DOKK / manpages / debian 12 / mkosi / mkosi.1.en
mkosi(1) mkosi(1)

mkosi — Build Bespoke OS Images

mkosi [options...] build

mkosi [options...] clean

mkosi [options...] summary

mkosi [options...] shell [command line...]

mkosi [options...] boot [nspawn settings...]

mkosi [options...] qemu

mkosi [options...] ssh

mkosi [options...] serve

mkosi [options...] bump

mkosi [options...] genkey

mkosi [options...] help

mkosi is a tool for easily building customized OS images. It’s a fancy wrapper around dnf --installroot, debootstrap, pacstrap and zypper that may generate disk images with a number of bells and whistles.

The following command line verbs are known:

This builds the image, based on the settings passed in on the command line or read from a mkosi.default file. This verb is the default if no verb is explicitly specified. This command must be executed as root. Any arguments passed after build are passed as arguments to the build script (if there is one).
Remove build artifacts generated on a previous build. If combined with -f, also removes incremental build cache images. If -f is specified twice, also removes any package cache.
Outputs a human-readable summary of all options used for building an image. This will parse the command line and mkosi.default file as it would do on build, but only output what it is configured for and not actually build anything.`
This builds the image if it is not built yet, and then invokes systemd-nspawn to acquire an interactive shell prompt in it. If this verb is used an optional command line may be specified which is then invoked in place of the shell in the container. Combine this with -f in order to rebuild the image unconditionally before acquiring the shell, see below. This command must be executed as root.
Similar to shell but boots the image up using systemd-nspawn. If this verb is used an optional command line may be specified which is passed as “kernel command line” to the init system in the image.
Similar to boot but uses qemu to boot up the image, i.e. instead of container virtualization VM virtualization is used. This verb is only supported on images that contain a boot loader, i.e. those built with Bootable=yes (see below). This command must be executed as root unless the image already exists and -f is not specified.
When the image is built with the Ssh=yes option, this command connects to a booted (boot, qemu verbs) container/VM via SSH. Make sure to run mkosi ssh with the same config as mkosi build was run with so that it has the necessary information available to connect to the running container/VM via SSH.
This builds the image if it is not built yet, and then serves the output directory (i.e. usually mkosi.output/, see below) via a small embedded HTTP server, listening on port 8081. Combine with -f in order to rebuild the image unconditionally before serving it. This command is useful for testing network based acquisition of OS images, for example via machinectl pull-raw ... and machinectl pull-tar ....
Determines the current image version string (as configured with ImageVersion=/--image-version=), increases its last dot-separated component by one and writes the resulting version string to mkosi.version. This is useful for implementing a simple versioning scheme: each time this verb is called the version is bumped in preparation for the subsequent build. Note that --auto-bump/-B may be used to automatically bump the version after each successful build.
Generate a pair of SecureBoot keys for usage with the SecureBootKey=/--secure-boot-key= and SecureBootCertificate=/--secure-boot-certificate= options.
This verb is equivalent to the --help switch documented below: it shows a brief usage explanation.

Execution flow for mkosi build. Columns represent the execution context. Default values/calls are shown in parentheses. When building with --incremental mkosi creates a cache of the distribution installation for both images if not already existing and replaces the distribution installation in consecutive runs with data from the cached one.


HOST . BUILD . FINAL
. IMAGE . IMAGE
. .
start . .
| . .
v . . build script? -------exists-----> copy .
| . skeleton trees .
| . (mkosi.skeleton/) .
none . | .
| . v .
v . install .
skip . distribution, .
build image . packages and .
| . build packages, .
| . run .
| . prepare script .
| . (mkosi.prepare build) .
| . or if incremental .
| . use cached build image .
| . | .
| . v .
| . copy .
| . build sources .
| . (./) .
| . | .
| . v .
| . copy .
| . extra trees .
| . (mkosi.extra/) .
| . | .
| . v .
| . run .
| . postinstall script .
| . (mkosi.postinst build) .
| . | .
| .-------------------------' .
| | . .
| v . .
| run . .
| finalize script . .
|(mkosi.finalize build). .
| | . .
| '-------------------------. .
| . | .
| . v .
| . run .
| . build script .
| . (mkosi.build) .
| . | .
'-----------------------------------+------------------------.
. . |
. . v
. . copy
. . skeleton trees
. . (mkosi.skeleton/)
. . |
. . v
. . install
. . distribution
. . and packages,
. . run
. . prepare script
. . (mkosi.prepare final)
. . or if incremental
. . use cached final image
. . |
. . v
. . copy
. . build results
. . |
. . v
. . copy
. . extra trees
. . (mkosi.extra/)
. . |
. . v
. . run
. . postinstall script
. . (mkosi.postinst final)
. . |
. . v
. . |
. . perform cleanup
. . (remove files, packages,
. . package metadata)
. . |
.--------------------------------------------------'
| . .
v . .
run . .
finalize script . .
(mkosi.finalize final) . .
| . .
.---------' . .
| . .
v . .
end . .
. .
HOST . BUILD . FINAL
. IMAGE . IMAGE
. .

The following output formats are supported:

Raw GPT disk image, with ext4 as root (gpt_ext4)
Raw GPT disk image, with xfs as root (gpt_xfs)
Raw GPT disk image, with btrfs as root (gpt_btrfs)
Raw GPT disk image, with squashfs as read-only root (gpt_squashfs)
Plain squashfs image, without partition table, as read-only root (plain_squashfs)
Plain directory, containing the OS tree (directory)
btrfs subvolume, with separate subvolumes for /var, /home, /srv, /var/tmp (subvolume)
Tar archive (tar)
CPIO archive (cpio) in the format appropriate for a kernel initrd

When a GPT disk image is created, the following additional options are available:

A swap partition may be added in
The image may be made bootable on EFI and BIOS systems
Separate partitions for /srv and /home may be added in
The root, /srv and /home partitions may optionally be encrypted with LUKS.
A dm-verity partition may be added in that adds runtime integrity data for the root partition

The following settings can be set through configuration files (the syntax with SomeSetting=value) and on the command line (the syntax with --some-setting=value). For some command line parameters, a single-letter shortcut is also allowed. In the configuration files, the setting must be in the appropriate section, so the settings are grouped by section below.

Command line options that take no argument are shown without “=” in their long version. In the config files, they should be specified with a boolean argument: either “1”, “yes”, or “true” to enable, or “0”, “no”, “false” to disable.

