Voice Agents
Build an AI that talks with people in real time — on the phone or in the browser — by chaining speech-to-text, an LLM and text-to-speech.
Trends
LiveKit gained the most ground in package downloads: 3.8× against the median npm package over a year.
Public counts each month: pull requests on GitHub, package downloads, stars, Product Hunt followers. They show direction, not market share.
- Oct 1 · Action needed OpenAI removes gpt-5.1, gpt-5.3-codex and gpt-5.4-nano from the API on April 1, 2027; move to gpt-6-sol or gpt-6-luna.
- Oct 1 · Action needed OpenAI removes the tts-1, tts-1-hd and gpt-4o-mini-tts text-to-speech models on January 6, 2027; move to gpt-realtime-2.1-mini.
The real choice
Hosted platform or open-source framework. Hosted platforms handle phone numbers, turn-taking and provider wiring, and charge per minute on top of model costs. Open-source frameworks give you full control over the pipeline and cost, but you run the real-time servers yourself. A speech-to-speech model skips the chain entirely: one model hears and answers, with less delay but less control over each step.
Pick by situation
Tap the ones that are you — the tools that fit light up below.- You want a hosted voice agent but need to swap providers and call your own functions
Vapi
ElevenLabs - The agent answers or makes business phone calls, set up largely without code
Retell AI - Voice happens inside your web or mobile app, or you need video too
LiveKit
Daily - You want full control over every component of the pipeline, in Python
Pipecat
LiveKit - Response delay matters most and you're fine with one model doing the listening and talking
OpenAI API
Compare the tools
Grouped by the side of the choice they answer, not ranked. Open a row for when to use it, the trade-off and what makers say.
VapiplatformDevelopers who want to configure every part — models, voices, tools, phone numbers — through an API.


16+2 open source41k/wk
Use it whenYou want a hosted voice agent but need to swap providers and call your own functions.
Trade-offPlatform fee per minute plus the pass-through model costs, which makes the total harder to predict.
Used by Calling Clones, Hyring, Inworld, Omakase.ai, Relyable and 11 more · in 2 open-source projects.
THANKS! I mean it. You push me to create this product with your Challenge. I do love being able to call my clones! And I will keep on using VAPI to play with AI and phone calls!
Retell AIplatformPhone agents for bookings and customer service, set up largely through a visual builder.


4+1 open source142.1k/wk
Use it whenThe agent's job is answering or making business phone calls.
Trade-offLess room for custom pipelines than building on an open framework.
Used by Copperlane, Relyable, SantaCalls.ai, Cal.com · in 1 open-source project.
We chose Retell because it offers low-latency, interruption-handling voice conversations with reliable telephony integrations, making it far more production-ready for real phone workflows.
ElevenLabsapi · self-hostableVoice agents built on ElevenLabs voices, with turn-taking, tools, knowledge bases and phone numbers in one platform.No evidence yet380.7k/wk18.9× vs npm*
Use it whenVoice quality is a selling point and you want the agent platform from the same vendor as the voices.
Trade-offPer-minute pricing plus LLM costs, and the best fit is with ElevenLabs' own speech models.
In Lovable, ReplitNo maker's product we track shows it for this yet.
LiveKitplatform · open sourceReal-time voice and video in your own app, with an agents framework on top of its WebRTC infrastructure.


38+27 open source3.6M/wk3.8× vs npm
Use it whenVoice happens inside your web or mobile app, or you need video too.
Trade-offMore to learn and run than a hosted platform, unless you pay for LiveKit Cloud.
Used by April, Deeptrue, Extra Thursday, Firecoach AI, graph8 and 33 more · in 27 open-source projects.
Voice AI needs real-time, not near-real-time. LiveKit's open-source WebRTC gave us full pipeline control with sub-100ms latency. Daily.co and Dyte were too abstracted for what we needed.
Pipecatframework · open sourceA Python pipeline where you choose and swap every speech, LLM and voice component.

5+13 open source245.7k/wk
Use it whenYou want full control over the pipeline and your team works in Python.
Trade-offYou host and scale the real-time processes yourself.
Used by Gradient Bang, Here, let me AI that for you., Intersect Labs, Morning Commute, Product Hunt · in 13 open-source projects.
Open source, vendor neutral, framework for building voice and AI-native experiences
DailyplatformWebRTC audio and video transport plus Pipecat Cloud, which hosts Pipecat agents so you don't run the real-time servers.


16+5 open source68.7k/wk
Use it whenYou build with Pipecat and want hosting, or need browser and phone transport for your own pipeline.
Trade-offHosted only, billed per minute, and you still build the agent logic yourself.
Used by Cerebrium, Coval, Gradient Bang, Inworld, Tavus and 11 more · in 5 open-source projects.
Tavus utilizes Daily's WebRTC platform to power real-time video conversations with our digital replicas, enabling smooth and embeddable video chat experiences across web applications.
OpenAI APIapiThe Realtime API, where one model listens and speaks directly, over WebRTC, WebSocket or SIP phone calls.No evidence yet37.4M/wk+60% vs npm
Use it whenYou want the shortest delay and the most natural turn-taking, and can accept OpenAI's voices.
Trade-offLess control over each step than a chained pipeline, a small set of voices, and audio tokens cost more than text.
llms.txtIn Lovable, Replit, RorkNo maker's product we track shows it for this yet.
Before you choose
Get a working phone or browser call on a hosted platform first; you'll learn where latency and interruptions hurt before writing infrastructure. Watch total response delay — anything over about a second feels broken in conversation. Move to a framework when per-minute fees or missing control become the real constraint.
- Measuring each component's latency on its own and never the full turn, where speech-to-text, the model and text-to-speech add up to a pause users notice.
- Skipping interruption handling, so the agent keeps talking over a caller who has already started answering.
- Giving a phone agent tools that change bookings or accounts without confirming out loud what it's about to do.
Other options
Real choices most makers here won't need to weigh.
Decided alongside
What the 68 makers' products here chose for their other decisions.