSocket transports, and a hosted deployment option in Pipecat Cloud.

Where Pipecat is stronger than Wixzel Voice

  • Control. The pipeline is a list of processors you wrote, so a behaviour Wixzel Voice does not expose as a setting is an edit rather than a feature request.
  • Open source and free. Pipecat is permissively licensed and yours to modify and redistribute; Wixzel Voice's self-host license includes the source but is for your own business only.
  • Provider choice. Any transcription, model or voice service with a service class, including ones Wixzel Voice does not run, and you can write a service class for one that has none.
  • Wideband audio. A browser session over WebRTC carries higher-fidelity audio than Wixzel Voice's phone-quality (8 kHz) web sessions.
  • No dependency on one operator. If Daily disappeared tomorrow the framework would keep running; Wixzel Voice is an alpha run by a sole proprietor.

Wixzel Voice vs Pipecat, in detailEntry last updated

How were these voice agent APIs chosen and described?

Products were included if a developer could build a phone-calling AI agent on them today, and each was described from its own public documentation at the level of business model: who owns the provider accounts, who sells the telephony, how it is billed, and how mature it is. Rates were left out because they change by tier and by month. Carriers such as Twilio or Telnyx, and model vendors such as Deepgram or ElevenLabs, were left out because they are parts of these stacks, not alternatives to them. The author, Aqeel Shamsudheen, builds Wixzel Voice; the comparison pages linked from each entry go into more detail, and corrections are welcome via the contact page.

Frequently asked questions

What is the best voice agent API for developers?
There is no single best one; it depends on what you bring. Teams that want to buy numbers and design agents on a mature platform fit a hosted platform: Bland AI, Retell AI or Vapi. Developers who already have a SIP trunk, or want agents in their own web and mobile apps, and want one prepaid balance across engines fit Wixzel Voice. Teams that want to own the pipeline in open-source code fit LiveKit Agents or Pipecat.
Which voice agent API is cheapest?
Only comparable once telephony and provider costs are included, and those depend on your carrier and your choice of models. Wixzel Voice publishes generated prices from $0.0466 to $0.1851 per connected minute with telephony at your carrier's rate and no provider bills. The other platforms publish a platform fee with provider and telephony costs on top; check their pricing pages, which change by tier.
Do I need my own SIP trunk for these?
For phone calls on Wixzel Voice, yes; it does not sell numbers, though its web and app sessions need no trunk. Bland AI, Retell AI and Vapi will provision a number for you and also accept a trunk. LiveKit Agents and Pipecat need you to integrate telephony yourself or through a provider.
Which voice agent API is best for Indian languages?
Wixzel Voice's sarvam engine runs Sarvam's Saaras, Sarvam-105B and Bulbul models end to end for Hindi, Tamil, Telugu and other Indian languages. Other platforms can reach Indian-language models where they expose them as providers; check each catalogue.
How was this list put together?
By Aqeel Shamsudheen, the founder of Wixzel Voice, so read the Wixzel Voice entry as the interested party it is. Each other product is described from its own public documentation at the level of business model rather than rate, with the date its entry last changed shown beside it. Products are grouped by kind and listed alphabetically within each group, not ranked. Corrections are welcome at the address on the contact page.