Give me a disposable OCI micro VM, and I will deterministically and declaratively turn it into a lightweight, hardened Alpine machine.
alpine-fleet converts an Oracle Cloud (OCI) VM.Standard.E2.1.Micro instance from the stock Oracle Linux image into a minimal Alpine Linux install with key-only root SSH (no password). Once your OCI CLI is configured, a run is hands-off: it launches the instance, jumps it into Alpine with kexec, installs Alpine to the boot disk over the serial console, locks down SSH, and verifies the result. Proven on the E2.1.Micro across 50+ end-to-end runs.
⚠️ Destructive operation This project installs Alpine onto/dev/sdaand replaces the existing operating system. The orchestrator confirms the installer's "erase disk" prompt automatically. Use it only on disposable or recoverable instances.
pipx install 'alpine-fleet[oci]'
alpine-fleet doctor # checks your host, prints the install command for anything missing
alpine-fleet up # launch (if needed), convert, harden, verifyThe E2.1.Micro free-tier instance has 1 GB of RAM, so what the OS itself uses matters. Measured on OCI PHX over five runs:
| Metric | Stock Oracle Linux 7.9 | Alpine Linux 3.24 |
|---|---|---|
| RAM available to workloads | 288–316 MB | 778–792 MB |
| In use (total − available) | 351–379 MB | 167–181 MB |
| Total the kernel reports | 668 MB | 960 MB |
That is +472 to +500 MB (2.5–2.7×) more usable RAM.
The stock boot reports crashkernel=auto and exposes noticeably less memory to the running system than Alpine's kernel does (this project did not measure exactly how much of that difference is the crash-kernel reservation). The rest of the improvement comes from Alpine's much smaller userspace. A full automated run takes about 315–337 seconds.
pipx install 'alpine-fleet[oci]' # or: pip install 'alpine-fleet[oci]'
alpine-fleet doctor # add --fix to be offered the install commandpip cannot install system packages, so doctor checks them for you (Python 3.9+; Linux, macOS or WSL; native Windows is not supported).
| Tool | Needed for | How to get it |
|---|---|---|
| OCI CLI | Launch, teardown, capacity discovery | Installed by the [oci] extra, then run oci setup config |
jq |
JSON handling in the shell scripts | pacman -S jq, apt install jq, dnf install jq, brew install jq |
OpenSSH (ssh, scp, ssh-keygen) |
Remote staging and serial tunneling | System package (openssh-client on Debian/Ubuntu) |
| bash 4+ | The shell scripts | macOS ships 3.2: brew install bash |
podman or docker (optional) |
alpine-fleet cache (faster runs) |
System package manager |
You also need free E2.1.Micro capacity in your OCI region, a VCN with a public subnet, and an SSH keypair (default public key: ~/.ssh/id_ed25519.pub, and its private key must be your default SSH identity).
The OCI serial console proxy requires an RSA key (Ed25519 is rejected), by default ~/.ssh/oci-console-rsa. If you don't have one:
ssh-keygen -t rsa -b 4096 -f ~/.ssh/oci-console-rsa -N ''Working from a checkout instead: pip install -e '.[oci]'.
| Command | What it does |
|---|---|
alpine-fleet doctor [--fix] |
Check host requirements; print the install command for anything missing. |
alpine-fleet discover |
Show E2.1.Micro and A1.Flex headroom per availability domain, discovered dynamically. |
alpine-fleet up [--yes] [--debug] [--no-launch] |
Launch an instance if none is running (retries until capacity appears), then convert, harden and verify. Prints phase timings and a RAM comparison, and ends with the ssh root@<ip> command. |
alpine-fleet convert USER@IP INSTANCE_OCID |
Convert an existing instance. You must type its address to confirm (or pass --yes). |
alpine-fleet down [--yes] |
Terminate the instance recorded in the state file. |
alpine-fleet cache [--force] |
Build the Oracle Linux 7 kexec binary cache in a container (skips installing kexec-tools on the target). |
up and convert also accept --alpine-version (default v3.24), --ssh-pubkey, --console-key and --answerfile.
