Checkpointed workstations for coding agents
Unkindled is a Linux VM you SSH into, with Herdr preinstalled. Take a whole-machine checkpoint before your agent runs, fork it to try a second approach, and roll the disk back when something breaks. Not a sandbox that prevents damage — a machine that survives it. Your API keys stay in your home directory, not on our disk.
$5 credit on signup. Card or crypto. Ubuntu 24.04, root access.
agent@unk-04:~$ unkindled checkpoint --label "before refactor" ✓ ckpt_7f3a 12.4 GB disk · 1.8s agent@unk-04:~$ unkindled fork ckpt_7f3a --name unk-04-b ✓ unk-04-b ready 8 vCPU / 16 GB · 4.2s agent@unk-04:~$ unkindled restore ckpt_7f3a · attaching snapshot layer · restoring ext4 metadata 2.1 GB / 2.1 GB · restoring postgres data 3.4 GB / 3.4 GB · restarting postgresql, docker, herdr ✓ restored in 38.2s ssh agent@unk-04
The problem
A bad agent run does not stop at the repository. It installs packages, migrates the database, rewrites environment files, fills Docker layers, and leaves every open session in a state you cannot reconstruct.
apt packages, systemd units, /etc, kernel modules. None of it is in your diff.
Postgres rows, Redis keys, SQLite files, local caches. "Just revert the commit" does not touch them.
Node 18 to 22, Python 3.11 to 3.13, rebuilt native modules. Reproducing the build takes longer than fixing it.
Herdr panes, shells, environment variables, open files. Your agent's working context is not a git object.
On branch main Your branch is up to date with 'origin/main'. nothing to commit, working tree clean
! 3,412 files changed outside the repo ! postgres schema at rev 0441, expected 0438 ! node_modules built against node 18 ! 2 Herdr sessions orphaned ✓ git: clean
Same workspace. Fourteen minutes apart.
Sandbox vs. Unkindled
Die. Restore. Keep the progress. Both models shrink the blast radius; they pull in opposite directions. A sandbox contains the damage. Unkindled reverses it.
| Sandbox | Unkindled | |
|---|---|---|
| Safety model | Containment: stop the damage escaping | Reversibility: undo the damage |
| Assumes the agent is | Hostile or untrusted | Fallible, like you |
| Time horizon | Ephemeral, per run | Persistent, long-lived |
| State | A fresh template every time | Your real DB, caches, and sessions |
| When it breaks | Kill it and start over | Roll back to a checkpoint |
| What is protected | The host and other tenants | Your afternoon's work |
| The sentence | "You may not break anything." | "You may break this; we put it back." |
Contained at the edge, reversible inside. Abuse, egress, and cross-tenant access are still isolated. The difference is what happens to your work when a run goes wrong: nothing you checkpointed is lost.
How it works
There is no rebuild step and no image bake. Unkindled snapshots the disk, clones it, and puts it back.
Freeze the whole disk: OS, Postgres, Docker layers, toolchains, caches, and Herdr session state. Copy-on-write, so it takes about two seconds and does not double your storage.
1.8s median40 GB disk
Clone a running machine at the checkpoint. New hostname, new SSH key, identical disk state. Point a second agent at a second approach instead of arguing with one context window.
4.2s mediannew VM
Roll the machine back. Services restart, your SSH session reconnects, Herdr reattaches. This is not git reset with extra steps; the database rolls back too.
p50 38sp95 54s
Restore times are measured, not modeled: p50 38s and p95 54s across production restores in the last 30 days. Disk size and host load move those numbers. We publish both on the status page.
Agent skills
unkindled install-skills writes a portable skill into the harness you
already use. The agent learns when to checkpoint, how to fork, and when to stop and
ask before a restore. Plain Markdown; no plugin, no SDK.
$ unkindled install-skills --harness all --scope user ✓ claude ~/.claude/skills/unkindled/SKILL.md ✓ opencode ~/.config/opencode/skills/unkindled/SKILL.md ✓ codex ~/.codex/AGENTS.md (block updated) ✓ hermes ~/.hermes/skills/unkindled/SKILL.md ✓ cursor ~/.cursor/rules/unkindled.mdc ✓ gemini ~/.gemini/GEMINI.md (block updated) ✓ generic ~/.agents/skills/unkindled/SKILL.md 7 targets · idempotent · backups on write
--scope project writes into the repo instead;
--dry-run prints every path before touching disk.
Harnesses we write to
Not on the list? --dir <path> writes the same portable
SKILL.md anywhere, and --harness auto detects what you
already have. Codex and Gemini also get a marked block in
AGENTS.md / GEMINI.md.
Interactive demo
Pick a checkpoint and restore it. This runs the same control flow as the real command, compressed to about three seconds. No backend, no signup.
Simulated at 40× speed. A real restore is p50 38s.
Idle.
The math
A $6 VPS is cheaper right up until it costs you a day. Put your own numbers in.
Cost of lost time, per month
$720
That is 9.1× the Pro plan at $79/mo on annual.
Excludes the rebuild: reinstalling packages, replaying migrations, and re-reading the logs. That is usually the larger half.
Pricing
Every plan is a full VM with root, a public IPv4, and 1 TB of egress. Active hours are hours the machine is powered on. Hit the cap and it suspends. No overage invoice.
