Maritime

Developers building long-lived AI agents need a place for those agents to actually live and work: files, logins, installed packages, and memory that persist across sessions. Conventional deployment models (like serverless, stateless request handlers) don't fit agents that need to pick up a task a week later exactly where they left off. Keeping a full VM running for every agent at all times is wasteful since, per Maritime's own research note, a typical agent is only awake for about eleven minutes a day. Without a dedicated, isolated environment, agents can't safely install software, run arbitrary generated code, or maintain signed-in browser sessions between tasks.

Maritime hosts each agent in its own Firecracker micro-VM with a real kernel, persistent disk, and preloaded dependencies, so state like files, memory, logins, and packages survive sleep, restart, and redeploy. The platform snapshots idle VMs to disk and restores them on the next inbound message, with new machines booting in 0.5 seconds and sleeping ones waking in about 0.6 seconds mid-task, which keeps costs tied mainly to storage rather than idle compute. Agents can apt-get libraries, run headful Chrome, clone repos, train models, or serve web apps and APIs inside this sandboxed machine. Maritime also offers a desktop mode (noVNC-based) where a headful Linux desktop inside the agent's VM can be viewed live or controlled directly, and an MCP server option lets an existing hosted agent be handed a dedicated computer by user id. Access is available through a dashboard, CLI, TypeScript SDK, and Python SDK.