The distribution to install in the image. Takes one of the following arguments: fedora, debian, ubuntu, arch, opensuse, mageia, centos, centos_epel, clear, photon, openmandriva, rocky, rocky_epel, alma, alma_epel. If not specified, defaults to the distribution of the host.
The release of the distribution to install in the image. The precise syntax of the argument this takes depends on the distribution used, and is either a numeric string (in case of Fedora Linux, CentOS, ..., e.g. 29), or a distribution version name (in case of Debian, Ubuntu, ..., e.g. artful). If neither this option, nor Distribution= is specified, defaults to the distribution version of the host. If the distribution is specified, defaults to a recent version of it.
The mirror to use for downloading the distribution packages. Expects a mirror URL as argument.
Additional package repositories to use during installation. Expects one or more URLs as argument, separated by commas. This option may be used multiple times, in which case the list of repositories to use is combined. Use “!*” to remove all repositories from to the list or use e.g. “!repo-url” to remove just one specific repository. For Arch Linux, additional repositories must be passed in the form <name>::<url> (e.g. myrepo::https://myrepo.net).
This option is only applicable for RPM-based distributions: CentOS, Fedora Linux, Mageia, Photon, Rocky Linux, Alma Linux and OpenMandriva. Allows use of the host’s existing RPM repositories. By default, a hardcoded set of default RPM repositories is generated and used. Use --repositories= to identify a custom set of repositories to be enabled and used for the build.
This option can (for now) only be used with RPM-based istributions and Arch Linux. It identifies a directory containing extra repository definitions that will be used when installing packages. The files are passed directly to the corresponding package manager and should be written in the format expected by the package manager of the image’s distro.
The architecture to build the image for. Note that this currently only works for architectures compatible with the host’s architecture.

