Real-time infrastructure for multi-modal AI
People and agents trade video, audio and data, each over one QUIC connection. Recorders, captioners and other agents subscribe to the tracks they need.
WebRTC connects the people on a call. Caprid also sends each track to every process that subscribes to it.
What's included
| No rooms | Clients connect once and subscribe to the tracks they need. There's nothing to create, join or tear down. |
|---|---|
| Settings per subscriber | Two subscribers to the same track can get different trade-offs, such as lowest latency for one and every frame for the other. |
| Scoped tokens | Each listener's token covers only its own tracks and has its own expiry. |
| Sent only when subscribed | Caprid sends a track only after something subscribes to it, so no bandwidth goes to tracks nobody reads. |
You decide how each track is delivered
Set encoding and packaging for each track. Rank tracks so a tool result goes out before a video frame. Choose what Caprid drops first when the network is congested. Send generated speech faster than real time so the player can buffer ahead. On the same track, one subscriber can get only the newest frames while another gets every frame.
Show me how it worksAttach listeners to a session between a caller and an agent, and see which tracks reach each one.
switch tracks, or a whole side, on the caller and agent boxes; lines run in the same order as the tracks in each box
Under the hood
Our relays, or yours
Caprid runs routing, access control and keys. If media can't leave your network, run Self-hosted relays. Caprid still runs the control plane.
Operators that work on your tracks
Attach an Operator to a stream to transcribe, moderate or record it, or to package it as low-latency HLS for screens that can't run a real-time client.
From the founders of Mux
Caprid runs on QUIC and MoQ, open protocols from the IETF.