Listen to music together with up to seven other people, on the same couch or in different cities. Start a session and Duet gives you six words. Read them out on a call, let someone scan the code, or send the link. Pick a song from your phone, a YouTube video or a whole playlist, and it plays on everyone's phone at the same moment. Anyone can pause, skip, seek, vote or add their own songs.
Real screens, drawn by the app's own code.
| 🎶 In sync | Within a few milliseconds, and kept there: each phone eases its own speed instead of jumping. |
| 🌍 Anywhere | Phones connect directly over the internet. No account, and no server of ours. |
| 🛋️ Same room | On the same Wi-Fi, sessions show up by themselves and work with no internet at all. |
| Play a video or a whole playlist together, kept in step on every phone. | |
| 🎧 Play along | Line Duet up with a song playing in another music app on each phone. |
| 🗳️ Shared queue | Everyone adds songs from their own phone and votes the next one up. |
| 🎤 Lyrics | Synced lyrics scroll along for everyone. |
| 👑 Hand over | Leaving? Pass the session to someone else and the music carries on. |
| 💬 Chat & reactions | Floating emoji over the record, a little chat, and a heartbeat you can send. |
| 🔒 Lock screen controls | Keeps playing in the background. |
sequenceDiagram
participant H as Host phone
participant R as Public relay (MQTT over TLS)
participant G as Guest phone
H->>R: offer, sealed with a key from the six words
G->>R: answer, sealed the same way
Note over H,G: WebRTC connects the phones directly
H-->>G: songs, sent ahead over an encrypted data channel
loop five times a second
G->>G: where should the song be? (host clock + offset)
G->>G: ease speed, or jump if far off
end
The six words come from the BIP-39 list (66 bits). Stretched with PBKDF2 they give a topic name and an AES-256 key. Phones exchange WebRTC offers, answers and ICE candidates through HiveMQ's free public MQTT broker over TLS, sealed with that key, so the broker only ever sees noise under a random topic. Music never passes through it.
On the same Wi-Fi the host also advertises the session with NSD. A guest who has the code connects to it directly; a guest who doesn't can tap the session and ask, and on the host's approval receives the code sealed to a one-off X25519 key. No internet is involved.
Limit: with no relay server of our own (no TURN), roughly one in ten network pairs, mostly on strict mobile carriers, can't connect directly. Duet says so and suggests Wi-Fi on one side.
The host keeps the shared state: the queue, the current song, and its position at a moment of the host's clock. Guests estimate the offset to that clock the way NTP does, trusting only pings that came back about as fast as the fastest, and fitting a line over a few minutes of them so a clock running slightly fast or slow is followed too. Pings stuck behind song data are ignored.
Five times a second each phone compares where the song should be with where its player is. Small drift is eased out by playing up to 1.5% faster or slower; only a large, lasting gap causes a jump, aimed ahead by however long a jump has been taking. A per-output sync nudge lines up Bluetooth headphones, which add 150 to 250 ms, and Sync by ear can measure it with the microphone.
YouTube's embedded player reports its time late, so Duet watches for the moment the time changes and keeps the freshest reading; drift there is eased with a 5% speed change where YouTube allows it.
The host sends each song to every guest over a WebRTC data channel before it plays, and the next one while the current one plays. Songs are named by their SHA-256, checked on arrival, limited to 50 MB, and deleted when the session ends.
No account and no server of ours. Music goes directly between the phones, encrypted. The relay that helps phones find each other only sees sealed, unreadable data. Full policy: prempoddar.github.io/Duet/privacy.
Requires JDK 17 or 21 for Gradle.
./gradlew assembleDebug
./gradlew testDebugUnitTestTests cover the session code and key derivation, sealing, same-room pairing, clock estimation (including clock skew and pings delayed behind transfers), the YouTube position estimate, drift decisions, every playback command, the protocol, and whole sessions over in-memory links: joining, songs reaching everyone (and corrupt ones being rejected), guests' songs being forwarded, chat, reactions, hearts, hand-over, leaving and the eight-person cap. Connecting across networks needs two real phones.
As text
core/ pure Kotlin: session codes, sealing, pairing, clock sync, drift control,
playback state and commands, protocol, host and guest session logic, lyrics
net/ MQTT relay, local network (NSD + TCP), WebRTC data channels
data/ settings, on-disk song store, music library and import
play/ ExoPlayer sync loop, YouTube player, other apps, media session service
session/ wires a live session together
ui/ Compose screens, the record
Apache 2.0. Fraunces and DM Sans are used under the SIL Open Font License. The BIP-39 word list is MIT licensed.