The image format type to generate. One of directory (for generating OS images inside a local directory), subvolume (similar, but as a btrfs subvolume), tar (similar, but a tarball of the image is generated), cpio (similar, but a cpio archive is generated), gpt_ext4 (a block device image with an ext4 file system inside a GPT partition table), gpt_xfs (similar, but with an xfs file system), gpt_btrfs (similar, but with an btrfs file system), gpt_squashfs (similar, but with a squashfs file system), plain_squashfs (a plain squashfs file system without a partition table).
The manifest format type or types to generate. A comma-delimited list consisting of json (the standard JSON output format that describes the packages installed), changelog (a human-readable text format designed for diffing). Defaults to json.
Path for the output image file to generate. Takes a relative or absolute path where the generated image will be placed. If neither this option nor OutputDirectory= is used, the image is generated under the name image, but its name suffixed with an appropriate file suffix (e.g. image.raw.xz in case gpt_ext4 is used in combination with xz compression). If the ImageId= option is configured it is used instead of image in the default output name. If an image version is specified via ImageVersion=, it is included in the default name, e.g. a specified image version of 7.8 might result in an image file name of image_7.8.raw.xz.
Paths for the split-out output image files, when SplitArtifacts=yes is used. If unspecified, the relevant split artifact files will be named like the main image, but with .root, .verity, and .efi suffixes inserted (and in turn possibly suffixed by compression suffix, if compression is enabled).
Path to a directory where to place all generated artifacts (i.e. the generated image when an output path is not given, SHA256SUMS file, etc.). If this is not specified and the directory mkosi.output/ exists in the local directory, it is automatically used for this purpose. If the setting is not used and mkosi.output/ does not exist, all output artifacts are placed adjacent to the output image file.
Path to a directory where to store data required temporarily while building the image. This directory should have enough space to store the full OS image, though in most modes the actually used disk space is smaller. If not specified, and mkosi.workspace/ exists in the local directory, it is used for this purpose. Otherwise, a subdirectory in the temporary storage area is used ($TMPDIR if set, /var/tmp/ otherwise). The data in this directory is removed automatically after each build. It’s safe to manually remove the contents of this directory should an mkosi invocation be aborted abnormally (for example, due to reboot/power failure). If the btrfs output modes are selected this directory must be backed by btrfs too.
Replace the output file if it already exists, when building an image. By default when building an image and an output artifact already exists mkosi will refuse operation. Specify this option once to delete all build artifacts from a previous run before re-building the image. If incremental builds are enabled, specifying this option twice will ensure the intermediary cache files are removed, too, before the re-build is initiated. If a package cache is used (also see the “Files” section below), specifying this option thrice will ensure the package cache is removed too, before the re-build is initiated. For the clean operation this option has a slightly different effect: by default the verb will only remove build artifacts from a previous run, when specified once the incremental cache files are deleted too, and when specified twice the package cache is also removed.
Override the first usable LBA (Logical Block Address) within the GPT header. This defaults to 2048, which is actually the desired value. However, some tools, e.g. the prl_disk_tool utility from the Parallels virtualization suite require this to be set to 34, otherwise they might fail to resize the disk image and/or partitions inside it.
Generate a bootable image. By default this will generate an image bootable on UEFI systems. Use BootProtocols= to select support for a different boot protocol.
Pick one or more boot protocols to support when generating a bootable image, as enabled with Bootable=. Takes a comma-separated list of uefi or bios. May be specified more than once in which case the specified lists are merged. If uefi is specified the sd-boot UEFI boot loader is used, if bios is specified the GNU Grub boot loader is used. Use “!*” to remove all previously added protocols or “!protocol” to remove one protocol.
Use the specified kernel command line when building bootable images. By default command line arguments get appended. To remove all arguments from the current list pass “!*”. To remove specific arguments add a space separated list of “!” prefixed arguments. For example adding “!* console=ttyS0 rw” to a mkosi.default file or the command line arguments passes “console=ttyS0 rw” to the kernel in any case. Just adding “console=ttyS0 rw” would append these two arguments to the kernel command line created by lower priority configuration files or previous KernelCommandLine= command line arguments.
Sign the resulting kernel/initrd image for UEFI SecureBoot.
Path to the PEM file containing the secret key for signing the UEFI kernel image, if SecureBoot= is used.
Path to the X.509 file containing the certificate for the signed UEFI kernel image, if SecureBoot= is used.
Common name to be used when generating SecureBoot keys via mkosi’s genkey command. Defaults to mkosi of %u, where %u expands to the username of the user invoking mkosi.
Number of days that the keys should remain valid when generating SecureBoot keys via mkosi’s genkey command. Defaults to two years (730 days).
Set the read-only flag on the root partition in the partition table. Only applies to gpt_ext4, gpt_xfs, gpt_btrfs, subvolume output formats, and is implied on gpt_squashfs and plain_squashfs. The read-only flag is essentially a hint to tools using the image (see https://systemd.io/DISCOVERABLE_PARTITIONS/). In particular, all systemd tools like systemd-nspawn and systemd-gpt-auto-generator will mount such partitions read-only, but tools from other project may ignore the flag.
Attempt to make the resulting root file system as small as possible by removing free space from the file system. Only supported for gpt_ext4 and gpt_btrfs. For ext4 this relies on resize2fs -M, which reduces the free disk space but is not perfect and generally leaves some free space. For btrfs the results are optimal and no free space is left.
Encrypt all partitions in the file system or just the root file system. Takes either all or data as argument. If all, the root, /home and /srv file systems will be encrypted using dm-crypt/LUKS (with its default settings). If data, the root file system will be left unencrypted, but /home and /srv will be encrypted. The passphrase to use is read from the mkosi.passphrase file in the current working directory. Note that the UEFI System Partition (ESP) containing the boot loader and kernel to boot is never encrypted since it needs to be accessible by the firmware.
Add a “Verity” integrity partition to the image. Takes a boolean or the special value signed, and defaults to disabled. If enabled, the root partition (or /usr/ partition, in case UsrOnly= is enabled) is protected with dm-verity against offline modification, the verification data is placed in an additional GPT partition. Implies ReadOnly=yes. If this is enabled, the Verity root hash is written to an output file with .roothash or .usrhash suffix. If set to signed, Verity is also enabled, but the resulting root hash is then also signed (in PKCS#7 format) with the signature key configured with SecureBootKey=. Or in other words: the SecureBoot key pair is then used to both sign the kernel, if that is enabled, and the root//usr/ file system. This signature is then stored in an additional output file with the .roothash.p7s or .usrhash.p7s suffix in DER format. It is also written to an additional partition in the image. The latter allows generating self-contained signed disk images, implementing the Verity provisions described in the Discoverable Partitions Specification (https://systemd.io/DISCOVERABLE_PARTITIONS).
Enable or disable internal compression in the file system. Only applies to output formats with squashfs or btrfs. Takes one of zlib, lzo, zstd, lz4, xz or a boolean value as argument. If the latter is used compression is enabled/disabled and the default algorithm is used. In case of the squashfs output formats compression is implied, but this option may be used to select the algorithm.
Configure compression for the resulting image or archive. The argument can be either a boolean or a compression algorithm (xz, zstd). xz compression is used by default. Note that when applied to block device image types this means the image cannot be started directly but needs to be decompressed first. This also means that the shell, boot, qemu verbs are not available when this option is used. Implied for tar and cpio.
Enable compression. Using this option is equivalent to either CompressFs= or CompressOutput=; the appropriate type of compression is selected automatically.
Set the path to the mksquashfs executable to use. This is useful in case the parameters for the tool shall be augmented, as the tool may be replaced by a script invoking it with the right parameters, this way.
Encode the resulting image as QEMU QCOW2 image. This only applies to gpt_ext4, gpt_xfs, gpt_btrfs, gpt_squashfs. QCOW2 images can be read natively by qemu, but not by the Linux kernel. This means the shell and boot verbs are not available when this option is used, however qemu will work.
Set the image’s hostname to the specified name.
Configure the image version. This accepts any string, but it is recommended to specify a series of dot separated components. The version may also be configured in a file mkosi.version in which case it may be conveniently managed via the bump verb or the --auto-bump switch. When specified the image version is included in the default output file name, i.e. instead of image.raw the default will be image_0.1.raw for version 0.1 of the image, and similar. The version is also passed via the $IMAGE_VERSION to any build scripts invoked (which may be useful to patch it into /etc/os-release or similar, in particular the IMAGE_VERSION= field of it).
Configure the image identifier. This accepts a freeform string that shall be used to identify the image with. If set the default output file will be named after it (possibly suffixed with the version). If this option is used the root, /usr/ and Verity partitions in the image will have their labels set to this (possibly suffixed by the image version). The identifier is also passed via the $IMAGE_ID to any build scripts invoked (which may be useful to patch it into /etc/os-release or similar, in particular the IMAGE_ID= field of it).
If specified, mkosi does not build unified kernel images and instead installs kernels with a separate initrd and boot loader config to the efi or bootloader partition.
If specified, mkosi will run the tool to create the initrd such that a non-generic initrd is created that will only be able to run on the system mkosi is run on. Currently mkosi uses dracut for all supported distributions except Clear Linux and this option translates to enabling dracut’s hostonly option.
If specified, mkosi will only add the /usr/ directory tree (instead of the whole root file system) to the image. This is useful for fully stateless systems that come up pristine on every single boot, where /etc/ and /var/ are populated by systemd-tmpfiles/systemd-sysusers and related calls, or systems that are originally shipped without a root file system, but where systemd-repart adds one on the first boot.
If specified and building an image with a partition table, also write out the root file system partition, its Verity partition (if configured) and the generated unified kernel (if configured) into separate output files. This is useful in A/B update scenarios where an existing disk image shall be augmented with a new version of a root or /usr partition along with its Verity partition and unified kernel.
By default, if mkosi is run inside a sudo environment all generated artifacts have their UNIX user/group ownership changed to the user which invoked sudo. With this option this may be turned off and all generated files are owned by root.
If running on a SELinux-enabled system (Fedora Linux, CentOS, Rocky Linux, Alma Linux), files inside the container are tagged with SELinux context extended attributes (xattrs), which may interfere with host SELinux rules in building or further container import stages. This option strips SELinux context attributes from the resulting tar archive.
Set the machine’s ID to the specified value. If unused, a random ID will be used while building the image and the final image will be shipped without a machine ID.

