Video-generation gateways

A video-generation gateway is a hosted inference aggregator that resells many third-party video models (Veo, Kling, Seedance, Hailuo/MiniMax, Luma, Wan, LTX-Video, Pika, and others) behind a single account, a single API convention, and a single bill. You send a prompt and parameters to a gateway-owned endpoint; the gateway runs the underlying model on its own infrastructure and returns a hosted result URL. This skill covers fal.ai, Replicate, and WaveSpeed as the representative gateways, and the production decisions common to the category.

This skill is about routing and operating, not about art direction. Prompt craft, shot design, and motion direction for a specific model belong to model- or craft-specific skills. Here the questions are: which gateway, going through a gateway at all, how the async job contract works, how to discover a model's schema, how to pin a version, how to cap spend, how to fail over, and whose license governs the output.

When a gateway is the right route (and when it is not)