There are two important requirements for the way the MIME database is stored:

This specification uses the XDG Base Directory Specification[BaseDir] to define the prefixes below which the database is stored. In the rest of this document, paths shown with the prefix <MIME> indicate the files should be loaded from the mime subdirectory of every directory in XDG_DATA_HOME:XDG_DATA_DIRS.

For example, when using the default paths, “Load all the <MIME>/text/html.xml files” means to load /usr/share/mime/text/html.xml, /usr/local/share/mime/text/html.xml, and ~/.local/share/mime/text/html.xml (if they exist).

Each application that wishes to contribute to the MIME database will install a single XML file, named after the application, into one of the three <MIME>/packages/ directories (depending on where the user requested the application be installed). After installing, uninstalling or modifying this file, the application MUST run the update-mime-database command, which is provided by the freedesktop.org shared database[SharedMIME].

update-mime-database is passed the mime directory containing the packages subdirectory which was modified as its only argument. It scans all the XML files in the packages subdirectory, combines the information in them, and creates a number of output files.

Where the information from these files is conflicting, information from directories lower in the list takes precedence. Any file named Override.xml takes precedence over all other files in the same packages directory. This can be used by tools which let the user edit the database to ensure that the user's changes take effect.

The files created by update-mime-database are:

The format of these generated files and the source files in packages are explained in the following sections. This step serves several purposes. First, it allows applications to quickly get the data they need without parsing all the source XML files (the base package alone is over 700K). Second, it allows the database to be used for other purposes (such as creating the /etc/mime.types file if desired). Third, it allows validation to be performed on the input data, and removes the need for other applications to carefully check the input for errors themselves.

Each application provides only a single XML source file, which is installed in the packages directory as described above. This file is an XML file whose document element is named mime-info and whose namespace URI is http://www.freedesktop.org/standards/shared-mime-info. All elements described in this specification MUST have this namespace too.

The document element may contain zero or more mime-type child nodes, in any order, each describing a single MIME type. Each element has a type attribute giving the MIME type that it describes.

Each mime-type node may contain any combination of the following elements, and in any order:

Applications may also define their own elements, provided they are namespaced to prevent collisions. Unknown elements are copied directly to the output XML files like comment elements. A typical use for this would be to indicate the default handler application for a particular desktop ("Galeon is the GNOME default text/html browser"). Note that this doesn't indicate the user's preferred application, only the (fixed) default.

Here is an example source file, named diff.xml:

<?xml version="1.0"?>
<mime-info xmlns='http://www.freedesktop.org/standards/shared-mime-info'>
  <mime-type type="text/x-diff">
    <comment>Differences between files</comment>
    <comment xml:lang="af">verskille tussen lêers</comment>
    ...
    <magic priority="50">
      <match type="string" offset="0" value="diff\t"/>
      <match type="string" offset="0" value="***\t"/>
      <match type="string" offset="0" value="Common subdirectories: "/>
    </magic>
    <glob pattern="*.diff"/>
    <glob pattern="*.patch"/>
  </mime-type>
</mime-info>

In practice, common types such as text/x-diff are provided by the freedesktop.org shared database. Also, only new information needs to be provided, since this information will be merged with other information about the same type.

These files have a mime-type element as the root node. The format is as described above. They are created by merging all the mime-type elements from the source files and creating one output file per MIME type. Each file may contain information from multiple source files. The magic, glob and root-XML elements will have been removed.

The example source file given above would (on its own) create an output file called <MIME>/text/x-diff.xml containing the following:

<?xml version="1.0" encoding="utf-8"?>
<mime-type xmlns="http://www.freedesktop.org/standards/shared-mime-info" type="text/x-diff">
<!--Created automatically by update-mime-database. DO NOT EDIT!-->
  <comment>Differences between files</comment>
  <comment xml:lang="af">verskille tussen lêers</comment>
  ...
</mime-type>

This is a simple list of lines containing a MIME type and pattern, separated by a colon. For example:

# This file was automatically generated by the
# update-mime-database command. DO NOT EDIT!
...
text/x-diff:*.diff
text/x-diff:*.patch
...

Applications MUST first try a case-sensitive match, then try again with the filename converted to lower-case if that fails. This is so that main.C will be seen as a C++ file, but IMAGE.GIF will still use the *.gif pattern.

If several patterns match then the longest pattern SHOULD be used. In particular, files with multiple extensions (such as Data.tar.gz) MUST match the longest sequence of extensions (eg '*.tar.gz' in preference to '*.gz'). Literal patterns (eg, 'Makefile') must be matched before all others. It is suggested that patterns beginning with `*.' and containing no other special characters (`*?[') should be placed in a hash table for efficient lookup, since this covers the majority of the patterns. Thus, patterns of this form should be matched before other wildcarded patterns.

There may be several rules mapping to the same type. They should all be merged. If the same pattern is defined twice, then they MUST be ordered by the directory the rule came from, as described above.

Lines beginning with `#' are comments and should be ignored. Everything from the `:' character to the newline is part of the pattern; spaces should not be stripped. The file is in the UTF-8 encoding. The format of the glob pattern is as for fnmatch(3). The format does not allow a pattern to contain a literal newline character, but this is not expected to be a problem.

Common types (such as MS Word Documents) will be provided in the X Desktop Group's package, which MUST be required by all applications using this specification. Since each application will then only be providing information about its own types, conflicts should be rare.

The magic data is stored in a binary format for ease of parsing. The old magic database had complex escaping rules; these are now handled by update-mime-database.