Takes a boolean or the special value conditional. If true, automatically install packages to ensure basic functionality, as appropriate for the given image type. For example, systemd is always included, systemd-udev and dracut if the image is bootable, and so on. If false, only packages specified with Packages= will be installed. If conditional, the list of packages to install will be extended with boolean dependencies (c.f. https://rpm.org/user_doc/boolean_dependencies.html), to install specific packages when other packages are in the list. For example, systemd-udev may be automatically included if the image is bootable and systemd is installed. With this, various “base” packages still need to be specified if they should be included, but the corresponding “extension” packages will be added automatically when appropriate. This feature depends on support in the package manager, so it is not implemented for all distributions.
Install the specified distribution packages (i.e. RPM, DEB, ...) in the image. Takes a comma separated list of package specifications. This option may be used multiple times in which case the specified package lists are combined. Packages specified this way will be installed both in the development and the final image. Use BuildPackages= to specify packages that shall only be used for the image generated in the build image, but that shall not appear in the final image. The types and syntax of “package specifications” that are allowed depend on the package installer (e.g. dnf or yum for rpm-based distros or apt for deb-based distros), but may include package names, package names with version and/or architecture, package name globs, paths to packages in the file system, package groups, and virtual provides, including file paths. To remove a package e.g. added by a mkosi.default configuration file prepend the package name with !. For example -p “!apache2” would remove the apache2 package. To replace the apache2 package by the httpd package just add -p “!apache2,httpd” to the command line arguments. To remove all packages use “!*”. Example: when using an distro that uses dnf, Packages=meson libfdisk-devel.i686 git-* prebuilt/rpms/systemd-249-rc1.local.rpm /usr/bin/ld @development-tools python3dist(mypy) would install the meson package (in the latest version), the 32-bit version of the libfdisk-devel package, all available packages that start with the git- prefix, a systemd rpm from the local file system, one of the packages that provides /usr/bin/ld, the packages in the “Development Tools” group, and the package that contains the mypy python module.
Include documentation in the image built. By default if the underlying distribution package manager supports it documentation is not included in the image built. The $WITH_DOCS environment variable passed to the mkosi.build script indicates whether this option was used or not.
If set to false (or when the command-line option is used), the $WITH_TESTS environment variable is set to 0 when the mkosi.build script is invoked. This is supposed to be used by the build script to bypass any unit or integration tests that are normally run during the source build process. Note that this option has no effect unless the mkosi.build build script honors it.
Takes a path to a directory to use as package cache for the distribution package manager used. If this option is not used, but a mkosi.cache/ directory is found in the local directory it is automatically used for this purpose. The directory configured this way is mounted into both the development and the final image while the package manager is running.
Takes a path to a directory to copy into the OS tree before invoking the package manager. Use this to insert files and directories into the OS tree before the package manager installs any packages. If this option is not used, but the mkosi.skeleton/ directory is found in the local directory it is automatically used for this purpose (also see the “Files” section below). Instead of a directory, a tar file may be provided. In this case it is unpacked into the OS tree before the package manager is invoked. This mode of operation allows setting permissions and file ownership explicitly, in particular for projects stored in a version control system such as git which retain full file ownership and access mode metadata for committed files. If the tar file mkosi.skeleton.tar is found in the local directory it will be automatically used for this purpose.
Takes a path to a directory to copy on top of the OS tree the package manager generated. Use this to override any default configuration files shipped with the distribution. If this option is not used, but the mkosi.extra/ directory is found in the local directory it is automatically used for this purpose (also see the “Files” section below). As with the skeleton tree logic above, instead of a directory, a tar file may be provided too. mkosi.skeleton.tar will be automatically used if found in the local directory.
Enable/disable removal of package manager databases, caches, and logs at the end of installation. Can be specified as true, false, or “auto” (the default). With “auto”, files will be removed if the respective package manager executable is not present at the end of the installation.
Takes a comma-separated list of globs. Files in the image matching the globs will be purged at the end.
Takes a comma-separated list of package specifications for removal, in the same format as Packages=. The removal will be performed as one of the last steps. This step is skipped if CleanPackageMetadata=no is used. This option is currently only implemented for distributions using dnf.
Adds variables to the environment that the build/prepare/postinstall/finalize scripts are executed with. Takes a space-separated list of variable assignments or just variable names. In the latter case, the values of those variables will be passed through from the environment in which mkosi was invoked. This option may be specified more than once, in which case all listed variables will be set. If the same variable is set twice, the later setting overrides the earlier one.
Takes a path to a source tree to copy into the development image, if the build script is used. This only applies if a build script is used, and defaults to the local directory. Use SourceFileTransfer= to configure how the files are transferred from the host to the container image.
Takes a path of a directory to use as build directory for build systems that support out-of-tree builds (such as Meson). The directory used this way is shared between repeated builds, and allows the build system to reuse artifacts (such as object files, executable, ...) generated on previous invocations. This directory is mounted into the development image when the build script is invoked. The build script can find the path to this directory in the $BUILDDIR environment variable. If this option is not specified, but a directory mkosi.builddir/ exists in the local directory it is automatically used for this purpose (also see the “Files” section below).
Takes a path of a directory to use as the include directory. This directory is mounted at /usr/include when building the build image and running the build script. This means all include files installed to /usr/include will be stored in this directory. This is useful to make include files available on the host system for use by language servers to provide code completion. If this option is not specified, but a directory mkosi.includedir/ exists in the local directory, it is automatically used for this purpose (also see the “Files” section below).
Takes a path of a directory to use as the install directory. The directory used this way is shared between builds and allows the build system to not have to reinstall files that were already installed by a previous build and didn’t change. The build script can find the path to this directory in the $DESTDIR environment variable. If this option is not specified, but a directory mkosi.installdir exists in the local directory, it is automatically used for this purpose (also see the “Files” section below).
Similar to Packages=, but configures packages to install only in the first phase of the build, into the development image. This option should be used to list packages containing header files, compilers, build systems, linkers and other build tools the mkosi.build script requires to operate. Note that packages listed here are only included in the image created during the first phase of the build, and are absent in the final image. Use Packages= to list packages that shall be included in both. Packages are appended to the list. Packages prefixed with “!” are removed from the list. “!*” removes all packages from the list.
Set the password of the root user. By default the root account is locked. If this option is not used, but a file mkosi.rootpw exists in the local directory, the root password is automatically read from it.
Indicate that the password supplied for the root user has already been hashed, so that the string supplied with Password= or mkosi.rootpw will be written to /etc/shadow literally.
Enable autologin for the root user on /dev/pts/0 (nspawn), /dev/tty1 (QEMU) and /dev/ttyS0 (QEMU with QemuHeadless=yes) by patching /etc/pam.d/login.
Causes the (second) final image build stage to be skipped. This is useful in combination with a build script, for when you care about the artifacts that were created locally in $BUILDDIR, but ultimately plan to discard the final image.
Takes a path to an executable that is used as build script for this image. If this option is used the build process will be two-phased instead of single-phased. The specified script is copied onto the development image and executed inside an systemd-nspawn container environment. If this option is not used, but the mkosi.build file found in the local directory it is automatically used for this purpose (also see the “Files” section below). Specify an empty value to disable automatic detection.
Takes a path to an executable that is invoked inside the image right after installing the software packages. It is the last step before the image is cached (if incremental mode is enabled). This script is invoked inside a systemd-nspawn container environment, and thus does not have access to host resources. If this option is not used, but an executable script mkosi.prepare is found in the local directory, it is automatically used for this purpose. Specify an empty value to disable automatic detection.
Takes a path to an executable that is invoked inside the final image right after copying in the build artifacts generated in the first phase of the build. This script is invoked inside a systemd-nspawn container environment, and thus does not have access to host resources. If this option is not used, but an executable mkosi.postinst is found in the local directory, it is automatically used for this purpose. Specify an empty value to disable automatic detection.
Takes a path to an executable that is invoked outside the final image right after copying in the build artifacts generated in the first phase of the build, and after having executed the mkosi.postinst script (see PostInstallationScript=). This script is invoked directly in the host environment, and hence has full access to the host’s resources. If this option is not used, but an executable mkosi.finalize is found in the local directory, it is automatically used for this purpose. Specify an empty value to disable automatic detection.
Configures how the source file tree (as configured with BuildSources=) is transferred into the container image during the first phase of the build. Takes one of copy-all (to copy all files from the source tree), copy-git-cached (to copy only those files git ls-files --cached lists), copy-git-others (to copy only those files git ls-files --others lists), mount to bind mount the source tree directly. Defaults to copy-git-cached if a git source tree is detected, otherwise copy-all. When you specify copy-git-more, it is the same as copy-git-cached, except it also includes the .git/ directory.
Same as SourceFileTransfer=, but for the final image instead of the build image. Takes the same values as SourceFileFransfer= except mount. By default, sources are not copied into the final image.
If given, any symbolic links in the source file tree are resolved and the file contents are copied to the build image. If not given, they are left as symbolic links. This only applies if SourceFileTransfer= is copy-all. Defaults to leaving them as symbolic links.
Same as SourceResolveSymlinks=, but for the final image instead of the build image.
When true, enables network connectivity while the build script mkosi.build is invoked. By default, the build script runs with networking turned off. The $WITH_NETWORK environment variable is passed to the mkosi.build build script indicating whether the build is done with or without network. If specified as never, the package manager is instructed not to contact the network for updating package data. This provides a minimal level of reproducibility, as long as the package data cache is already fully populated.
Specifies a .nspawn settings file for systemd-nspawn to use in the boot and shell verbs, and to place next to the generated image file. This is useful to configure the systemd-nspawn environment when the image is run. If this setting is not used but an mkosi.nspawn file found in the local directory it is automatically used for this purpose.

