Let your users talk to your app

Micdrop streams your users’ speech to your Node server and plays the answer from the AI models you choose.

npm install @micdrop/web @micdrop/react

import { Micdrop } from '@micdrop/web'
import { useMicdropState } from '@micdrop/react'

const start = () => Micdrop.start({ url: 'wss://your-server.com' })

function Call() {
  const { isStarted, conversation } = useMicdropState()
  return (
    <>
      <button onClick={isStarted ? Micdrop.stop : start}>
        {isStarted ? 'Hang up' : 'Start the call'}
      </button>
      {conversation.map((item, i) =>
        'content' in item && (
          <p key={i}>{item.role}: {item.content}</p>
        )
      )}
    </>
  )
}

npm install @micdrop/react-native @micdrop/react react-native-audio-api

import { Micdrop } from '@micdrop/react-native'
import { useMicdropState } from '@micdrop/react'
import { Button, Text, View } from 'react-native'

const start = () => Micdrop.start({ url: 'wss://your-server.com' })

function Call() {
  const { isStarted, conversation } = useMicdropState()
  return (
    <View>
      <Button
        title={isStarted ? 'Hang up' : 'Start the call'}
        onPress={isStarted ? Micdrop.stop : start}
      />
      {conversation.map((item, i) =>
        'content' in item && (
          <Text key={i}>{item.role}: {item.content}</Text>
        )
      )}
    </View>
  )
}

npm install @micdrop/web

import { Micdrop } from '@micdrop/web'

Micdrop.start({ url: 'wss://your-server.com' })

Micdrop.on('StateChange', (state) => {
  console.log(state.conversation)
  console.log(state.isAssistantSpeaking)
})

Micdrop.on('EndCall', () => {
  console.log('Call ended by assistant')
})

// Hang up
Micdrop.stop()

npm install @micdrop/server @micdrop/openai @micdrop/gladia @micdrop/elevenlabs

import { MicdropServer } from '@micdrop/server'
import { OpenaiAgent } from '@micdrop/openai'
import { GladiaSTT } from '@micdrop/gladia'
import { ElevenLabsTTS } from '@micdrop/elevenlabs'

const { OPENAI_KEY, GLADIA_KEY, TTS_KEY, VOICE } =
  process.env as Record<string, string>
const systemPrompt = 'You are a helpful assistant'

wss.on('connection', (socket) => {
  new MicdropServer(socket, {
    firstMessage: 'How can I help you today?',
    agent: new OpenaiAgent({ apiKey: OPENAI_KEY, systemPrompt }),
    stt: new GladiaSTT({ apiKey: GLADIA_KEY }),
    tts: new ElevenLabsTTS({ apiKey: TTS_KEY, voiceId: VOICE }),
  })
})

Hear Micdrop hold a conversation

Every demo is open source in the repository, ready to run with your own API keys.

A voice assistant on OpenAI Realtime

One speech to speech model hears and answers, and Micdrop still handles the turns, the interruptions and the transcript.

See the code

A voice assistant on French models

The call speaks French, on French providers from the transcription to the voice.

See the code

A voice assistant on your own machine

Transcription, agent and voice run locally, with no API key, and the audio stays on the laptop.

See the code

A robot that obeys your voice

Two robots hear the same sentence. One reads it with Jev, the other with Claude, and a clock under each shows which one moves first.

See the code

A game that plays a famous person

Ask yes or no questions out loud to find who the game is playing. It answers in the first person, and Jev reads each question in a few hundred milliseconds.

See the code

A word to find by asking questions

The game thinks of something and answers each question with yes, probably, I don’t know, probably not or no.

See the code

Micdrop handles what happens between two sentences

A voice agent feels slow or rude when it cuts you off mid-thought, waits too long to answer, or talks over you.

The browser hears you
Voice activity detection runs in the browser, which streams the audio to your server while you are still talking.
The agent waits through a pause
Turn detection sees that the sentence is unfinished, so the agent waits for you to say the day.
The agent runs your code
Before it answers, the agent calls a tool you wrote, which runs in your Node server next to your database.
The voice starts early
Text to speech reads the first words of the answer while the model is still writing the rest.
You can interrupt the assistant
Playback stops as soon as you talk. Echo cancellation keeps the assistant’s voice out of your microphone.

Your app code stays the same whichever models answer

Browser or phoneYour app

Your Node server

Speech to textGladia

AgentOpenAI

Text to speechElevenLabs

import { MicdropServer } from '@micdrop/server'
import { GladiaSTT } from '@micdrop/gladia'
import { OpenaiAgent } from '@micdrop/openai'
import { ElevenLabsTTS } from '@micdrop/elevenlabs'

new MicdropServer(socket, {
  stt: new GladiaSTT({ apiKey: GLADIA_KEY }),
  agent: new OpenaiAgent({ apiKey: OPENAI_KEY, systemPrompt }),
  tts: new ElevenLabsTTS({ apiKey: ELEVENLABS_KEY, voiceId }),
})

Each stage streams into the next. To swap a provider, you change its line. Read the docs →

