Skip to main content
Video work takes seconds to minutes, so the API is queue-first:
remove-background is synchronous: no polling at all. tighten and launch-cut are queued and resolve through the jobs envelope.

​
Job states

queued → running → completed | refused | failed
  • completed carries the full envelope: finalUrl, canvasUrl, recipeVersion, timings.
  • refused carries a typed reason: free.
  • failed is on us: see Infrastructure errors; failed jobs are not charged.

​
Estimate first

GET /v1/estimate?recipeId={id} is free and returns { credits, etaSeconds, creditsRemaining, blockers[] } with typed blockers (INSUFFICIENT_CREDITS, IN_FLIGHT, SCOPE_DENIED, RECIPE_NOT_ALLOWED, SPEND_CAP_REACHED). Agents should estimate, ask consent, then run.
An estimate validates the PRICE and your ACCOUNT (credits, scopes, in-flight caps), not your material: material checks (duration, readability, subject presence) happen at run time, and when they fail you get a free typed refusal, never a charge.

​
Polling etiquette

Poll every 5-15s; each template page posts an etaSeconds to calibrate expectations. Webhooks (callback_url) are on the roadmap for fire-and-forget integrations.