Simcast

Real-time AI video infrastructure

Generated video, streamed as it renders.

Simcast is an API for low-latency AI video generation and live streaming — models that produce frames continuously, delivered to viewers while they're still being made. Built for developers who need generation to keep pace with a live audience, not a render queue.

inference latency

What Simcast does

One endpoint, from prompt to playing pixels. You send a prompt or a stream of prompts; Simcast runs the generation pipeline, encodes the output into a standards-based live stream (HLS), and hands your viewers a URL. No render-and-wait, no stitching clips together.

The hard parts — keeping a diffusion pipeline fed in real time, holding the stream seam-free across segment boundaries, and reacting to live input without stalling playback — are the product. You call the API; the latency engineering is ours.

Built around three constraints

01 / latency

Streamed, not rendered

Frames leave the pipeline continuously and reach viewers as a live HLS stream — generation runs ahead of playback, never behind a queue.

02 / interaction

Reacts to live input

Prompts and viewer signals can change the stream on the fly, injected into upcoming segments without breaking playback continuity.

03 / quality

Seam-free continuity

Segments join without frozen frames or timeline drift, so a stream that runs for hours holds together like a single continuous take.

The interface

stream.py
# start a live stream from a prompt
import simcast

stream = simcast.Stream.create(
    prompt="a calm host talking to the camera, warm room",
    resolution="720p",
)

# viewers watch here, live, while it generates
print(stream.hls_url)

# steer it in real time
stream.send("she picks up a guitar and starts to play")