Writing Extensions for Python-Markdown

Python-Markdown includes an API for extension writers to plug their own custom functionality and/or syntax into the parser. There are preprocessors which allow you to alter the source before it is passed to the parser, inline patterns which allow you to add, remove or override the syntax of any inline elements, and postprocessors which allow munging of the output of the parser before it is returned. If you really want to dive in, there are also blockprocessors which are part of the core BlockParser.

As the parser builds an ElementTree object which is later rendered as Unicode text, there are also some helpers provided to ease manipulation of the tree. Each part of the API is discussed in its respective section below. Additionally, reading the source of some Available Extensions may be helpful. For example, the Footnotes extension uses most of the features documented here.


Preprocessors munge the source text before it is passed into the Markdown core. This is an excellent place to clean up bad syntax, extract things the parser may otherwise choke on and perhaps even store it for later retrieval.

Preprocessors should inherit from markdown.preprocessors.Preprocessor and implement a run method with one argument lines. The run method of each Preprocessor will be passed the entire source text as a list of Unicode strings. Each string will contain one line of text. The run method should return either that list, or an altered list of Unicode strings.

A pseudo example:

from markdown.preprocessors import Preprocessor

class MyPreprocessor(Preprocessor):
    def run(self, lines):
        new_lines = []
        for line in lines:
            m = MYREGEX.match(line)
            if m:
                # do stuff
        return new_lines

Inline Patterns

Inline Patterns implement the inline HTML element syntax for Markdown such as *emphasis* or [links](http://example.com). Pattern objects should be instances of classes that inherit from markdown.inlinepatterns.Pattern or one of its children. Each pattern object uses a single regular expression and must have the following methods:

  • getCompiledRegExp():

    Returns a compiled regular expression.

  • handleMatch(m):

    Accepts a match object and returns an ElementTree element of a plain Unicode string.

Note that any regular expression returned by getCompiledRegExp must capture the whole block. Therefore, they should all start with r'^(.*?)' and end with r'(.*?)!'. When using the default getCompiledRegExp() method provided in the Pattern you can pass in a regular expression without that and getCompiledRegExp will wrap your expression for you and set the re.DOTALL and re.UNICODE flags. This means that the first group of your match will be m.group(2) as m.group(1) will match everything before the pattern.

For an example, consider this simplified emphasis pattern:

from markdown.inlinepatterns import Pattern
from markdown.util import etree

class EmphasisPattern(Pattern):
    def handleMatch(self, m):
        el = etree.Element('em')
        el.text = m.group(3)
        return el

As discussed in Integrating Your Code Into Markdown, an instance of this class will need to be provided to Markdown. That instance would be created like so:

# an oversimplified regex
MYPATTERN = r'\*([^*]+)\*'
# pass in pattern and create instance
emphasis = EmphasisPattern(MYPATTERN)

Actually it would not be necessary to create that pattern (and not just because a more sophisticated emphasis pattern already exists in Markdown). The fact is, that example pattern is not very DRY. A pattern for **strong** text would be almost identical, with the exception that it would create a ‘strong’ element. Therefore, Markdown provides a number of generic pattern classes that can provide some common functionality. For example, both emphasis and strong are implemented with separate instances of the SimpleTagPattern listed below. Feel free to use or extend any of the Pattern classes found at markdown.inlinepatterns.

Generic Pattern Classes

  • SimpleTextPattern(pattern):

    Returns simple text of group(2) of a pattern.

  • SimpleTagPattern(pattern, tag):

    Returns an element of type “tag” with a text attribute of group(3) of a pattern. tag should be a string of a HTML element (i.e.: ‘em’).

  • SubstituteTagPattern(pattern, tag):

    Returns an element of type “tag” with no children or text (i.e.: ‘br’).

There may be other Pattern classes in the Markdown source that you could extend or use as well. Read through the source and see if there is anything you can use. You might even get a few ideas for different approaches to your specific situation.


Treeprocessors manipulate an ElemenTree object after it has passed through the core BlockParser. This is where additional manipulation of the tree takes place. Additionally, the InlineProcessor is a Treeprocessor which steps through the tree and runs the InlinePatterns on the text of each Element in the tree.

A Treeprocessor should inherit from markdown.treeprocessors.Treeprocessor, over-ride the run method which takes one argument root (an Elementree object) and returns either that root element or a modified root element.

