Portal, from the San Francisco startup Clusy, is runtime infrastructure for Python jobs that have to keep their place when the machine under them disappears. It records a running process from inside the interpreter, including variables, tensors, open files, and the exact step in the loop, then resumes that step on another provider, chip, or CPU and GPU mix. The launch is for researchers, ML engineers, and coding agents running training, simulation, and large in-memory analysis on spot capacity, where a preemption or an out-of-memory kill normally throws the work away. Migration stays inside the application. There are no VM snapshots and nothing to install in the hypervisor, so a job can leave an x86_64 GPU VM and continue on a different architecture, with each tensor checked against its digest on arrival. Portal attaches to cloud accounts the user already has, and a coding agent can be pointed at a setup prompt on the site. In one published night of runs across AWS, Modal, and Lambda, five workloads cost $203, and the spot jobs saved 60% against the same machines on demand. A 728 GB ERA5 climate array saved 71%. Karpathy's autoresearch ran at 22% of the AWS on-demand price. A ship-tracking job moved from an AWS spot instance to a Lambda GPU box and back with no steps lost, and ESMFold folded 16,650 human proteins across eight GPU shards at 27% under AWS list price for those GPU hours. Clusy already ships an AI-native notebook and is backed by Founders, Inc. Portal is the layer that lets those jobs, or jobs started from a CLI, move across hardware without a restart. A dedicated deployment keeps the controller and checkpoints in the customer's own account. On a customized setup, the company says it can move up to 5 TB of state inside one preemption notice. Ju Lin, co-founder and CEO, posted the launch.