Jobs and Sandboxes

Two ways to run code on MEGA hardware without standing up a Space.

Jobs run a container to completion on the hardware you pick and then stop. They are meant for the work that surrounds a model rather than serves it: evaluation sweeps, dataset preparation, fine-tuning runs. A job can mount buckets and repositories, and its logs and exit status are kept against your account.

Sandboxes are short-lived execution environments attached to a bucket or repository, intended for an assistant rather than a person — fast startup, scoped filesystem access, no persistent state between sessions. They are what backs code execution when an MCP client asks to analyse a dataset or try a change against a repository.

Both bill from the same prepaid compute balance as Spaces, visible under billing settings.