A pseudo example:

from markdown.treeprocessors import Treeprocessor

class MyTreeprocessor(Treeprocessor):
    def run(self, root):
        #do stuff
        return my_modified_root

For specifics on manipulating the ElementTree, see Working with the ElementTree below.


Postprocessors manipulate the document after the ElementTree has been serialized into a string. Postprocessors should be used to work with the text just before output.

A Postprocessor should inherit from markdown.postprocessors.Postprocessor and over-ride the run method which takes one argument text and returns a Unicode string.

Postprocessors are run after the ElementTree has been serialized back into Unicode text. For example, this may be an appropriate place to add a table of contents to a document:

from markdown.postprocessors import Postprocessor

class TocPostprocessor(Postprocessor):
    def run(self, text):
        return MYMARKERRE.sub(MyToc, text)


Sometimes, pre/tree/postprocessors and Inline Patterns aren’t going to do what you need. Perhaps you want a new type of block type that needs to be integrated into the core parsing. In such a situation, you can add/change/remove functionality of the core BlockParser. The BlockParser is composed of a number of Blockproccessors. The BlockParser steps through each block of text (split by blank lines) and passes each block to the appropriate Blockprocessor. That Blockprocessor parses the block and adds it to the ElementTree. The Definition Lists extension would be a good example of an extension that adds/modifies Blockprocessors.

A Blockprocessor should inherit from markdown.blockprocessors.BlockProcessor and implement both the test and run methods.

The test method is used by BlockParser to identify the type of block. Therefore the test method must return a boolean value. If the test returns True, then the BlockParser will call that Blockprocessor’s run method. If it returns False, the BlockParser will move on to the next BlockProcessor.

The test method takes two arguments:

  • parent: The parent etree Element of the block. This can be useful as the block may need to be treated differently if it is inside a list, for example.

  • block: A string of the current block of text. The test may be a simple string method (such as block.startswith(some_text)) or a complex regular expression.

The run method takes two arguments:

  • parent: A pointer to the parent etree Element of the block. The run method will most likely attach additional nodes to this parent. Note that nothing is returned by the method. The Elementree object is altered in place.

  • blocks: A list of all remaining blocks of the document. Your run method must remove (pop) the first block from the list (which it altered in place - not returned) and parse that block. You may find that a block of text legitimately contains multiple block types. Therefore, after processing the first type, your processor can insert the remaining text into the beginning of the blocks list for future parsing.

Please be aware that a single block can span multiple text blocks. For example, The official Markdown syntax rules state that a blank line does not end a Code Block. If the next block of text is also indented, then it is part of the previous block. Therefore, the BlockParser was specifically designed to address these types of situations. If you notice the CodeBlockProcessor, in the core, you will note that it checks the last child of the parent. If the last child is a code block (<pre><code>...</code></pre>), then it appends that block to the previous code block rather than creating a new code block.

Each BlockProcessor has the following utility methods available:

  • lastChild(parent):

    Returns the last child of the given etree Element or None if it had no children.

  • detab(text):

    Removes one level of indent (four spaces by default) from the front of each line of the given text string.

  • looseDetab(text, level):

    Removes “level” levels of indent (defaults to 1) from the front of each line of the given text string. However, this methods allows secondary lines to not be indented as does some parts of the Markdown syntax.

Each BlockProcessor also has a pointer to the containing BlockParser instance at self.parser, which can be used to check or alter the state of the parser. The BlockParser tracks it’s state in a stack at parser.state. The state stack is an instance of the State class.

State is a subclass of list and has the additional methods:

  • set(state):

    Set a new state to string state. The new state is appended to the end of the stack.

  • reset():

    Step back one step in the stack. The last state at the end is removed from the stack.

  • isstate(state):

    Test that the top (current) level of the stack is of the given string state.

Note that to ensure that the state stack doesn’t become corrupted, each time a state is set for a block, that state must be reset when the parser finishes parsing that block.

An instance of the BlockParser is found at Markdown.parser. BlockParser has the following methods:

  • parseDocument(lines):

    Given a list of lines, an ElementTree object is returned. This should be passed an entire document and is the only method the Markdown class calls directly.

  • parseChunk(parent, text):

    Parses a chunk of markdown text composed of multiple blocks and attaches those blocks to the parent Element. The parent is altered in place and nothing is returned. Extensions would most likely use this method for block parsing.

