The platform, verb by verb.
What each command does, where it runs, and what you see in the terminal. Features marked “Coming soon” are not open to every account yet. The rest work today in the panel and through the REST API; the cf commands shown here come with the cf CLI.
git branch, but for the running server.
Spinning up a staging clone of a 200 GB DB is a weekend project. So nobody does it. So you patch in prod.
cf branch prod — a running copy of a ComputeForge server in a minute or two, on its own address behind a firewall that lets SSH in only from yours. Run the destructive thing on the copy. Diff against the original. Throw it away, or merge its file changes back. In the panel today: a server's Branches tab. Copying a server from another provider onto our cloud, or branching at DigitalOcean, Hetzner or AWS with your own API token: Coming soon
Stop SSH'ing in to figure out what changed.
Something broke. You don't know when, you don't know who, you don't know what. The logs were rotated. auditd was off. You're on a forensic dig with a flashlight.
Every change to every tracked path is recorded. cf log shows the full history. cf diff @{12.hours.ago} shows what's different from this morning. cf blame /etc/nginx/api.conf shows which login changed each line last, and when. In the panel today: a machine's History, Compare and Files tabs.
Tag the known-good state. Roll back when the next deploy is wrong.
The deploy went sideways. The "last good state" is a snapshot you took two weeks ago and a memory of what you fixed since. Restore is an ordeal of rebuilds and runbooks.
cf tag prod --as before-pg17 names the current state. cf checkout prod --tag before-pg17 restores it: tracked files come back with no downtime, on any connected server where you allow remote changes. On our cloud, a whole server goes back to a tagged checkpoint on the same IP addresses: a short swap for newer servers, an in-place rebuild of a few minutes (longer for bigger disks) for older ones. Need to revert one change instead? cf revert /etc/nginx/api.conf @{yesterday}. In the panel today: a machine's Tags tab, and Undo this change in its History.
git push, but it's your VM going to a new provider.
Moving a VM between hosts is a weekend project nobody wants. Backup, provision, restore, fix the bootloader, swap DNS, hope the cutover doesn't tear.
cf push hetzner://target does the whole dance. We create the target, copy the running server across while it keeps serving, copy what changed since, stage DNS at low TTL, and hand you the cutover button. --commit stops your services for one short final copy and swaps DNS; --rollback stays open for 24 hours. Works in any direction — TO us, AWAY from us, BETWEEN third parties — between servers on the same Linux distribution and major version.
DNS, probes and alerts in one place. Failover when a probe goes red.
Cloudflare DNS over here, UptimeRobot over there, a Slack webhook in another tab, a runbook in a fourth. For what should be one thing.
Forge DNS holds your zones. Lifeguard probes any VM (yours or ours) from our datacenter at TCP/HTTP/HTTPS/ping, or runs your custom script on the box through its agent. When a probe goes red, the record switches to your standby, and alerts go out by e-mail, Slack, Discord, ntfy or webhook. Undoing a failover (cf incident undo) is a manual override: Lifeguard keeps the record on the primary until that incident resolves, even if the probe is still failing.
Stop being on call. Start being protected.
Polling-based monitoring is 2010s. By the time you get the page, your customers already noticed. Then you SSH in, triage, fix, write a postmortem. You're paid to be a smoke detector — not to design, not to ship.
Every VM's agent keeps a persistent connection to our control plane, so we hear about a change as it happens, not on the next poll. AI baselines what "normal" looks like for each VM. When the pattern deviates, we trigger cf scan automatically, seal a forensic capsule, and propose the right action: fail DNS over to a standby, branch the box for inspection, or check it out to its last good tag. You choose: full-autonomous, propose-and-approve, or alert-only. Either way, you're not the smoke detector anymore.
Stop spinning up servers by hand.
You know what each box does. You know what depends on what. That knowledge is in your head, your wiki, and a runbook nobody's read in eight months. When something breaks at 3am, the runbook is wrong, the wiki is wrong, and you're the only person who can re-derive the topology from memory.
Forge Topology is a declarative description of your stack — services, roles, instance counts, dependencies, health probes, standbys, autonomy modes. Write it in YAML, or describe your stack to Forge AI in plain English and review the generated spec. Connect existing VMs and we infer the topology from what's already running. Once declared, anomaly + topology = the right action. "prod-web-2 compromised → fail over to the standby via Forge DNS, branch prod-web-2 for forensics, check it out to its last good tag."
services: prod-web: role: webapp instances: 3 image: golden/ubuntu-24.04-app depends_on: [prod-db, prod-redis] health: { http: /healthz, expect: 200 } standby: prod-web-standby autonomy: propose-and-approve prod-db: role: postgres instances: 1 autonomy: propose-and-approve # never auto on db staging: role: webapp instances: 1 autonomy: full-autonomous
Every server you connect, on the same private network.
