Machine0 Launch Offers Persistent CPU and GPU VMs for AI Agents via CLI

According to the post, the service offers configurations up to 60 vCPUs, 240 GB RAM, and GPUs including H100 and H200, with a 99.99% VM-level uptime. Agents can manage machines directly via the CLI or through the MCP (Model Context Protocol). The tool is designed for long-horizon agent workloads, such as coding agents running six to eight hours, orchestrated training and reinforcement learning runs that take days, and always-on agents like OpenClaw and Hermes.

The post outlines three primary use cases. In agent fleets, a pilot agent scopes work and delegates it to sub-agents, each on its own VM; one customer reportedly runs hundreds of machines at once, spun up and torn down from the CLI. For model optimization and RL environments, machine0 supports agent-orchestrated RL environments and inference-speed optimizations, with one customer running RL environments on 60-vCPU machines for days and another keeping a suspended H100 for overnight optimization work. In product infrastructure, one company builds its product on top of machine0, launching thousands of machines per user session, most alive for about two minutes.

Every machine is a full KVM virtual machine, not a container or sandbox, with the actual GPU exposed to the guest and no syscall-interception layer. The stack is described as deliberately dull: TypeScript, Postgres, Redis. The tool also supports suspend, snapshot, and resume, persistent block storage from 10 GB to 16 TB, and reproducible builds via NixOS flakes or Ansible playbooks.

Barnaby states he has built cloud infrastructure for 15 years, dropped out of a PhD at Imperial College London, and spent six years as co-founder and CTO of Upflow, a YC W20 company. He says machine0 started as a personal tool and he is his own first user.