Use the specified directory or file system image as the base image, and create the output image that consists only of changes from this base. The base image is attached as the lower file system in an overlayfs structure, and the output filesystem becomes the upper layer, initially empty. Thus files that are not modified compared to the base image are not present in the output image. This option may be used to create systemd “system extensions” or portable services. See https://systemd.io/PORTABLE_SERVICES/#extension-images for more information.
Takes a size in bytes for the root file system. The specified numeric value may be suffixed with K, M, G to indicate kilo-, mega- and gigabytes (all to the base of 1024). This applies to output formats gpt_ext4, gpt_xfs, gpt_btrfs. Defaults to 3G.
Similar to RootSize=, configures the size of the UEFI System Partition (ESP). This is only relevant if the Bootable= option is used to generate a bootable image. Defaults to 256 MB.
Similar to RootSize=, configures the size of a swap partition on the image. If omitted, no swap partition is created.
Similar to RootSize=, configures the size of the /home partition. If omitted, no separate /home partition is created.
Similar to RootSize=, configures the size of the /srv partition. If omitted, no separate /srv partition is created.

Generate a SHA256SUMS file of all generated artifacts after the build is complete.
Sign the generated SHA256SUMS using gpg after completion.
Select the gpg key to use for signing SHA256SUMS. This key must be already present in the gpg keyring.
Generate a bmap file for usage with bmaptool from the generated image file.

List of colon-separated paths to look for tools in, before using the regular $PATH search path.
When used with the build verb, this option adds console=ttyS0 to the image’s kernel command line and sets the terminal type of the serial console in the image to the terminal type of the host (more specifically, the value of the $TERM environment variable passed to mkosi). This makes sure that all terminal features such as colors and shortcuts still work as expected when connecting to the qemu VM over the serial console (for example via -nographic). When used with the qemu verb, this option adds the -nographic option to qemu’s command line so qemu starts a headless vm and connects to its serial console from the current terminal instead of launching the VM in a separate window.
When used with the qemu verb, this options sets qemu’s -smp argument which controls the number of guest’s CPUs. Defaults to 2.
When used with the qemu verb, this options sets qemu’s -m argument which controls the amount of guest’s RAM. Defaults to 1G.
When used with the qemu verb, this option specifies whether QEMU should use KVM acceleration. Defaults to yes if the host machine supports KVM acceleration, no otherwise.
When used, this option instructs underlying calls of systemd-nspawn to use the current unit scope, instead of creating a dedicated transcient scope unit for the containers. This option should be used when mkosi is run by a service unit.
When used with the boot or qemu verbs, this option creates a virtual ethernet link between the host and the container/VM. The host interface is automatically picked up by systemd-networkd as documented in systemd-nspawn’s man page: https://www.freedesktop.org/software/systemd/man/systemd-nspawn.html#-n
When used with the shell, boot, or qemu verbs, this option runs the specified verb on a temporary snapshot of the output image that is removed immediately when the container terminates. Taking the temporary snapshot is more efficient on file systems that support subvolume snapshots or `reflinks' natively (“btrfs” or new “xfs”) than on more traditional file systems that do not (“ext4”).
If specified, installs and enables sshd in the final image and generates a SSH keypair and adds the public key to root’s authorized_keys in the final image. The private key is stored in mkosi’s output directory. When building with this option and running the image using mkosi boot or mkosi qemu, the mkosi ssh command can be used to connect to the container/VM via SSH.
If specified, use the given private key when connecting to the guest machine via mkosi ssh. This requires the public key counterpart to be present in the same location, suffixed with .pub (as done by ssh-keygen). If this option is not present, mkosi generates a new key pair automatically.
If specified as a path, use the given socket to connect to the ssh agent when building an image and when connecting via mkosi ssh instead of hard-coding a key. If specified as true, $SSH_AUTH_SOCK will be parsed instead (hint: use sudo with -E). The keys listed by ssh-add -L will be installed as authorized keys in the built image. The ssh invocation done by mkosi ssh will inherit $SSH_AUTH_SOCK for authentication purposes.
In the image, sshd will be configured to listen on this port. mkosi ssh will connect to this port.
When used with the ssh verb, mkosi will attempt to retry the SSH connection up to given timeout (in seconds) in case it fails. This option is useful mainly in scripted environments where the qemu and ssh verbs are used in a quick succession and the virtual device might not get enough time to configure itself.