Your three VMs are at three providers. Stitching them into a private network is YOUR problem. Hosted mesh services charge per user once you outgrow the free tier. WireGuard from scratch is a yak shave.
One click puts a connected server on your private WireGuard overlay. Resolve db from web (full name db.<account>.mesh) regardless of which provider hosts which. Extends to your laptop, off-platform machines, your homelab. Laptops join with the standard Tailscale client; the mesh runs on open-source WireGuard tooling that we host. In the panel today: Mesh, and Add to Forge Mesh on a machine's Settings tab.
AI that reads the diff.
You ran cf diff and got 47 changed lines in postgresql.conf. What do they mean? Which one's risky? You ask Stack Overflow.
Forge AI reads the diff and explains in plain English. "You raised max_wal_size from 1GB to 4GB — this allows larger checkpoint intervals before WAL recycling, useful for write-heavy workloads." Then proposes the exact cherry-pick. Apply with auto-rollback if Lifeguard fails health within 60 seconds.
postgresql.conf · shared_buffers 256MB → 1GB
Real backups in a format you own.
Snapshots usually stay inside the host that made them. You can't download a DigitalOcean snapshot, so it won't restore at Hetzner.
Tag a server on our cloud and export that checkpoint into your own Borg repository over SSH, or to Restic: encrypted with a passphrase only you keep, restorable on any Linux box, mountable with borg mount. DNS zones export as standard BIND. Mesh config exports as plain WireGuard, and we never hold your private keys. Your data, your format.
Stop trusting the OS to tell you the truth.
A rootkit's whole job is lying to the OS. ps doesn't show it. ls doesn't show it. Your in-VM antivirus reads the same lies. You find out weeks later — when a customer calls.
cf scan checks your disk from outside the running OS: on our cloud, out-of-band, from a copy of the disk read by another machine; on other providers, in-guest, by cross-checking what the OS reports against what is on disk. Hashes binaries against the package manager's published hashes. It looks for hidden files, modified system binaries, unexpected kernel modules and persistence entries. Works on any provider. In the panel today: a machine's Integrity tab (Scan now; on our cloud, a scheduled integrity watch too).
More CPU and RAM on a running server. Billed by the second.
Your plan was fine until launch day. A bigger plan means a restart at the worst moment, and then you pay for the big plan long after the rush is over.
Raise CPU and RAM on a server while it runs, and lower them again when the load drops. No restart, no new server, same addresses. The extra is billed by the second, only while the server holds it. Pick when it ends, or keep it until the next restart; either way the server goes back to its plan size, and you get a warning before it does.
RAM that follows the load, inside limits you set.
Size for the peak and you pay for idle memory all month. Size for the average and the out-of-memory killer picks a process for you on the busiest night.
Turn on auto-RAM and set a floor, a ceiling and a spending limit. When the server stays short of memory it gets more; when it has been quiet for a while it gives a step back. It never goes past your ceiling or your limit, and a change you make by hand pauses it.
Grow a disk while it runs.
The disk fills up at 2 a.m., and stopping the server to resize it means downtime at the worst moment. Going back to a smaller plan later means a new server and a weekend of copying.
Grow disk makes the system disk bigger while the server keeps running; the panel shows how to grow the file system, at the next restart or right away with one command. Shrink disk Coming soon moves a server to a plan with a smaller disk without a reinstall: the data comes along, the server is checked to boot, and you can undo it for a while if something looks wrong. Disks it cannot shrink safely are refused before anything changes.
Boot your own ISO.
The OS you need is not in the list, or the server will not boot and you want your own rescue tools, and all you can do is open a support ticket.
Add an ISO from a web address, an upload, or a file in your own object storage bucket, and boot a server from it. The server's disk stays attached, so you can install, repair or copy files off it. Detach the ISO and the server boots from its own disk again. You can also create a new server straight from an ISO and install your OS of choice.
Archive a server. Take it with you, or bring it back.
A server you no longer run still costs a full plan, or you delete it and hope you never need it again. And a complete copy you can take somewhere else is hard to get.
Archive a server, an image or a snapshot: the disks are compressed into protected storage in your account, every byte is checked, and only then is the original deleted. Download the archive as a standard compressed disk image (qcow2), or restore it as an image and create a server from it. The archive is billed as object storage. When you promote a branch, the snapshots it was made from can go the same way instead of staying as snapshots: choose it under Source snapshots, and they are removed only after the copy is checked in full.