fal-ailisted

# fal.ai generation fal.ai hosts many independent, frequently-changing models behind one client pattern (submit a job, poll or subscribe, retrieve result URLs). This skill teaches the pattern and the operational rules; it does not pin a model catalog, because model IDs, capabilities, and pricing change on their own schedule and going stale here would silently break every production that trusted it. Some endpoints are explicitly **realtime**, not queued jobs. When the selected official schema requires WebRTC/WMA (for example H3 Max Director), do not use the submit/poll/subscribe patterns below. Use the endpoint's realtime client contract through an application-authenticated server proxy, keep the Fal key server-side, and treat the result as live media until a documented export has been retained. Read the route-specific skill before offering that mode. This skill owns the client call once a model is already chosen and the shot's direction is already decided. Read `video-model-selection` first to choose between this provider and `google-ai`, and `video-cinematography` to turn the storyboard beat into the actual prompt/camera/lighting direction you pass in. fal.ai's officially supported client is Node.js (`@fal-ai/client`), not Python. `npx` is already this product's proven way to run a managed Node CLI non-interactively (see `hyperframes-cli`) -- use the same pattern here rather than reaching for a Python SDK that isn't fal.ai's first-class offering. ## Never invent a model