Those settings cannot be configured in the configuration files.

Takes a path to a directory. mkosi switches to this directory before doing anything. Note that the various mkosi.* files are searched for only after changing to this directory, hence using this option is an effective way to build a project located in a specific directory.
Loads additional settings from the specified settings file. Most command line options may also be configured in a settings file. See the table below to see which command line options match which settings file option. If this option is not used, but a file mkosi.default is found in the local directory it is automatically used for this purpose. If a setting is configured both on the command line and in the settings file, the command line generally wins, except for options taking lists in which case both lists are combined.
Iterate through all files mkosi.* in the mkosi.files/ subdirectory, and build each as if --default=mkosi.files/mkosi.... was invoked. This is a quick way to build a large number of images in one go. Any additional specified command line arguments override the relevant options in all files processed this way.
If specified, overrides the directory the --all logic described above looks for settings files in. If unspecified, defaults to mkosi.files/ in the current working directory.
Enable incremental build mode. This only applies if the two-phase mkosi.build build script logic is used. In this mode, a copy of the OS image is created immediately after all OS packages are unpacked but before the mkosi.build script is invoked in the development container. Similarly, a copy of the final image is created immediately before the build artifacts from the mkosi.build script are copied in. On subsequent invocations of mkosi with the -i switch these cached images may be used to skip the OS package unpacking, thus drastically speeding up repetitive build times. Note that when this is used and a pair of cached incremental images exists they are not automatically regenerated, even if options such as Packages= are modified. In order to force rebuilding of these cached images, combine -i with -ff to ensure cached images are first removed and then re-created.
Enable additional debugging output. Takes a comma-separated list of arguments specifying the area of interest. Pass any invalid value (e.g. empty) to list currently accepted values.
Show package version.
Show brief usage information.
If specified, after each successful build the the version is bumped in a fashion equivalent to the bump verb, in preparation for the next build. This is useful for simple, linear version management: each build in a series will have a version number one higher then the previous one.

Images may be created containing installations of the following operating systems:

Fedora Linux
Debian
Ubuntu
Arch Linux
openSUSE
Mageia
CentOS
Clear Linux
Photon
OpenMandriva
Rocky Linux
Alma Linux
Gentoo

In theory, any distribution may be used on the host for building images containing any other distribution, as long as the necessary tools are available. Specifically, any distribution that packages debootstrap may be used to build Debian or Ubuntu images. Any distribution that packages dnf may be used to build Fedora Linux, Mageia or OpenMandriva images. Any distro that packages pacstrap may be used to build Arch Linux images. Any distribution that packages zypper may be used to build openSUSE images. Any distribution that packages yum (or the newer replacement dnf) may be used to build CentOS, Rocky Linux, or Alma Linux images. Any distribution that packages emerge may be used to build Gentoo images.

Currently, Fedora Linux packages all relevant tools as of Fedora 28.

Legacy concepts are avoided: generated images use GPT disk labels (and no MBR labels), and only systemd-based images may be generated.

All generated GPT disk images may be booted in a local container directly with:

systemd-nspawn -bi image.raw
    

Additionally, bootable GPT disk images (as created with the --bootable flag) work when booted directly by EFI and BIOS systems, for example in KVM via:

qemu-kvm -m 512 -smp 2 -bios /usr/share/edk2/ovmf/OVMF_CODE.fd -drive format=raw,file=image.raw
    

EFI bootable GPT images are larger than plain GPT images, as they additionally carry an EFI system partition containing a boot loader, as well as a kernel, kernel modules, udev and more.

All directory or btrfs subvolume images may be booted directly with:

systemd-nspawn -bD image
    

To make it easy to build images for development versions of your projects, mkosi can read configuration data from the local directory, under the assumption that it is invoked from a source tree. Specifically, the following files are used if they exist in the local directory:

The mkosi.default file provides the default configuration for the image building process. For example, it may specify the distribution to use (fedora, ubuntu, debian, arch, opensuse, mageia, openmandriva, gentoo) for the image, or additional distribution packages to install. Note that all options encoded in this configuration file may also be set on the command line, and this file is hence little more than a way to make sure invoking mkosi without further parameters in your source tree is enough to get the right image of your choice set up.

Additionally, if a mkosi.default.d/ directory exists, each file in it is loaded in the same manner adding/overriding the values specified in mkosi.default. If mkosi.default.d/ contains a directory named after the distribution being built, each file in that directory is also processed.

The file format is inspired by Windows .ini files and supports multi-line assignments: any line with initial whitespace is considered a continuation line of the line before. Command-line arguments, as shown in the help description, have to be included in a configuration block (e.g. “[Content]”) corresponding to the argument group (e.g. “Content”), and the argument gets converted as follows: “--with-network” becomes “WithNetwork=yes”. For further details see the table above.

The mkosi.skeleton/ directory or mkosi.skeleton.tar archive may be used to insert files into the image. The files are copied before the distribution packages are installed into the image. This allows creation of files that need to be provided early, for example to configure the package manager or set systemd presets.

When using the directory, file ownership is not preserved: all files copied will be owned by root. To preserve ownership, use a tar archive.

The mkosi.extra/ directory or mkosi.extra.tar archive may be used to insert additional files into the image, on top of what the distribution includes in its packages. They are similar to mkosi.skeleton/ and mkosi.skeleton.tar, but the files are copied into the directory tree of the image after the OS was installed.

