docs/sandboxes/bootstrap.mdx
Before a sandbox starts doing real work, you can choose its root storage and prepare its environment. Flat root disks trade layer-level deduplication for a direct ext4 root. Scripts bundle reusable commands. Patches modify the layered rootfs before the VM boots. A custom init system (systemd, OpenRC, s6) can run as PID 1 instead of microsandbox's minimal agent.
OCI sandboxes use stitched EROFS lower layers with a writable ext4 upper and guest OverlayFS by default. A flat root disk is an explicit alternative: microsandbox merges the OCI layers into one deterministic ext4 base, caches it by content, creates a private clone for each sandbox, grows that clone to the requested capacity, and mounts it directly as the guest root. The immutable cached base is never attached writable.
<CodeGroup> ```rust Rust use microsandbox::sandbox::FlatClone; use microsandbox::Sandbox;let sb = Sandbox::builder("worker") .image("python:3.12") .root_disk_with(|disk| disk.flat().size(8192).clone_strategy(FlatClone::Auto)) .create() .await?;
```typescript TypeScript
import { Sandbox } from "microsandbox";
await using sb = await Sandbox.builder("worker")
.image("python:3.12")
.rootDisk((disk) => disk.flat().size(8192).cloneStrategy("auto"))
.create();
from microsandbox import FlatClone, Image, RootDisk, Sandbox
sb = await Sandbox.create(
"worker",
image=Image.oci(
"python:3.12",
root_disk=RootDisk.flat(8192, clone=FlatClone.AUTO),
),
)
sb, err := m.CreateSandbox(ctx, "worker",
m.WithImage("python:3.12"),
m.WithRootDisk(m.RootDisk.Flat(m.RootDiskFlatOptions{
SizeMiB: 8192,
Clone: m.FlatCloneAuto,
})),
)
auto first asks the host filesystem for a native copy-on-write clone and falls back to an independent sparse copy when cloning is unsupported. Use copy when independence from the base's physical extents matters more than provisioning latency, or reflink when clone support is an operational requirement and failure is preferable to fallback. Linux uses FICLONE, macOS uses clonefile, and Windows uses block cloning only on volumes that advertise block refcounting.
The normal layered root remains the default and retains layer-level deduplication. Flat v1 is local-only, ext4-only, and incompatible with rootfs patches and snapshots; unsupported combinations fail explicitly.
See Optimization for workload tradeoffs, host filesystem expectations, and how to verify the clone strategy actually used.
Scripts are files mounted at /.msb/scripts/ inside the sandbox. The directory is on PATH, so each script is callable by name through exec() or shell().
It provides a clean way to bundle setup procedures or entry points with a sandbox without baking them into the image.
<CodeGroup> ```rust Rust use indoc::indoc; use microsandbox::Sandbox;let sb = Sandbox::builder("worker") .image("ubuntu") .script("setup", indoc! {" #!/bin/bash apt-get update && apt-get install -y python3 curl "}) .script("start", indoc! {" #!/bin/bash exec python3 /app/main.py "}) .create() .await?;
sb.shell("setup").await?; let output = sb.shell("start").await?;
```typescript TypeScript
import { Sandbox } from "microsandbox";
await using sb = await Sandbox.builder("worker")
.image("ubuntu")
.script("setup", "#!/bin/bash\napt-get update && apt-get install -y python3 curl")
.script("run", "#!/bin/bash\nexec python3 /app/main.py")
.create();
await sb.shell("setup");
const output = await sb.shell("run");
from microsandbox import Sandbox
sb = await Sandbox.create(
"worker",
image="ubuntu",
scripts={
"setup": "#!/bin/bash\napt-get update && apt-get install -y python3 curl",
"start": "#!/bin/bash\nexec python3 /app/main.py",
},
)
await sb.shell("setup")
output = await sb.shell("start")
sb, err := m.CreateSandbox(ctx, "worker",
m.WithImage("ubuntu"),
m.WithScripts(map[string]string{
"setup": "#!/bin/bash\napt-get update && apt-get install -y python3 curl",
"start": "#!/bin/bash\nexec python3 /app/main.py",
}),
)
_, err = sb.Shell(ctx, "setup")
output, err := sb.Shell(ctx, "start")
Patches modify the rootfs before the VM boots. Write config files, copy directories from the host, create symlinks, append to existing files, remove things you don't need. The base image stays untouched since patches are written to the writable layer on top.
