FAQ
Commonly asked questions
What developers and teams ask us most, grouped by what you're trying to decide.
Getting started
What is boxd?
Persistent Linux machines in the cloud for developers and their agents. Each one is a real virtual machine with its own kernel, root access, SSH and an HTTPS URL. It keeps its files and running processes between sessions, and you can fork, snapshot or hibernate it.
How do I get started?
Install the CLI with curl -fsSL https://boxd.sh/downloads/install.sh | sh, run boxd auth login, then boxd machine new my-app. It's ready in under 10ms. Connect with ssh my-app.boxd. New accounts get €30 of free credits.
Is there an API or SDK?
Yes. There are Python and TypeScript SDKs, with a gRPC API underneath. Every CLI command also takes --json, so scripts and agents can drive it directly.
What's installed on a machine?
Ubuntu 24.04 with systemd, passwordless sudo and Docker. Claude Code, Codex and OpenCode come preinstalled, along with Node, Python, Go, git, headless Chrome, ffmpeg and nginx.
Can I use my Claude Code subscription?
Yes. Sign in once and new machines come up signed in. For many machines in parallel, run claude setup-token and store the token as the boxd secret CLAUDE_CODE_OAUTH_TOKEN.
Can I run Docker?
Yes. Every machine is a full VM with Docker installed and your user in the docker group, so containers run the way they do on your laptop.
Can I use a distro other than Ubuntu?
The distro, yes. Machines boot from a container image, so a base like SUSE with your packages works. We set it up with you for now. The kernel is ours, because fast boot, fork and snapshots depend on it.
Does the CLI work on Windows?
Not yet. The CLI runs on macOS (Apple Silicon) and Linux, and Windows is on the list. You can still SSH into a machine from Windows.
Machines and state
How long can a machine run?
As long as you want. There is no session limit. By default a machine with no inbound traffic for 4 hours hibernates and wakes on the next connection with its processes intact. For jobs that run without network traffic, turn that off with boxd machine config set <name> auto-hibernate.timeout 0.
What is a fork?
A copy of a running machine, memory, processes and disk included, made in under 200ms. The original keeps running. Use it to try two ideas at once or to hand out copies of one prepared machine.
What's the difference between a snapshot and a checkpoint?
A checkpoint is a save point on the same machine: roll it back to exactly where it was. A snapshot is a saved image you start new machines from.
What's the difference between standby, hibernate and stop?
Standby keeps memory warm for an instant resume and pays for RAM and disk. Hibernate writes memory to disk, pays for disk only and wakes on the next connection. Stop also pays for disk only, ends running processes and stays off until you start it.
Can I get a bigger machine?
Yes. Machines come in 1 vCPU / 4 GiB, 2 vCPU / 8 GiB (the default) and 4 vCPU / 16 GiB, and you can pick any of them yourself. For 8 vCPU / 32 GiB or larger, email contact@boxd.sh. Resize an existing machine with boxd machine resize <name> --vcpu 4 --memory 16G, which restarts it.
How many machines can I run?
50 per organization to start, including forks and hibernated machines. Email contact@boxd.sh to raise it.
Pricing and billing
How does billing work?
Credit-based. You buy credits and your organization's machines draw them down at the rate card: €0.039 per vCPU-hour, €0.012 per GiB-hour of RAM, and €0.00008 per GiB-hour of disk. RAM bills on memory in use and disk on what you have written to, not on the provisioned size. A default machine (2 vCPU / 8 GiB) is about €0.17 per hour at full memory while it runs, and about €0.1 holding 2 GiB with 20 GiB written to disk. There is no monthly plan and no per-seat fee.
Do you bill on reserved or actual memory and disk?
Actual, for both. RAM bills on the memory a machine actually has resident, sampled every second, while it is running or in standby, not on the 8 GiB it was provisioned with. Disk bills on the disk you have written to, in every state, not on the 100 GiB provisioned: a machine that has written 20 GiB pays for 20 GiB. vCPU is the one component billed on the machine's size: its full vCPU count while it runs, nothing in standby, hibernated or stopped.
What do I get for free?
€30 of credits. Nothing is charged when you add a card. When the credits run out, auto top-up adds €20 at a time. Change the amount or switch it off under Billing.
Do I pay while a machine is idle?
Only for its disk. A hibernated or stopped machine pays for the disk it has written and nothing for CPU or RAM.
How are forks billed?
Making a fork is free. A fork then bills like any other machine for the time it runs.
How do teams work?
Invite your team into an organization to share machines and draw from one credit balance. Everyone's usage bills to the organization.
Security and isolation
How is a machine isolated?
Every machine is a KVM virtual machine with its own Linux kernel, not a container on a shared host kernel. Code inside has root over its own machine and no path to the host or to other customers' machines.
Can I run untrusted, agent-written code?
Yes, that's what it's built for. The agent gets root inside its own machine, and the hardware boundary keeps it away from the host and other customers.
Can I lock down what a machine can reach?
Yes. Give a machine an egress allowlist and it can only connect to the hosts you name. Forks inherit the allowlist. Create a machine with --isolated and boxd leaves out the in-machine boxd CLI, your integrations and your agent logins, and it can't reach your other machines.
Can I keep a machine off the public internet?
Yes. Every machine gets an HTTPS URL by default. To keep machines private, put your organization behind Tailscale or serve them on your own domain.
How are secrets and tokens stored?
Integration tokens and secrets are encrypted at rest and stay on boxd's side, not in your machine. GitHub credentials are fetched only when git or gh runs. Secrets you set with boxd env set --secret are sealed and never returned by the API.
What happens to my data when I delete a machine?
Its disk is removed straight away. Its backups stay until their retention runs out, 7 days by default, or you can delete them first. Backups are kept in EU object storage and sent over TLS.