Browser or phoneYour app

Your Node server

Speech to speechGemini Live

import { MicdropServer } from '@micdrop/server'
import { GeminiLive } from '@micdrop/gemini'

new MicdropServer(socket, {
  realtime: new GeminiLive({
    apiKey: GEMINI_KEY,
    systemPrompt,
  }),
})

One model hears the audio and answers in audio. Micdrop still handles the turns, the interruptions and the order of the transcript. Read the docs →

Browser or phoneYour app

Your Node server

Speech to textGladia

AgentMistral

Text to speechGradium

import { MicdropServer } from '@micdrop/server'
import { GladiaSTT } from '@micdrop/gladia'
import { MistralAgent } from '@micdrop/mistral'
import { GradiumTTS } from '@micdrop/gradium'

new MicdropServer(socket, {
  stt: new GladiaSTT({ apiKey: GLADIA_KEY }),
  agent: new MistralAgent({ apiKey: MISTRAL_KEY, systemPrompt }),
  tts: new GradiumTTS({ apiKey: GRADIUM_KEY, voiceId, region: 'eu' }),
})

Three European providers keep the audio and the conversation inside the EU. They handle French and 90 other languages. Read the docs →

Browser or phoneYour app

Your machine

Speech to textWhisper

AgentOllama

Text to speechKokoro

import { MicdropServer } from '@micdrop/server'
import { WhisperSTT } from '@micdrop/whisper'
import { AiSdkAgent } from '@micdrop/ai-sdk'
import { KokoroTTS } from '@micdrop/kokoro'
import { createOpenAI } from '@ai-sdk/openai'

// Ollama serves the OpenAI protocol on /v1
const ollama = createOpenAI({ baseURL: 'http://localhost:11434/v1', apiKey: 'ollama' })

new MicdropServer(socket, {
  stt: new WhisperSTT({ model: 'base' }),
  agent: new AiSdkAgent({ model: ollama.chat('qwen3:4b-instruct'), systemPrompt }),
  tts: new KokoroTTS({ voice: 'britishFemale' }),
})

Whisper and Kokoro run inside your Node process while Ollama runs the agent. The call works offline and needs no API key. The audio stays on your machine. Read the docs →

Micdrop is a library that runs inside your server

	Micdrop	Hosted platformsVapi, Retell AI	Agent frameworksPipecat, LiveKit Agents
Where the call runs	In the Node server you already deploy	On the vendor’s servers	In an agent service next to your app, often behind a WebRTC server
Languages	TypeScript, from the browser and the phone to the server	Any, through a REST API and webhooks	Python first. LiveKit Agents also has a Node SDK.
Your agent logic	A class in your codebase, next to your database and your auth	A public endpoint that the platform calls	Code in the agent process, apart from your app
Cost	Free under the MIT license. You pay each AI provider directly.	A fee per minute on top of the providers, $0.05 on Vapi	Free and open source. You pay for the servers that carry the audio, or for a cloud plan.
Phone numbers and dashboard	You add them yourself. Micdrop covers calls in web and mobile apps.	Included, with compliance certifications and an SLA	Telephony through SIP or Twilio

Read the detailed comparisons withVapi Retell AI Pipecat LiveKit Agents

Raconte runs its voice interviews on Micdrop

With Raconte, a team sends a link, an AI interviews each person by voice, and the team gets back transcripts, sentiment and insights.

Each interview runs on Micdrop in the browser, with Gladia, Mistral, Gradium, OpenAI and ElevenLabs on the server. Raconte and Micdrop have the same author.

Visit raconte.ai

Raconte showing the transcript of a voice interview, with the recording, the AI summary and the sentiment over time

Micdrop also handles the rest of the call

In the browser and on the phone

Voice activity detection Turn detection Call state Live transcript Microphone and speaker selection Mute Pause and resume React hooks Latency settings

On the server

Tool calls Intent classification Data extraction Dictation First message Automatic hang up Authentication and parameters Noise filtering Audio recording Resumed conversations

Any provider, with a fallback

Micdrop ships integrations for OpenAI, Gemini, Mistral, Gladia, ElevenLabs, Cartesia, Gradium and local models. You can plug any other provider into the same interfaces. When a provider fails, a fallback takes over.

Provided integrations Local models Sovereign voice AI Custom agent Custom speech to text Custom text to speech Agent fallback Speech to text fallback Text to speech fallback

Just call Micdrop.start() and your app holds a conversation.

Start building

Micdrop is free and open source under the MIT license. Bring your own API keys, or run the models on your own machine.

Hear Micdrop hold a conversation

A voice assistant on OpenAI Realtime

A voice assistant on French models

A voice assistant on your own machine

A robot that obeys your voice

A game that plays a famous person

A word to find by asking questions

Micdrop handles what happens between two sentences

The browser hears you

The agent waits through a pause

The agent runs your code

The voice starts early

You can interrupt the assistant

Your app code stays the same whichever models answer

Micdrop is a library that runs inside your server

Raconte runs its voice interviews on Micdrop

Micdrop also handles the rest of the call

In the browser and on the phone

On the server

Any provider, with a fallback