Patches are applied in order and work with OCI images and bind-mounted rootfs. They're not supported with disk image roots (QCOW2, Raw).
<Tip> By default, patching a path that already exists in the image will error. Pass `replace: true` on the operation to allow it. `Mkdir` and `Remove` are idempotent and won't error either way. </Tip> <CodeGroup> ```rust Rust use microsandbox::Sandbox;let sb = Sandbox::builder("worker") .image("alpine") .patch(|p| p .text("/etc/greeting.txt", "Hello from a patched rootfs!\n", None, false) .text("/etc/motd", "Custom message of the day.\n", None, true) // replace existing .mkdir("/app", Some(0o755)) .text("/app/config.json", r#"{"debug": true}"#, Some(0o644), false) .copy_file("./cert.pem", "/etc/ssl/cert.pem", None, false) .append("/etc/hosts", "127.0.0.1 myapp.local\n") ) .create() .await?;
```typescript TypeScript
import { Sandbox } from "microsandbox";
await using sb = await Sandbox.builder("worker")
.image("alpine")
.patch((p) => p
.text("/etc/greeting.txt", "Hello from a patched rootfs!\n")
.text("/etc/motd", "Custom message of the day.\n", { replace: true })
.mkdir("/app", { mode: 0o755 })
.text("/app/config.json", '{"debug": true}', { mode: 0o644 })
.copyFile("./cert.pem", "/etc/ssl/cert.pem")
.append("/etc/hosts", "127.0.0.1 myapp.local\n"),
)
.create();
from microsandbox import Patch, Sandbox
sb = await Sandbox.create(
"worker",
image="alpine",
patches=[
Patch.text("/etc/greeting.txt", "Hello from a patched rootfs!\n"),
Patch.text("/etc/motd", "Custom message of the day.\n", replace=True),
Patch.mkdir("/app", mode=0o755),
Patch.text("/app/config.json", '{"debug": true}', mode=0o644),
Patch.copy_file("./cert.pem", "/etc/ssl/cert.pem"),
Patch.append("/etc/hosts", "127.0.0.1 myapp.local\n"),
],
)
mode755 := uint32(0o755)
mode644 := uint32(0o644)
sb, err := m.CreateSandbox(ctx, "worker",
m.WithImage("alpine"),
m.WithPatches(
m.Patch.Text("/etc/greeting.txt", "Hello from a patched rootfs!\n", m.PatchOptions{}),
m.Patch.Text("/etc/motd", "Custom message of the day.\n", m.PatchOptions{Replace: true}),
m.Patch.Mkdir("/app", m.PatchOptions{Mode: &mode755}),
m.Patch.Text("/app/config.json", `{"debug": true}`, m.PatchOptions{Mode: &mode644}),
m.Patch.CopyFile("./cert.pem", "/etc/ssl/cert.pem", m.PatchOptions{}),
m.Patch.Append("/etc/hosts", "127.0.0.1 myapp.local\n"),
),
)
The patch builder appends operations in the order you call them; calls are chainable. Available operations across SDKs: text, file, mkdir, append, copyFile / copy_file, copyDir / copy_dir, symlink, remove.
For per-language signatures and option shapes, see the SDK references:
PatchBuilderPatchBuilderPatch (factory) and PatchConfig (the value type)Patch (helpers) and PatchConfig (the value type)Transparent huge pages (THP) let the guest kernel back eligible anonymous memory with larger pages. This can reduce address-translation overhead for large memory working sets, especially inside a virtual machine, but always may allocate memory in 2 MiB units for sparsely touched mappings. Microsandbox therefore defaults to the density-conscious madvise policy and lets each sandbox choose its policy at creation time.
The selected value is applied as the Linux transparent_hugepage= kernel boot parameter before agentd or the user workload starts. It is persisted with the sandbox and takes effect again on every start. It cannot be changed with msb modify; create or replace the sandbox to select another policy.
let sb = Sandbox::builder("memory-worker") .image("python:3.12") .thp(TransparentHugePagePolicy::Always) .create() .await?;
```typescript TypeScript
const sb = await Sandbox.builder("memory-worker")
.image("python:3.12")
.thp("always")
.create();
sb = await Sandbox.create(
"memory-worker",
image="python:3.12",
thp="always",
)
sb, err := m.CreateSandbox(ctx, "memory-worker",
m.WithImage("python:3.12"),
m.WithTHP(m.THPAlways),
)
msb create python:3.12 --name memory-worker --thp always
Use always when the workload has large, densely touched anonymous allocations and memory bandwidth or TLB pressure matters. Keep madvise for general-purpose and high-density sandboxes. Use never for diagnostics or workloads that must avoid transparent huge pages entirely.