How do I report a vulnerability?
Email contact@boxd.sh with the details.
Hosting and compliance
Where do my machines run?
On bare-metal servers in Germany, so your machines and their data stay in the EU. You can also self-host boxd in your own datacenter, cloud or country.
Is there a US region?
It's coming. Join the waitlist at boxd.sh/us-waitlist.
Do you sign a DPA?
Yes. Email contact@boxd.sh and we'll send our Data Processing Agreement for signature.
Are you SOC 2 or ISO 27001 certified?
We're working on both. Email contact@boxd.sh if you need our security details for a review.
Do you support SSO?
You sign in with GitHub or Google. Talk to us about SAML SSO for Enterprise.
Self-hosting
Can I self-host boxd?
Yes. boxd ships as a single binary that runs every role, on your own KVM-capable Linux hosts. No Kubernetes, no external Postgres, no Redis.
What do I need to self-host?
Linux hosts with KVM and Intel VT-x or AMD-V exposed. Most clouds and bare-metal providers support it. A production cluster runs at least three control nodes.
How is self-hosting licensed?
An annual licence, priced for your setup. Email contact@boxd.sh with your team size and where you want to run it.
Agents
Why do agents need a persistent machine?
Agents run for hours, not seconds, and need somewhere to come back to. A sandbox that resets loses the agent's work and makes it reinstall everything. A boxd machine keeps files, packages and processes, so the agent stops at 2am and picks up at 9am.
How does forking help an agent?
Fork the machine mid-task and give each copy a different approach. Every copy has the full running stack. Keep the one that works and delete the rest.
Can an agent manage other machines?
Yes. Every machine has the boxd CLI, already signed in, so an agent can create, fork and run commands on other machines. Create the machine with --isolated to turn this off.
boxd vs the alternatives
Most tools give you a sandbox that lives for a task or a dev container that bills while idle. boxd gives you a real machine that remembers.
boxd vs E2B
E2B is a sandbox SDK with a cap on how long each run lasts. boxd is a machine you and your agents can work in for as long as you need.
Pick boxd when
- Your agent runs for days. E2B caps a continuous run at 1 hour on Hobby and 24 hours on Pro.
- You want plain SSH. On E2B you build an SSH template and tunnel over WebSocket.
- You want Claude Code, Codex and OpenCode installed when the machine starts.
- You want EU hosting by default. E2B's EU cluster is Pro and up, on request.
Pick E2B when
You want short code runs from an SDK on an open-source runtime.
boxd vs GitHub Codespaces
A codespace is a dev container opened from a GitHub repo. A boxd machine is a VM for you and your agents, with any repo or none.
Pick boxd when
- You want to self-host. GitHub says Codespaces can't be.
- Your agents run unattended. A codespace stops after at most 4 hours idle.
- You want to fork a running machine per developer or pull request, not rebuild one from a prebuild.
Pick GitHub Codespaces when
Your team lives in GitHub and VS Code and sets up environments with devcontainer.json.
boxd vs Vercel Sandbox
Vercel Sandbox runs sessions of up to 24 hours and restores the disk between them. A boxd machine wakes with its processes still running.
Pick boxd when
- Your processes have to survive a pause. Vercel restores the files and you restart services yourself.
- You want to fork without stopping the original. A Vercel snapshot shuts the sandbox down.
- You don't want a session limit. Vercel stops a session at 45 minutes on Hobby and 24 hours on Pro.
- You want machines that aren't tied to one hosting platform.
Pick Vercel Sandbox when
Your app already lives on Vercel and you need to run a short piece of code next to it.
boxd vs Modal
Modal sandboxes are containers with a 24-hour maximum life. A boxd machine is a VM that stays until you delete it.
Pick boxd when
- Your work runs past 24 hours. A Modal sandbox ends at 24 hours at most.
- You want to fork a running machine. Modal's memory snapshots are experimental, end the sandbox and expire after 7 days.
- You want to SSH in and work on the machine yourself.
- You want EU hosting at the normal price. Modal charges 1.15x to 1.75x to pin a region.
Pick Modal when
You need GPUs or want to fan a Python job out across thousands of containers.
boxd vs Fly.io Sprites
Sprites are persistent VMs on Fly.io at one fixed size. boxd adds live forks, a choice of sizes and a self-host option.
Pick boxd when
- You want to fork a running machine. Sprites have no fork, and checkpoints save disk, not memory.
- You want a sleeping machine to wake with its processes intact. A Sprite drops its memory when it goes cold.
- You want to run it on your own servers.
Pick Fly.io Sprites when
You're already on Fly.io and want a fixed 8 vCPU, 100 GB machine with nothing to size.
boxd vs exe.dev
exe.dev is the closest to boxd: persistent VMs over SSH with coding agents installed. The difference is what a copy keeps.
Pick boxd when
- You want a copy to keep memory and running processes. exe.dev copies disk and config.
- You want to pay for the RAM your machine uses, not a fixed capacity tier.
- You want to run it on your own servers.
- You want everything in the EU. exe.dev runs its main server in the US for every account.
Pick exe.dev when
You want a flat monthly plan with their Shelley agent built in.
boxd vs Daytona
A Daytona sandbox is a container unless you pick a VM snapshot. Every boxd machine is a full VM.
Pick boxd when
- You want every machine to fork and pause with memory. On Daytona only the VM class can.
- You want to self-host a maintained release. Daytona's open-source repo stopped updating in June 2026.
- You want Claude Code, Codex and OpenCode installed, with GitHub, Linear and Slack one click from connected.
Pick Daytona when
You need GPUs, Windows machines or SDKs in Ruby, Go or Java.