Firecracker microVMs at one cent per GB-hour: two cents an hour for a small one, billed by the minute. Give your agent the luxury of a sandbox all its own. It will thank you for it.
| size | machine | while it runs |
|---|---|---|
| small | 1 vCPU · 2 GB | €0.02 / h |
| default | 2 vCPU · 4 GB | €0.04 / h |
| large | 4 vCPU · 8 GB | €0.08 / h |
| xlarge | 8 vCPU · 16 GB | €0.16 / h |
At zero, running boxes pause with their memory kept — nothing is lost, nothing is overdrawn.
Each task gets a box of its own. The moment the agent hands back to you, the box pauses with everything in memory and the meter stops, for as long as you take. Your reply wakes it where it was. Nothing merges until you press Accept. It runs on your Anthropic API key: we sell the machine, not the tokens.
Your repos, your agent, your clouds, your team’s tools. Each one is an app the organization installs, and a task reaches it without ever holding its keys.
beta open to early organizations, on request.
Three things a box does here: a pause that keeps its memory (processes, sessions and open terminals come back, and a paused box doesn't count against your plan), a fork of a box while it runs, and snapshots you start new boxes and Claude Code tasks from. And each box is a dev environment a person can open — an IDE in the browser, terminals, a URL for every port. What it doesn't do yet: SSH, moving a snapshot or a fork to another machine, public port URLs, inbound ports, IPv6, and images from a private registry.
On our own servers, in a single location for now. Each box is a Firecracker microVM with its own kernel (a 6.18 we build and pin), its own network namespace, its own unprivileged user on the host and its own CPU limits.
That's the point of a microVM: a kernel per box rather than a shared one. On top of it, private address ranges and cloud metadata are blocked, outbound port 25 is closed, and the network is shared fairly — a box is capped at 200 Mbit/s each way. Not yet: per-box disk I/O limits.
Anything that can call an HTTP API: create a box, run commands in it, pause, fork, snapshot. For conversations in Agents, Claude Code is built in — Opus, Sonnet, Haiku or Fable — on your own Anthropic API key, stored as a secret. Any other command-line agent runs as a custom command that reads its prompt on stdin. We sell the machine, not the tokens.
Spunto never does it for you. The agent works on a branch of its own and is told to push there — nothing more. The conversation then stops at “in review”, its box paused and off the meter, until you press Accept (which runs your command, say gh pr merge) or Drop.
One cent per GB-hour, while a box runs: it pays for its RAM, by the started minute — two cents an hour for a small one, four cents for an agent's default box. Paused, stopped or building, it costs nothing, and disks are never billed. You start with a 7-day trial and €2 of credit: card required, nothing charged for 7 days, cancel anytime. The subscription is €20 a month, credited in full to your usage, with up to 16 GB running at once; top-ups of €10, €20 or €50 never expire. At zero, running boxes pause with their memory kept.
Start a box or a conversation on the trial credit. When you want more, €20 a month goes straight back into your usage, and a paused box never counts.
Got a server already? Connect it — your own machines run sandboxes free, no credit used.