By default the microsandbox agent runs as PID 1 inside the guest: small, fast, minimal. For workloads that expect a real init (systemd, OpenRC, s6, runit, etc.), --init hands PID 1 over to the init binary of your choice.
Common reasons to opt in: long-lived daemons, system service tests, anything that talks to dbus or expects systemctl to work.
Use auto to pick a known init binary from the image, or pass an absolute path when you need to pin the entry point for reproducible CI. When an OCI image declares a known init path as the first ENTRYPOINT token, such as /init in s6-overlay images, auto hands PID 1 to that path. For attached msb run, microsandbox preserves the OCI launch contract by passing the remaining ENTRYPOINT plus CMD or trailing command to that init instead of direct-executing the wrapper through agentd. Otherwise, auto falls back to probing common distro paths: /sbin/init, /lib/systemd/systemd, and /usr/lib/systemd/systemd.
let sb = Sandbox::builder("worker") .image("ghcr.io/superradcompany/debian-systemd:12") .memory(1024) .cpus(2) .init("auto") .create() .await?;
```typescript TypeScript
import { MiB, Sandbox } from "microsandbox";
await using sb = await Sandbox.builder("worker")
.image("ghcr.io/superradcompany/debian-systemd:12")
.memory(MiB(1024))
.cpus(2)
.init("auto")
.create();
from microsandbox import Sandbox
sb = await Sandbox.create(
"worker",
image="ghcr.io/superradcompany/debian-systemd:12",
memory=1024,
cpus=2,
init="auto",
)
sb, err := m.CreateSandbox(ctx, "worker",
m.WithImage("ghcr.io/superradcompany/debian-systemd:12"),
m.WithMemory(1024),
m.WithCPUs(2),
m.WithInit(m.Init.Auto()),
)
msb run ghcr.io/superradcompany/debian-systemd:12 \
-m 1G -c 2 \
--init auto \
-- bash
To verify the handoff worked, check /proc/1/comm inside the sandbox:
$ msb run ghcr.io/superradcompany/debian-systemd:12 --init auto -- cat /proc/1/comm
systemd
If --init=auto cannot find an init binary from the image ENTRYPOINT or the guest-side probe list, boot fails with a clear error in kernel.log. Use an explicit path when you know where the init lives.
Pass extra argv and env to the init:
init_with(...) / initWith(...).InitConfig to init=.--init-arg once per entry, and --init-env KEY=VAL once per env var.argv[0] is the init command; extra argv entries are appended after it. Env is merged on top of the inherited environment.
await using sb = await Sandbox.builder("worker")
.image("ghcr.io/superradcompany/debian-systemd:12")
.initWith("/lib/systemd/systemd", (i) => i
.args(["--unit=multi-user.target"])
.env("container", "microsandbox"))
.create();
from microsandbox import InitConfig, Sandbox
sb = await Sandbox.create(
"worker",
image="ghcr.io/superradcompany/debian-systemd:12",
init=InitConfig(
cmd="/lib/systemd/systemd",
args=("--unit=multi-user.target",),
env={"container": "microsandbox"},
),
)
sb, err := m.CreateSandbox(ctx, "worker",
m.WithImage("ghcr.io/superradcompany/debian-systemd:12"),
m.WithInit(m.Init.Cmd("/lib/systemd/systemd",
m.InitOptions{
Args: []string{"--unit=multi-user.target"},
Env: map[string]string{"container": "microsandbox"},
},
)),
)
msb run ghcr.io/superradcompany/debian-systemd:12 \
--init /lib/systemd/systemd \
--init-arg --unit=multi-user.target \
--init-env container=microsandbox \
-- bash
Most slim Docker base images do not include systemd or another full init. Use an image that ships the init you want, or build a small custom image that installs it.
--init controls PID 1. --entrypoint and the trailing command normally control the workload you run after boot, so the two can be combined. An image-declared init entrypoint such as /init is the special case. With --init auto and attached msb run, microsandbox passes the trailing command to PID 1 as part of the image's own launch contract.