DocsLimits

Reference

Limits

What Cloudex sandboxes can't do: no public IP or inbound TCP and UDP, no /dev/kvm or GPU, Docker on the host network only, size and time limits, and how saves expire.

Sandboxes are Cloudflare containers, not full virtual machines. If you need any of the things below, a hosted VM service is a better fit. Sandboxes on your own nodes have other limits, listed at the end.

Network

  • No public IP address, and no inbound TCP or UDP ports. Nothing on the internet can connect to a sandbox directly. Reach it through cloudex ssh and ssh port forwarding, the API’s TCP WebSocket, or HTTPS preview links.
  • Outbound connections to the internet work.
  • Provider keys cover Anthropic and OpenAI only, over HTTPS.
  • Preview links reach web servers in the sandbox over HTTPS, on one domain per deployment, and only while the sandbox runs. Anyone with a link can use it.

Hardware

  • No /dev/kvm. You can’t run virtual machines or hardware-accelerated emulators inside a sandbox.
  • No GPU.
  • At most 4 vCPU, 12 GiB of memory and a 20 GB disk, with standard-4. The disk includes the system image. See Sizes.
  • The image is for x86-64 only.

Docker

  • Containers need the host network: docker run --network host. A sandbox has no iptables, so bridge networks don’t work. With Docker Compose, set network_mode: host on each service.
  • Rootless Docker doesn’t start.
  • Docker images live outside /home. A save keeps them, but a restore from backup doesn’t.

Time

  • A sandbox stops after --idle without activity, 30 minutes by default and at most 6 hours.
  • Whatever --idle says, Cloudflare stops a container after at most 6 hours of inactivity; an open connection such as cloudex ssh keeps it up. See The 6-hour limit.
  • cloudex exec runs a command for at most 1 hour, with no terminal or standard input. Its output streams without a limit; with --json, or through the API without stream, it keeps the first 1 MiB of each stream. Background commands have no time limit of their own.
  • An upload, with cloudex cp, cloudex run or the files API, is at most 100 MB, compressed: Cloudflare’s limit on a request body.
  • Big disks resume slower: a sandbox with 15 GB of files can take about two minutes to come back, and saving a full disk takes about three.

Saves and backups

  • A save expires 30 days after it was made or last restored. After that, resume restores /home from the backup onto a fresh image, and anything installed elsewhere needs installing again.
  • A save is tied to the image it was made on. Resuming from a save doesn’t pick up a newer image from cloudex deploy; cloudex resume --from-backup does, for /home only.
  • Backups hold /home (without ~/.cache) and the SSH host keys, nothing else, and need R2 turned on for the account.
  • Processes don’t survive a stop. Start servers and agents again after a resume.
  • Deleting a sandbox deletes its backups at once. Its last save stays in your account’s container registry until it expires, within 30 days.

Forks and accounts

  • A fork request makes at most 10 sandboxes.
  • A deployment has one admin key with full control, and access tokens with one of three roles. There are no finer permissions, users or teams.
  • A stopped sandbox whose save has expired has to be resumed before it can be forked.

Nodes

A sandbox on a node has no 6-hour limit, and as many CPUs and as much memory as the node gives it. It has these limits instead:

  • Nodes are Linux on x86-64, with systemd, Docker 28 or later and root. arm64, Synology and Unraid aren’t supported yet.
  • No backups or sizes. Make a named save to copy a sandbox off the node, or to move it.
  • Forks stay on their node. A running sandbox pauses while its disk is copied, about 2 to 3 seconds per GB it has written, or 12 to 13 on Docker’s containerd image store. The copy takes about that much disk again, twice on the containerd store, until its last fork is deleted. Forks of forks go about 100 levels deep. See Forks on a node.
  • No Docker inside a sandbox under plain Docker; with Sysbox it works.
  • Plain Docker is for trusted code only: code that escapes its container is root on the node. See Isolation.
  • No disk limit per sandbox on most machines. Sandboxes share Docker’s disk on the node; the node refuses new ones when fewer than 10 GB are free, and stops the sandbox that wrote the most when fewer than 5 GB are (see Disk). Where Docker’s storage driver has quotas, the installer’s --disk-gb caps each one.
  • Everything but ssh goes through Cloudflare. Commands, previews and the dashboard are a little slower than on Cloudflare. ssh goes straight to the node from its LAN or tailnet, and through Cloudflare elsewhere. A sandbox can’t be reached through Cloudflare while its node is offline.
  • One inbound port: direct connections. The node publishes port 7022 on its private and tailnet addresses. Anything on your LAN or tailnet can reach it and get a TLS handshake, and nothing more without a ticket from your deployment. --direct off closes it. See The direct port.
  • Direct sessions are capped. A sandbox takes 64 open direct sessions, a node 256, and one token 64 on a node. A token gets at most 64 tickets for one sandbox a minute; past any of these, ssh goes through your deployment instead.
  • No previews, cp or exec over the direct path. Forward a port over ssh instead: cloudex ssh n1 -- -L 3000:localhost:3000.
  • Direct addresses don’t follow DHCP. When the node’s LAN address changes, run the installer again, or use its tailnet address.
  • A sandbox stays on its node. Moving it means a named save and a new sandbox, with a new id and new preview links.

For agents and LLMs: this page as Markdown,llms.txt and all docs in one file.