The file starts with the magic string "MIME-Magic\0\n". There is no version number in the file. Incompatible changes will be handled by creating both the current `magic' file and a newer `magic2' in the new format. Where possible, compatible changes only will be made. All numbers are big-endian, so need to be byte-swapped on little-endian machines.

The rest of the file is made up of a sequence of small sections. Each section is introduced by giving the priority and type in brackets, followed by a newline character. Higher priority entries come first. Example:

[50:text/x-diff]\n

Each line in the section takes the form:

[ indent ] ">" start-offset "=" value
[ "&" mask ] [ "~" word-size ] [ "+" range-length ] "\n"

Note that the value, value length and mask are all binary, whereas everything else is textual. Each of the elements begins with a single character to identify it, except for the indent level.

The word size is used for byte-swapping. Little-endian systems should reverse the order of groups of bytes in the value and mask if this is greater than one. This only affects `host' matches (`big32' entries still have a word size of 1, for example, because no swapping is necessary, whereas `host32' has a word size of 4).

The indent, range-length, word-size and mask components are optional. If missing, indent defaults to 0, range-length to 1, the word-size to 1, and the mask to all 'one' bits.

Indent corresponds to the nesting depth of the rule. Top-level rules have an indent of zero. The parent of an entry is the preceding entry with an indent one less than the entry.

If an unknown character is found where a newline is expected then the whole line should be ignored (there will be no binary data after the new character, so the next line starts after the next "\n" character). This is for future extensions.

The text/x-diff above example would (on its own) create this magic file:

00000000  4d 49 4d 45 2d 4d 61 67  69 63 00 0a 5b 35 30 3a  |MIME-Magic..[50:|
00000010  74 65 78 74 2f 78 2d 64  69 66 66 5d 0a 3e 30 3d  |text/x-diff].>0=|
00000020  00 05 64 69 66 66 09 0a  3e 30 3d 00 04 2a 2a 2a  |..diff..>0=..***|
00000030  09 0a 3e 30 3d 00 17 43  6f 6d 6d 6f 6e 20 73 75  |..>0=..Common su|
00000040  62 64 69 72 65 63 74 6f  72 69 65 73 3a 20 0a     |bdirectories: .|

Each XMLnamespaces file is a list of lines in the form:

namespaceURI " " localName " " MIME-Type "\n"

For example:

http://www.w3.org/1999/xhtml html application/xhtml+xml

The lines are sorted (using strcmp) and there are no lines with the same namespaceURI and localName in one file. If the localName was empty then there will be two spaces following the namespaceURI.

An implementation MAY also get a file's MIME type from the user.mime_type extended attribute. The type given here should normally be used in preference to any guessed type, since the user is able to set it explicitly. Applications MAY choose to set the type when saving files. Since many applications and filesystems do not support extended attributes, implementations MUST NOT rely on this method being available.

A type is a subclass of another type if any instance of the first type is also an instance of the second. For example, all image/svg files are also text/xml, text/plain and application/octet-stream files. Subclassing is about the format, rather than the catagory of the data (for example, there is no 'generic spreadsheet' class that all spreadsheets inherit from).

Some subclass rules are implicit:

In addition to these rules, explicit subclass information may be given using the sub-class-of element.

Note that some file formats are also compressed files (application/x-jar files are also application/zip files). However, this is different to a case such as a compressed postscript file, which is not a valid postscript file itself (so application/x-gzpostscript does not inherit from application/postscript, because an application that can handle the latter may not cope with the former).

Some types may or may not be instances of other types. For example, a spreadsheet file may be compressed or not. It is a valid spreadsheet file either way, but only inherits from application/x-gzip in one case. This information cannot be represented statically; instead an application interested in this information should run all of the magic rules, and use the list of types returned as the subclasses.

Because different applications have different requirements, they may choose to use the various methods provided by this specification in any order. However, the RECOMMENDED order to perform the checks is:

There are several reasons for checking most of the glob patterns before the magic. Some applications don't check the magic at all, and this makes it more likely that both will get the same type. Users can easily understand why calling their text file README.mp3 makes the system think it's an MP3, whereas they have trouble understanding why their computer thinks README.txt is a PostScript file. If the system guesses wrongly, the user can often rename the file to fix the problem.

Sometimes it is useful to assign MIME types to other objects in the filesystem, such as directories, sockets and device files. This could be useful when looking up an icon for a type, or for providing a textual description of one of these objects. The media type 'inode' is provided for this purpose, with the following types corresponding to the standard types of object found in a Unix filesystem:

An inode/mount-point is a subclass of inode/directory. It can be useful when adding extra actions for these directories, such as 'mount' or 'eject'. Mounted directories can be detected by comparing the 'st_dev' of a directory with that of its parent. If they differ, they are from different devices and the directory is a mount point.

The system described in this document is intended to allow different programs to see the same file as having the same type. This is to help interoperability. The type determined in this way is only a guess, and an application MUST NOT trust a file based simply on its MIME type. For example, a downloader should not pass a file directly to a launcher application without confirmation simply because the type looks `harmless' (eg, text/plain).

Do not rely on two applications getting the same type for the same file, even if they both use this system. The spec allows some leeway in implementation, and in any case the programs may be following different versions of the spec.

The MIME database is NOT intended to store user preferences. Users should never edit the database. If they wish to make corrections or provide MIME entries for software that doesn't provide these itself, they should do so by means of the Override.xml mentioned in the section called “Directory layout”. Information such as "text/html files need to be opened with Mozilla" should NOT go in the database.