When using the directory, file ownership is not preserved: all files copied will be owned by root. To preserve ownership, use a tar archive.

mkosi.build may be an executable script. If it exists, the image will be built twice: the first iteration will be the development image, the second iteration will be the final image. The development image is used to build the project in the current working directory (the source tree). For that the whole directory is copied into the image, along with the mkosi.build script. The script is then invoked inside the image (via systemd-nspawn), with $SRCDIR pointing to the source tree. $DESTDIR points to a directory where the script should place any files generated it would like to end up in the final image. Note that make/automake/meson based build systems generally honor $DESTDIR, thus making it very natural to build source trees from the build script. After the development image was built and the build script ran inside of it, it is removed again. After that the final image is built, without any source tree or build script copied in. However, this time the contents of $DESTDIR are added into the image.

When the source tree is copied into the build image, all files are copied, except for mkosi.builddir/, mkosi.cache/ and mkosi.output/. That said, .gitignore is respected if the source tree is a git checkout. If multiple different images shall be built from the same source tree it is essential to exclude their output files from this copy operation, as otherwise a version of an image built earlier might be included in a later build, which is usually not intended. An alternative to excluding these built images via .gitignore entries is to use the mkosi.output/ directory, which is an easy way to exclude all build artifacts.

The $MKOSI_DEFAULT environment variable will be set inside of this script so that you know which mkosi.default (if any) was passed in.

The mkosi.prepare script is invoked directly after the software packages are installed, from within the image context, if it exists. It is once called for the development image (if this is enabled, see above) with the “build” command line parameter, right before copying the extra tree. It is called a second time for the final image with the “final” command line parameter. This script has network access and may be used to install packages from other sources than the distro’s package manager (e.g. pip, npm, ...), after all software packages are installed but before the image is cached (if incremental mode is enabled). This script is executed within $SRCDIR. In contrast to a general purpose installation, it is safe to install packages to the system (pip install, npm install -g) instead of in $SRCDIR itself because the build image is only used for a single project and can easily be thrown away and rebuilt so there’s no risk of conflicting dependencies and no risk of polluting the host system.
The mkosi.postinst script is invoked as the penultimate step of preparing an image, from within the image context, if it exists. It is called first for the development image (if this is enabled, see above) with the “build” command line parameter, right before invoking the build script. It is called a second time for the final image with the “final” command line parameter, right before the image is considered complete. This script may be used to alter the images without any restrictions, after all software packages and built sources have been installed. Note that this script is executed directly in the image context with the final root directory in place, without any $SRCDIR/$DESTDIR setup.
The mkosi.finalize script, if it exists, is invoked as last step of preparing an image, from the host system. It is once called for the development image (if this is enabled, see above) with the “build” command line parameter, as the last step before invoking the build script, after the mkosi.postinst script is invoked. It is called the second time with the “final” command line parameter as the last step before the image is considered complete. The environment variable $BUILDROOT points to the root directory of the installation image. Additional verbs may be added in the future, the script should be prepared for that. This script may be used to alter the images without any restrictions, after all software packages and built sources have been installed. This script is more flexible than mkosi.postinst in two regards: it has access to the host file system so it’s easier to copy in additional files or to modify the image based on external configuration, and the script is run in the host, so it can be used even without emulation even if the image has a foreign architecture.
The mkosi.mksquashfs-tool script, if it exists, will be called wherever mksquashfs would be called.
The mkosi.nspawn nspawn settings file will be copied into the same place as the output image file, if it exists. This is useful since nspawn looks for settings files next to image files it boots, for additional container runtime settings.
The mkosi.cache/ directory, if it exists, is automatically used as package download cache, in order to speed repeated runs of the tool.
The mkosi.builddir/ directory, if it exists, is automatically used as out-of-tree build directory, if the build commands in the mkosi.build script support it. Specifically, this directory will be mounted into the build container, and the $BUILDDIR environment variable will be set to it when the build script is invoked. The build script may then use this directory as build directory, for automake-style or ninja-style out-of-tree builds. This speeds up builds considerably, in particular when mkosi is used in incremental mode (-i): not only the disk images, but also the build tree is reused between subsequent invocations. Note that if this directory does not exist the $BUILDDIR environment variable is not set, and it is up to build script to decide whether to do in in-tree or an out-of-tree build, and which build directory to use.
The mkosi.includedir/ directory, if it exists, is automatically used as an out-of-tree include directory for header files. Specifically, it will be mounted in the build container at /usr/include/ when building the build image and when running the build script. After building the (cached) build image, this directory will contain all the files installed to /usr/include. Language servers or other tools can use these files to provide a better editing experience for developers working on a project.
The mkosi.installdir/ directory, if it exists, is automatically used as the install directory. Specifically, this directory will be mounted into the container at /root/dest when running the build script. After running the build script, the contents of this directory are installed into the final image. This is useful to cache the install step of the build. If used, subsequent builds will only have to reinstall files that have changed since the previous build.
The mkosi.rootpw file can be used to provide the password or hashed password (if --password-is-hashed is set) for the root user of the image. The password may optionally be followed by a newline character which is implicitly removed. The file must have an access mode of 0600 or less. If this file does not exist, the distribution’s default root password is set (which usually means access to the root user is blocked).
The mkosi.passphrase file provides the passphrase to use when LUKS encryption is selected. It should contain the passphrase literally, and not end in a newline character (i.e. in the same format as cryptsetup and /etc/crypttab expect the passphrase files). The file must have an access mode of 0600 or less. If this file does not exist and encryption is requested, the user is queried instead.
The mkosi.secure-boot.crt and mkosi.secure-boot.key files contain an X.509 certificate and PEM private key to use when UEFI SecureBoot support is enabled. All EFI binaries included in the image’s ESP are signed with this key, as a late step in the build process.
The mkosi.output/ directory will be used for all build artifacts, if the image output path is not configured (i.e. no --output= setting specified), or configured to a filename (i.e. a path containing no / character). This includes the image itself, the root hash file in case Verity is used, the checksum and its signature if that’s enabled, and the nspawn settings file if there is any. Note that this directory is not used if the image output path contains at least one slash, and has no effect in that case. This setting is particularly useful if multiple different images shall be built from the same working directory, as otherwise the build result of a preceding run might be copied into a build image as part of the source tree (see above).
The mkosi.reposdir/ directory, if it exists, is automatically used as the repository directory for extra repository files. See the RepositoryDirectory option for more information.

All these files are optional.

