Decart

AI video/image generation platform with three APIs: Realtime (WebRTC, sub-500ms), Queue (async batch video), Process (sync image). SDKs for JavaScript, Python, and Swift. Auth via API key from platform.decart.ai — set DECART_API_KEY env var. Docs: https://docs.platform.decart.ai

When to use

Realtime video transformation — camera effects, video conferencing filters, AR/VR overlays, photo booths, live streaming
Batch video generation — marketing clips, social media content, product demos, animations from text or image
Image generation/editing — mockups, thumbnails, creative assets, image-to-image transformation
Avatar animation — virtual presenters, AI customer agents, talking heads animated with audio
Character transformation — transform user's live camera feed into characters (anime, fantasy, etc.)
Framework integration — Vercel AI SDK, TanStack AI, LangChain.js all supported

API selection

You want to...	API	Method	Best for
Transform live camera/video	Realtime	`client.realtime.connect()`	Interactive apps, below 500ms latency
Generate video from text/image	Queue	`client.queue.submitAndPoll()`	Content pipelines, batch jobs
Edit existing video	Queue	`client.queue.submitAndPoll()`	Video transformation
Generate/edit image	Process	`client.process()`	Sync results, thumbnails
Animate avatar with audio	Realtime	`client.realtime.connect()`	Virtual presenters, no camera needed

Model selection

Use case	Model	Type
Character transform live	`lucy_2_rt`	Realtime
Artistic style transfer	`mirage_v2`	Realtime
Animate portrait with audio	`live_avatar`	Realtime
Text-to-video (best quality)	`lucy-pro-t2v`	Batch
Image-to-video	`lucy-pro-i2v`	Batch
Video-to-video (best)	`lucy-pro-v2v`	Batch
Video-to-video (fast)	`lucy-fast-v2v`	Batch
Motion control	`lucy-motion`	Batch
Text-to-image	`lucy-pro-t2i`	Batch
Image editing	`lucy-pro-i2i`	Batch

Quick start patterns

Realtime (connect + transform)

const model = models.realtime("lucy_2_rt");
const stream = await navigator.mediaDevices.getUserMedia({
  video: { frameRate: model.fps, width: model.width, height: model.height }
});
const rt = await client.realtime.connect(stream, {
  model,
  onRemoteStream: (s) => { videoEl.srcObject = s; }
});
await rt.set({ prompt: "Transform into anime character", image: refImage, enhance: true });

Queue (video generation)

const result = await client.queue.submitAndPoll({
  model: models.video("lucy-pro-t2v"),
  prompt: "A golden retriever in a meadow, cinematic",
  resolution: "720p",
});

Process (image generation)

const image = await client.process({
  model: models.image("lucy-pro-t2i"),
  prompt: "Cozy coffee shop, warm lighting",
  resolution: "720p",
});

Authentication

Server-side: Set DECART_API_KEY env var. The SDK picks it up automatically.
Client-side: Generate short-lived tokens on your backend via client.tokens.create(), then send token.apiKey to the frontend.
NEVER expose permanent API keys in browser or mobile code.

Camera constraints

Always use model.fps, model.width, model.height from the SDK to avoid scaling artifacts.

Boundaries

What agents CAN do

Generate integration code for all three APIs (Realtime, Queue, Process)
Help select models, configure SDK, set up authentication
Write WebRTC realtime streaming code
Build framework integrations (Vercel AI SDK, TanStack AI, LangChain.js)
Troubleshoot connection, moderation, and billing issues

What agents CANNOT do

Create or manage Decart accounts, API keys, or billing
Access the Decart dashboard or platform settings
Test realtime connections (requires actual WebRTC + camera)
Verify credit balance or usage
Bypass content moderation policies

Common gotchas

NEVER expose permanent API keys in client code. Use client tokens for browser/mobile.
ALWAYS use model.fps, model.width, model.height for camera constraints. Mismatched resolution causes latency and artifacts.
ALWAYS call disconnect() and stop media tracks when done. Leaving connections open causes memory leaks and billing.
Catch errors from set(), setPrompt(), setImage(). Moderation rejections throw here.
Listen to connectionChange events. The SDK auto-reconnects, but your UI must reflect the state.
Keep enhance: true (default) unless you need exact prompt control.
Use submitAndPoll() for queue jobs. Don't manually poll faster than every 2 seconds.
Image editing (i2i) only works with Process API, not Queue API. Queue is for video only.
Realtime sessions end silently when credits run out. Listen for disconnect events.
For Lucy 2 character transform, use clear portrait photos with good lighting.

Verification checklist

Resources

Full docs navigation: https://docs.platform.decart.ai/llms.txt
Models overview: https://docs.platform.decart.ai/getting-started/models
Realtime best practices: https://docs.platform.decart.ai/models/realtime/streaming-best-practices
SDK reference: https://docs.platform.decart.ai/sdks/javascript-realtime