One agent at a time, with room to undo.
billed monthly · cancel anytime
Most popular
For developers running agents most of the day.
billed monthly · cancel anytime
Parallel forks for small teams and heavy builds.
billed monthly · cancel anytime
Hard cap, hard stop At your active-hour cap the VM suspends. Resume next cycle or upgrade mid-cycle and pay the prorated difference. Nothing bills over the cap.
Checkpoints, priced 3 included on Solo, 5 on Pro, 10 on Studio. Extra retention is $0.10/GB-month. Unlimited snapshots are how a $39 plan becomes a $12 storage bill; we cap it and show the meter.
Support, priced Docs and community are included on every plan. Human support is a $39/month add-on, included on Pro and Studio. No 24/7 on-call; a small team answering email.
Annual by default Save 20%, billed once a year. Annual prepay funds capacity up front; that is why the annual rate is $39 and the monthly rate is $49. Cancel anytime; unused time is not refunded.
Teams
SSO, shared billing, a shared checkpoint policy, and priority support. From $299/mo for five seats, annual by default.
Trust and security
Unkindled is built for people who read the privacy page before signing up. Here is what we do, and what we do not.
Tenancy
Each workspace is a KVM guest with dedicated volumes. No shared kernels, no noisy neighbors, no cross-tenant snapshot store.
Keys
Model API keys are written inside the VM at ~/.config and never copied
to the control plane or image store. We do not proxy model calls.
Logging
We do not inspect your disk, terminal, or agent traffic. Metrics are CPU, disk, network, and uptime. Debug access needs a time-boxed grant from you.
Portability
Download the full disk as qcow2 and boot it in QEMU or Proxmox. No exit interview and no migration fee.
Deletion
Workspaces are destroyed on request. Backups and snapshots are purged within 30 days. The deletion receipt lists every object removed.
Operations
status.unkindled.dev carries incident history and postmortems. Restore latency is on the same page, including the bad weeks.
Assurance
Monthly image and dependency scan reports with SBOMs. Run
unkindled scan inside your own workspace at any time; uploads are
redacted and opt-in.
Supply chain
Skill bundles are scanned before install-skills writes them. Critical
findings block the install; high findings require an explicit, audited override.
Limits, stated plainly
Comparison
If you have the time, a Hetzner box plus a snapshot script is cheaper. We sell the part you do not want to build, test, and carry a pager for.
| Feature | Unkindled | Hetzner + DIY | Codespaces | E2B-style sandbox |
|---|---|---|---|---|
| Floor price | $39/mo, 160 h, annual | $6–15/mo + your scripts | $0–39/user/mo, metered | Metered per sandbox-hour |
| Herdr preinstalled | Yes | No, you install it | No | No |
| Snapshot scope | Whole machine | Whole disk, if you script it | Repo + devcontainer | Sandbox filesystem |
| Fork a running machine | Yes, 4.2s | No, manual clone | No | New sandbox from a template |
| Typical restore | p50 38s | Depends on your scripts | Container rebuild, minutes | Re-provision from image |
| SSH + root | Yes, default | Yes | Browser or CLI tunnel | SDK-first |
| Agent sessions survive a restore | Yes | Only if you snapshot disk state | No | No |
| Keys stay on your disk | Yes | Yes | Org secret store | Env vars per sandbox |
| Export the disk | qcow2 | Raw disk | No | No |
| You maintain the snapshot tooling | No | Yes | No | No |
| Support | Community; human add-on | None | GitHub issues | Paid tiers |
Scroll sideways to compare all four. The feature column stays put.
FAQ
You can, and it will cost less. A comparable CX-series server is €8.49–€16/mo depending on the size. You will write the snapshot script, test restores, handle qcow2 consistency, wire up session state, rotate SSH keys, and be on call when a restore corrupts Postgres. Unkindled is that work, priced at $39–$149. If your time is free, DIY wins. If you bill more than $39 an hour, it does not.
No. We do not read your disk, we do not log terminal or agent traffic, and there is no model training anywhere in the product. We collect host metrics and billing counters. Support access requires a time-boxed grant from you.
The VM's disk is rolled back to the checkpoint's copy-on-write layer. Files,
packages, Postgres data, Docker layers, and /home come back. Running
processes do not; services restart and your SSH connection reconnects, usually in
under a minute. Anything written after the checkpoint is gone. That is the point.
Yes. Download the disk as qcow2 from the dashboard while the VM is running, then boot it in QEMU, Proxmox, or another KVM host. There is no exit fee; the download counts toward your 1 TB of egress.
The workspace suspends. You can resume next cycle or upgrade mid-cycle and pay the prorated difference. There is no overage rate and no surprise invoice. Active means powered on, not SSH-connected, so a forgotten idle VM still counts toward the cap.
Docs and community are included on every plan. Human support is $39/month (email, one business day), included on Pro and Studio. Studio adds a 45-minute onboarding call and a priority queue. There is no 24/7 on-call; we are a small team and we say so. Support is the largest cost line in a business like this, so we price it instead of hiding it in a higher sticker price.
If all you need is a box with Herdr installed, buy theirs. It is a good deal, and we will not pretend otherwise. Unkindled is the layer above it: checkpoint the whole machine, fork a running workspace, and restore in under a minute when an agent breaks the OS, the database, or the toolchain. A $6.49 VPS does not include snapshot orchestration, restore testing, or someone to call when a rollback corrupts Postgres. You are paying for the undo button and the pager, not the CPU.