Concepts

This page explains the concepts you will encounter while building with AxonKit. If you are coming from LiveKit or mediasoup, most terms will feel familiar.

SFU, not mesh

AxonKit is a Selective Forwarding Unit (SFU). Unlike a mesh or peer-to-peer architecture, clients do not send media directly to each other. Each client uploads its media to the server once; the server then forwards a copy to every subscriber.

This is what makes multi-party calls scale: a publisher’s uplink cost stays constant regardless of how many people are watching.

Mesh (P2P)                    SFU (AxonKit)
  A ────── B                    A ──►┐
  │  ╲   ╱  │                        │  Server
  │   ╳   │                    B ◄───┤  forwards
  │  ╱   ╲  │                    C ◄───┘
  C ────── D

Nodes, clusters, regions

AxonKit organizes its runtime into three levels of locality:

• Node — a single process running a mediasoup worker and an embedded NATS JetStream instance. Nodes serve clients directly.
• Cluster — a group of nodes that are topologically close (e.g. the same data center). Intra-cluster traffic is preferred.
• Region — a geographic grouping of clusters (e.g. US, EU).

Nodes discover each other automatically via NATS KV, so a new node joins the topology without static configuration.

Rooms

A room is the unit of isolation for a session of participants. The room a client joins is derived from the JWT token’s room grant, not from a separate room name argument — this keeps authorization and routing bound to the token.

Participants

A participant is a client connected to a room. There are two kinds:

• LocalParticipant — the participant representing the current client.
• RemoteParticipant — everyone else in the room.

Participants expose metadata, a display name, attributes, and the set of tracks they publish.

Tracks

A track is a single media stream — an audio track, a video track, or a data track. Tracks flow through a publication:

• TrackPublication wraps a published track and its subscription state.
• Publishing a track makes it available to the room; subscribing downloads it from the server.

Simulcast & SVC

• Simulcast encodes a single video source into multiple quality layers (spatial layers). Subscribers receive the layer that matches their bandwidth and viewport.
• SVC (Scalable Video Coding, e.g. VP9/AV1) achieves the same idea with a single encoding that can be truncated.

Dynacast

With dynacast, the server pauses simulcast layers (or whole tracks for SVC) that have no subscribers. Publishers save CPU and bandwidth when nobody is watching their high-quality layers.

Adaptive Stream

Adaptive Stream lets the client optimize subscriptions based on UI state: element visibility and rendered size drive which layer to subscribe to, so a hidden or tiny tile does not consume full resolution.

Routing: RIB and FIB

When a publisher and a subscriber are on different nodes, AxonKit forwards media between their mediasoup routers over a pipe transport.

• RIB (Routing Information Base) advertises the streams a node produces.
• FIB (Forwarding Information Base) determines where a subscriber should fetch each stream from.

Together they provide dynamic, low-latency routing that prefers local-cluster traffic and probes RTT for cross-region path selection.

Signaling: DataChannel RPC

After the initial HTTP join, signaling happens over a DataChannel using the AXON/1.0 protocol — a text-based request/response format inspired by the NATS wire protocol.