Note that the location of all these files may also be configured during invocation via command line switches, and as settings in mkosi.default, in case the default settings are not acceptable for a project.

If no build script mkosi.build (see above) is used the build consists of a single phase only: the final image is generated as the combination of mkosi.skeleton/ (see above), the unpacked distribution packages and mkosi.extra/.

If a build script mkosi.build is used the build consists of two phases: in the the first development phase an image that includes necessary build tools (i.e. the combination of Packages= and BuildPackages= is installed) is generated (i.e. the combination of mkosi.skeleton/ and unpacked distribution packages). Into this image the source tree is copied and mkosi.build executed. The artifacts the mkosi.build generates are saved. Then, the second final phase starts: an image that excludes the build tools (i.e. only Packages= is installed, BuildPackages= is not) is generated. This time the build artifacts saved from the first phase are copied in, and mkosi.extra copied on top, thus generating the final image.

The two-phased approach ensures that source tree is executed in a clean and comprehensive environment, while at the same the final image remains minimal and contains only those packages necessary at runtime, but avoiding those necessary at build-time.

Note that only the package cache mkosi.cache/ is shared between the two phases. The distribution package manager is executed exactly once in each phase, always starting from a directory tree that is populated with mkosi.skeleton but nothing else.

mkosi supports three different caches for speeding up repetitive re-building of images. Specifically:

1.
The package cache of the distribution package manager may be cached between builds. This is configured with the --cache= option or the mkosi.cache/ directory. This form of caching relies on the distribution’s package manager, and caches distribution packages (RPM, DEB, ...) after they are downloaded, but before they are unpacked.
2.
If an mkosi.build script is used, by enabling incremental build mode with --incremental, a cached copy of the development and final images can be made immediately before the build sources are copied in (for the development image) or the artifacts generated by mkosi.build are copied in (in case of the final image). This form of caching allows bypassing the time-consuming package unpacking step of the distribution package managers, but is only effective if the list of packages to use remains stable, but the build sources and its scripts change regularly. Note that this cache requires manual flushing: whenever the package list is modified the cached images need to be explicitly removed before the next re-build, using the -f switch.
3.
Finally, between multiple builds the build artifact directory may be shared, using the mkosi.builddir/ directory. This directory allows build systems such as Meson to reuse already compiled sources from a previous built, thus speeding up the build process of the mkosi.build build script.

The package cache (i.e. the first item above) is unconditionally useful. The latter two caches only apply to uses of mkosi with a source tree and build script. When all three are enabled together turn-around times for complete image builds are minimal, as only changed source files need to be recompiled: an OS image rebuilt will be almost as quick to build the source tree only.

The build script mkosi.build receives the following environment variables:

$SRCDIR contains the path to the sources to build.
$DESTDIR is a directory into which any artifacts generated by the build script shall be placed.
$BUILDDIR is only defined if mkosi.builddir and points to the build directory to use. This is useful for all build systems that support out-of-tree builds to reuse already built artifacts from previous runs.
$WITH_DOCS is either 0 or 1 depending on whether a build without or with installed documentation was requested (WithDocs=yes). The build script should suppress installation of any package documentation to $DESTDIR in case $WITH_DOCS is set to 0.
$WITH_TESTS is either 0or 1 depending on whether a build without or with running the test suite was requested (WithTests=no). The build script should avoid running any unit or integration tests in case $WITH_TESTS is 0.
$WITH_NETWORK is either 0or 1 depending on whether a build without or with networking is being executed (WithNetwork=no). The build script should avoid any network communication in case $WITH_NETWORK is 0.

Create and run a raw GPT image with ext4, as image.raw:

# mkosi --bootable --incremental boot
    

Create and run a bootable btrfs GPT image, as foobar.raw:

# mkosi --format gpt_btrfs --bootable -o foobar.raw
# mkosi --output foobar.raw boot
# mkosi --output foobar.raw qemu
    

Create and run a Fedora Linux image into a plain directory:

# mkosi --distribution fedora --format directory boot
    

Create a compressed image image.raw.xz and add a checksum file, and install SSH into it:

# mkosi --distribution fedora --format gpt_squashfs --checksum --compress --package=openssh-clients
    

Inside the source directory of an automake-based project, configure mkosi so that simply invoking mkosi without any parameters builds an OS image containing a built version of the project in its current state:

# cat >mkosi.default <<EOF
[Distribution]
Distribution=fedora
Release=24
[Output]
Format=gpt_btrfs
Bootable=yes
[Content]
Packages=openssh-clients,httpd
BuildPackages=make,gcc,libcurl-devel
EOF
# cat >mkosi.build <<EOF
#!/bin/sh
cd $SRCDIR
./autogen.sh
./configure --prefix=/usr
make -j `nproc`
make install
EOF
# chmod +x mkosi.build
# mkosi --bootable --incremental boot
# systemd-nspawn -bi image.raw
    

To create a Fedora Linux image with hostname:

# mkosi --distribution fedora --hostname image
    

Also you could set hostname in configuration file:

# cat mkosi.default
...
[Output]
Hostname=image
...
    

mkosi is packaged for various distributions: Debian, Ubuntu, Arch Linux, Fedora Linux, OpenMandriva, Gentoo. It is usually easiest to use the distribution package.

The current version requires systemd 233 (or actually, systemd-nspawn of it).

When not using distribution packages make sure to install the necessary dependencies. For example, on Fedora Linux you need:

dnf install arch-install-scripts btrfs-progs debootstrap dosfstools edk2-ovmf e2fsprogs squashfs-tools gnupg python3 tar veritysetup xfsprogs xz zypper sbsigntools
    

On Debian/Ubuntu it might be necessary to install the ubuntu-keyring, ubuntu-archive-keyring and/or debian-archive-keyring packages explicitly, in addition to debootstrap, depending on what kind of distribution images you want to build. debootstrap on Debian only pulls in the Debian keyring on its own, and the version on Ubuntu only the one from Ubuntu.

Note that the minimum required Python version is 3.7.

Primary mkosi git repository on GitHub (https://github.com/systemd/mkosi/)
mkosi — A Tool for Generating OS Images (http://0pointer.net/blog/mkosi-a-tool-for-generating-os-images.html) introductory blog post by Lennart Poettering
The mkosi OS generation tool (https://lwn.net/Articles/726655/) story on LWN

systemd-nspawn(1), dnf(8), debootstrap(8)

The mkosi Authors.

2016-