Pre-release. HareBox is a working local build with no public service yet. Statuses on every feature say what runs today; speed and reliability figures are targets until measured. Status
HareBox gives each agent its own full Ubuntu VM: root, systemd, Docker and nested KVM, with its own kernel. Put it to sleep and it wakes with its programs still running. Fork it and the copy is exact. Pay by the second, with free egress, and keep your data in the EU.
$0.03564 an hour for 4 vCPU and 8 GB. EU only. A 7-day trial with a card, charged nothing until day 7.
An illustration of the flow. Timings are targets until measured.
Agents install packages, run Docker, start servers, open browsers and work for hours. A container or a function can't hold all of that. A VM with its own kernel can, and HareBox makes it cheap to keep, copy and pause.
Stopping a VM saves its disk and ends its programs. Sleep saves memory too, so a dev server, a debugger or an agent halfway through a task carries on the moment it wakes. Use it for quick returns; stop the machine to end the bill.
Real VMs, locallyCopy a running VM with its memory and latest writes. Each copy gets a new identity, keys and entropy.
Real VMs, locallyFollow an agent over SSE, WebSocket or signed webhooks, resuming from a cursor after any disconnect.
Works locallyThe sandbox holds a placeholder. The real key is added on the way out, only for the host its rule names.
Works locallyHard dollar budgets on an account, project, fleet, label or key, plus a cap on every prompt's AI spend.
Works locallyUbuntu 24.04 on HareBox's own Linux 6.18 LTS kernel, passwordless sudo, systemd, Docker with BuildKit and Compose, and nested /dev/kvm for QEMU, Firecracker or an Android emulator.
Rules are API objects kept outside the VM, so your tenant can be root inside and still can't change them.
SimulatedThe API, CLI, SDKs, dashboard, webhooks and logs use one set of state names. Billing runs per second from the moment a machine is assigned until a stop begins, sleep included. A stopped disk is free for 14 days, then moves to paid cold storage instead of being deleted.
Each guide says who it's for, why it pays, what it costs, and the exact commands, in the CLI and, where the SDKs cover it, Python and TypeScript.
Give every user or project of your product its own persistent computer, agent and preview URL, and bill them for exactly what they use.
See the stepsPrepare one environment, fork it into many exact copies, and give each copy its own task. Keep the winner, delete the rest.
See the stepsHand an issue to Claude Code, Codex or OpenCode in a clean VM, follow it live, cap what it can spend, and get a pull request back.
See the stepsRun real Chrome with persistent profiles on a real desktop, record sessions for audit, and fork a logged-in browser into many.
See the stepsRun tests that need Docker, systemd or KVM on a fresh VM per job, from GitHub Actions or any CI, billed by the second.
See the stepsLet an agent work on a task for days, with its files, tools and browser intact, waking on a schedule or a webhook.
See the stepsFour sizes, one price on every tier, and no regional uplift. Every plan dollar comes back as machine time, a stopped disk is free for 14 days, and egress is free.
| Size | Shape | Per hour | Always on, per month | Runs on |
|---|---|---|---|---|
| small | 2 vCPU · 4 GB · 51 GB | $0.01782 | $13.01 | Trial and every plan |
| default | 4 vCPU · 8 GB · 51 GB | $0.03564 | $26.02 | Trial and every plan |
| large | 8 vCPU · 16 GB · 111 GB | $0.07128 | $52.03 | Every plan |
| xlarge | 16 vCPU · 32 GB · 251 GB | $0.198 | $144.54 | Team plan and up |
Dollars per hour, billed per second. Rate card of 2026-11-07. Plans, fees and a cost calculator.
Every error says what to do next. Output turns into JSON in a pipe. Prices and limits are live API facts, so nothing has to be remembered. An agent can even sign itself up and hand a human the claim link.
llms.txt, an agent skill and an AGENTS.md template$ harebox new --json
{"event": "error",
"code": "start_rate_limited",
"message": "This account has used its 12 starts for this minute.",
"fix": "Retry after 14 seconds, send placement: \"queue\" to wait in line, or move to a larger plan.",
"retry_after_seconds": 14}
HareBox is pre-release. Every feature on this site carries its status, so you can see what runs today and what is still being built.
Works locally Control plane, billing, budgets, the CLI and three SDKs. All 203 API operations answer, some through a labelled simulator.
Real VMs, locally Ubuntu 24.04 with Docker and nested KVM boots on HareBox's own kernel, with sleep, exact fork, stop and resume checked on a development machine.
Not yet Rented servers in France, uptime history, audits, and published benchmarks.
No. It's a working local build and a specification.
With KVM and the Cloud Hypervisor adapter. Otherwise the labelled simulator runs, and every response says simulated: true.
No, that's the point. Open connections may break, and timers inside the VM don't fire.
Yes, like a running one, and it still stops at its auto-stop. Stop it to end the bill.
At least 14 days for free, then it moves to paid cold storage. It's never deleted because of idleness alone.
The default size, 4 vCPU and 8 GB, is $0.03564 an hour, billed per second, at the same price on every tier. Plans start at $20 a month, and every plan dollar comes back as machine time. Egress is free. See Pricing.
Yes: 7 days with a card, including 25 hours of machine time and 2 sandboxes at once on the small and default sizes. Nothing is charged until day 7. See Free trial.
Yes. harebox fork copies it exactly, memory included, and each copy gets a new identity, IP addresses, host keys and entropy. See Parallel agents.
In the EU. HareBox launches in France, at Roubaix and Gravelines, and places each sandbox itself. Data, including snapshots, stays in the EU.
Start from the quickstart, try the dashboard on the local build, or read how other teams use it.