Models

A Reactor model is a real-time generative model (typically a video or world model). You connect to it over a session, steer it with commands, and receive streamed output. Every model on Reactor shares the same wire protocol:

• You open a session and connect() over WebRTC
• You send commands to control what the model does (e.g. set_prompt, send_image, move)
• The model emits media tracks (video, audio) and messages (structured JSON events)
• You close the session with disconnect()

The Reactor base class implements this shared protocol. Sessions, auth, file uploads, tracks, and events work the same regardless of which model you’re driving. A session can also be shared by multiple clients; see Sessions. What does differ is the schema: which commands a model accepts, what their parameters mean, what events it emits, and any model-specific state rules.

​

Per-model command and event reference

Each model publishes its own command and event schema. The model catalog lists every model you can connect to. Each model’s section documents its full schema (every command, parameter, and event) plus a prompt guide and tutorial.

​

Driving a model: typed SDK vs base class

Two ways to talk to a model:

• A typed SDK for the model (e.g. HeliosModel): adds named methods that mirror the model’s schema. Recommended for production
• The base Reactor class: generic, model-agnostic, sends raw JSON. Useful when no typed SDK exists yet, when prototyping, or when you want one client for several models

Both paths use the same wire protocol. See the SDK Reference overview for when each fits.