Start an async Creative Studio image or video job. Returns id + status=processing quickly — do not wait for pixels in this call. In Claude Connectors an inline MCP App widget shows loading then the image/video; do NOT paste CDN links, markdown images, or use “show/visualize file” tools for the result — the widget is the preview. Wait ~8–15s then call get_creative_generation sparingly (5–10s apart) until status is completed or failed. Workflow: (1) list_creative_models if model unknown, (2) get_creative_model when you need exact parameter names, (3) confirm the plan with the user if any detail (scope, aspect ratio, model choice) was not fully specified by them, (4) create with modelId + prompt + parameters, (5) poll sparingly. Call this ONCE per intended output — if the user provided several reference images and wants a single combined result, use one supportsMultiReferenceMedia model with all images in one mediaUrls entry instead of calling this once per image; only generate multiple separate outputs if the user explicitly asked for that. mediaUrls may include prior outputUrl values for image-to-video / edit chains. Attached/local files: prefer omitting mediaUploads and using mediaIntent (for optional params the user clearly wants) so the Creative Studio widget can collect the file in-browser without base64 through your context; coding agents with network access may use request_creative_media_upload (presigned PUT) or upload_creative_media only when bytes never surface in your response. IMPORTANT — optional media params: get_creative_model often marks image/video reference params as required=false (the model can also run as pure text-to-X). That does NOT mean you can skip a param the user clearly wants used — if they say "use this exact image" / "edit this photo" and you have no bytes for it, you MUST include its name in mediaIntent, or the job silently runs text-only and produces a fabricated result that only superficially resembles their photo instead of erroring. Requires creative:generate scope.