Skip to main content
Reactor Serve deploys your models on Reactor’s GPUs so applications can send inputs and receive outputs throughout a live session. Start with the quickstart to deploy a working model.

​
How it works

1

Port your model

Write a ReactorApp subclass, as in the runtime overview. Load weights once, and write one generate() that returns video or audio on named tracks.
2

Declare what clients can steer

Your model declares its own interface: the state a client can set, the tracks it sends, and the messages it sends back. The runtime validates every payload against that declaration and applies it between steps. It also publishes the declaration as the schema that clients and typed SDKs are built from.You never write transport code for any of it. See Commands and messages.The client gets set_prompt. Write an @event handler for what a field cannot carry.
3

Deploy it

Install the CLI, scaffold a workspace, and deploy. One command registers the model when needed, publishes the release with its weights, and activates it on Reactor’s GPUs.
4

Connect a client

Point the SDK at the registered org/name and pass a token, as in the quickstart. Mint the token on your server, as Authentication describes.
Reactor’s GPUs are the default target, so nothing above needs hardware of your own. When you do have a local GPU, the same workspace runs on it with reactor run, and Test locally covers that loop.

​
Guides

Deploy RLDX-1 to send camera views and robot state to a vision-language-action model on Reactor. Then test your deployment with synthetic inputs to check camera synchronization and action timing.

​
What the runtime handles

Real-time streaming

Frames reach clients over WebRTC as you generate them, not after the video is finished.

Live interaction

Clients change inputs mid-generation. Nothing restarts, and nothing is re-queued.

No transport code

You never import a WebRTC library, hold a WebSocket open, or encode video.

Validated inputs

Declare each command with types and constraints. The runtime checks every payload first.

​
Use these docs with a coding agent

Give the agent the quickstart and runtime overview, plus your inference code and dependency requirements. Include the weights guide when the model needs checkpoints. Specify the commands and track names the client needs. Ask the agent to verify a received frame and a command’s effect before it reports success. A successful image build alone does not test either behavior. Use each page’s Copy option, or append .md to its URL for Markdown. The docs MCP server is available at https://docs.reactor.inc/mcp.

​
Next

Runtime Overview

Tracks, the step loop, commands, and messages.

Model Anatomy

A working model read line by line.