Every run detects where the box is (stock OS, Alpine in RAM, or installed) and resumes from there, so an interrupted run can simply be repeated.
Where things live. Instance state and serial transcripts (instance IDs and IPs) are in ~/.local/state/alpine-fleet/; the kexec cache is in ~/.cache/alpine-fleet/kexec/. Nothing is written to the install directory.
Environment overrides.
| Variable | Effect |
|---|---|
SUBNET_ID, IMAGE_ID, COMPARTMENT_ID |
Override auto-discovery (default: first subnet found, newest compatible Oracle Linux image, tenancy root compartment). |
DISPLAY_NAME, SSH_KEY_FILE |
Instance name and the public key given to the stock image's opc user. |
OCI_BIN, OCI_CONFIG_FILE |
Use a specific oci binary or config file. |
ALPINE_FLEET_STATE_DIR, KEXEC_CACHE |
Relocate state and cache. |
NO_KEXEC_CACHE=1 |
Ignore the cache and install kexec-tools on the target. |
This project is intentionally destructive. The answerfile installs to /dev/sda and replaces the existing operating system. Do not run it against a VM that holds data you need. The orchestrator waits for the installer's erase prompt and confirms it, but it does not check which disk the prompt names.
The kexec transition is also disruptive: SSH to the stock OS disappears the moment the new kernel takes over. That is expected. The orchestrator treats the lost connection as a state transition and continues through the serial console.
- Launch: Provisions an E2.1.Micro in an availability domain with headroom and opens a serial console connection.
- Stage: Connects over SSH. Copies the cached
kexec(or installskexec-toolson the target if there is no cache), downloads the Alpine netboot kernel and initramfs in parallel, and loads them withkexec -l. - Jump: Triggers
kexec -e. Alpine boots into RAM and downloads its ~290 MB module image from the Alpine mirror; that download's speed varies from run to run. - Install: Over the serial console: logs in, pushes your public key, starts
sshd, pushes the answerfile (chunked base64) to/root/answers, and runssetup-alpine -e -f /root/answers. The root password prompt is bypassed with-e. - Reboot: Reboots off the physical block volume onto the installed disk.
- Harden and verify: Root's password field is set to
*,PermitRootLogin prohibit-passwordandPasswordAuthentication noare written tosshd_config, and a fresh key-only login is tested (retried whilesshdsettles). If that login is refused, the hardening changes (sshd_configand root's shadow entry) are rolled back through a session held open across the change, andsshd's log is printed.
BusyBox passwd -l writes a !-prefixed hash. Alpine's sshd is built without PAM, and without PAM, OpenSSH treats a ! hash as a locked account and refuses public-key logins too. Setting the field to * blocks password logins but still allows SSH keys. Use *.
- Provider support: OCI only, tested on VM.Standard.E2.1.Micro only. The bundled answerfile assumes
/dev/sdaand performs a destructive disk install. Other clouds, images and shapes need their own disk and device handling. - The free tier allows two E2.1.Micro instances;
upstops if two are already running. - The
kexeccache applies to the stock Oracle Linux 7 image. Any other image misses the cache and falls back to installingkexec-toolson the target, which is much slower. - Root ends key-only with no password. If your private key is lost, you must rebuild the instance.
- Boot time depends on the mirror; the orchestrator waits up to 10 minutes for the Alpine login prompt.
- macOS hosts need bash 4+ (see above) and have not been tested.
src/alpine_fleet/
cli.py, doctor.py, paths.py the `alpine-fleet` command
scripts/ orchestrate.py, bootstrap.sh, launch/teardown/discover/console helpers, lib/
answerfiles/ setup-alpine answerfile for the OCI micro
licenses/, THIRD_PARTY_NOTICES.txt
scripts/e2e-test.sh (a destructive full-cycle test) lives in the repo only and is not shipped in the wheel.
MIT. Third-party notices are in THIRD_PARTY_NOTICES.txt and licenses/. The kexec binary in the cache is built locally from Oracle's own package and is never distributed with this project.