kern

rootless, daemonless OCI container and resource-slicing runtime

TLDR

Start an interactive Alpine shell

$ kern box dev --image alpine -it -- sh

Cap a host process (no sandbox)

$ kern run --memory 256M --cpus 0.5 -- [./crunch]

Run a service with port publish and restart

$ kern box svc --image nginx:alpine -d -p 8080:80 --restart -- nginx -g 'daemon off;'

List running boxes

$ kern ps

Exec into a box

$ kern exec [svc] -it -- sh

Live TUI of CPU, memory, and volumes

$ kern top

Bring up a Compose file (kern or Docker format)

$ kern compose [stack.toml] up

Check whether this kernel can run boxes

$ kern doctor

kern is a static ~1.5 MB Linux runtime: real OCI images, always-rootless namespaces (user, pid, mount, net, uts, ipc), seccomp, and cgroup v2, with no daemon. Cold start of a box from an image is a few milliseconds. Resource profiles in `~/.config/kern/kern.toml` (`vcpu:`, `vdisk:`, `vgpio:`) attach by name to either a box or kern run.`--security-profile untrusted` is the hardened bundle (seccomp allowlist, `--cap-drop ALL`, `--read-only`). It is not a hypervisor: a kernel LPE is an escape, same class as Docker/Podman.Python/Node kern-sandbox and kern-mcp wrap the same binary for agent-generated code.

COMMANDS

box name [options] -- cmd

Create/run an isolated box from an OCI image (`--image`, `-it`, `-d`, `-p`, `-v`, `--restart`, `--security-profile untrusted`, resource tokens like `vcpu:heavy`).

run [limits] -- cmd

Apply CPU/memory (and optional Landlock) to a process without a sandbox.

ps / logs / exec / stop / inspect / stats / wait

Lifecycle and inspection. Most list verbs also take --json.

top

Live TUI.

compose file up|down

Run `kern-compose.toml` or a `docker-compose.yml` as one pod.

doctor / validate

Host capability check and config validation.

volume, images, pull, build, commit, push, save, load

OCI and volume operations (see project docs for the full matrix).

CAVEATS

Linux only (WSL2 is fine; no native Windows). Needs unprivileged user namespaces and cgroup v2. Image pull wants curl and tar on PATH (and often pasta / uidmap for non-root images). Bind-mounting `$HOME` is a trust decision, not a boundary. Not a CRI runtime and not a Docker Engine API clone.

HISTORY

Apache-2.0 project by Alex / getkern. First published release v0.7.0. Install: `curl -fsSL https://raw.githubusercontent.com/getkern/kern/main/install.sh | sh` or `cargo install --git https://github.com/getkern/kern getkern --locked`.

RESOURCES

Source code · Documentation