Skip to main content
A task is one run of a workflow. You submit it from a workflow’s detail page (or from your AI client over MCP, or from a script); Cloud schedules it onto one of your team’s agents; results land on the task’s own page.

​
Find your output

Once a task reaches completed, its detail page shows an Outputs panel — one entry per file the workflow produced (images, video, model files, JSON, anything the agent uploads).
  • Click an image to preview it; click the download icon to save.
  • Bulk-download isn’t there yet — use the API or webhooks if you need automation.
  • Output links are short-lived signed URLs. If a download returns an expired-token error, refresh the task page — Cloud mints a fresh URL each time.
Outputs sit in object storage indefinitely on the live deployments; retention policy is set per deployment. If you can’t find an old task’s output, check the Tasks list — completed runs stay there with their results until your team’s retention policy removes them.

​
Follow live progress

The task detail page streams progress as the agent runs. Live logs, step counters, partial previews — whatever the workflow’s runtime emits — appear within a second of being produced. If your network blocks server-sent events, the page falls back to polling instead. Either way, you don’t have to refresh.

​
Lifecycle

Every task moves through these states: completed, failed, and cancelled are terminal — the task won’t move past them.

​
Cancel a task

  • One task: open the task detail page, click Cancel.
  • Several: in the Tasks list, multi-select rows and use the bulk-actions bar.
Cancelling a running task asks the agent to stop; the runtime ends cleanly and any artifacts already produced still upload. Cancelling a queued task is instant — nothing’s running yet.

​
Run it again

Per-task re-run is in development. For now, to run the same workflow again, open the workflow’s detail page and click Run — the run dialog remembers your last-run parameters, so you start from where you left off. Each run creates a new task; earlier tasks stay in the Tasks list as history.

​
Drive it from code or an AI client

  • Code: see Generate an API key for the REST runtime endpoints.
  • AI client (Claude Code, Cursor, …): see Connect via MCP.
  • Notifications: subscribe to webhooks so your backend reacts when tasks finish, without polling.
  • Workflows — the templates tasks are based on.
  • Agents — what executes the task.
  • Notifications — in-app notifications for task outcomes.