  • parseBlocks(parent, blocks):

    Parses a list of blocks of text and attaches those blocks to the parent Element. The parent is altered in place and nothing is returned. This method will generally only be used internally to recursively parse nested blocks of text.

While is is not recommended, an extension could subclass or completely replace the BlockParser. The new class would have to provide the same public API. However, be aware that other extensions may expect the core parser provided and will not work with such a drastically different parser.

Working with the ElementTree

As mentioned, the Markdown parser converts a source document to an ElementTree object before serializing that back to Unicode text. Markdown has provided some helpers to ease that manipulation within the context of the Markdown module.

First, to get access to the ElementTree module import ElementTree from markdown rather than importing it directly. This will ensure you are using the same version of ElementTree as markdown. The module is found at markdown.util.etree within Markdown.

from markdown.util import etree

markdown.util.etree tries to import ElementTree from any known location, first as a standard library module (from xml.etree in Python 2.5), then as a third party package (Elementree). In each instance, cElementTree is tried first, then ElementTree if the faster C implementation is not available on your system.

Sometimes you may want text inserted into an element to be parsed by InlinePatterns. In such a situation, simply insert the text as you normally would and the text will be automatically run through the InlinePatterns. However, if you do not want some text to be parsed by InlinePatterns, then insert the text as an AtomicString.

from markdown.util import AtomicString
some_element.text = AtomicString(some_text)

Here’s a basic example which creates an HTML table (note that the contents of the second cell (td2) will be run through InlinePatterns latter):

table = etree.Element("table") 
table.set("cellpadding", "2")                      # Set cellpadding to 2
tr = etree.SubElement(table, "tr")                 # Add child tr to table
td1 = etree.SubElement(tr, "td")                   # Add child td1 to tr
td1.text = markdown.util.AtomicString("Cell content") # Add plain text content
td2 = etree.SubElement(tr, "td")                   # Add second td to tr
td2.text = "*text* with **inline** formatting."    # Add markup text
table.tail = "Text after table"                    # Add text after table

You can also manipulate an existing tree. Consider the following example which adds a class attribute to <a> elements:

def set_link_class(self, element):
    for child in element: 
        if child.tag == "a":
            child.set("class", "myclass") #set the class attribute
        set_link_class(child) # run recursively on children

For more information about working with ElementTree see the ElementTree Documentation (Python Docs).

Integrating Your Code Into Markdown

Once you have the various pieces of your extension built, you need to tell Markdown about them and ensure that they are run in the proper sequence. Markdown accepts an Extension instance for each extension. Therefore, you will need to define a class that extends markdown.extensions.Extension and over-rides the extendMarkdown method. Within this class you will manage configuration options for your extension and attach the various processors and patterns to the Markdown instance.

It is important to note that the order of the various processors and patterns matters. For example, if we replace http://... links with <a> elements, and then try to deal with inline html, we will end up with a mess. Therefore, the various types of processors and patterns are stored within an instance of the Markdown class in OrderedDicts. Your Extension class will need to manipulate those OrderedDicts appropriately. You may insert instances of your processors and patterns into the appropriate location in an OrderedDict, remove a built-in instance, or replace a built-in instance with your own.


The extendMarkdown method of a markdown.extensions.Extension class accepts two arguments:

  • md:

    A pointer to the instance of the Markdown class. You should use this to access the OrderedDicts of processors and patterns. They are found under the following attributes:

    • md.preprocessors
    • md.inlinePatterns
    • md.parser.blockprocessors
    • md.treeprocessors
    • md.postprocessors

    Some other things you may want to access in the markdown instance are:

    • md.htmlStash
    • md.output_formats
    • md.set_output_format()
    • md.output_format
    • md.serializer
    • md.registerExtension()
    • md.html_replacement_text
    • md.tab_length
    • md.enable_attributes
    • md.smart_emphasis
  • md_globals:

    Contains all the various global variables within the markdown module.


With access to the above items, theoretically you have the option to change anything through various monkey_patching techniques. However, you should be aware that the various undocumented parts of markdown may change without notice and your monkey_patches may break with a new release. Therefore, what you really should be doing is inserting processors and patterns into the markdown pipeline. Consider yourself warned!

A simple example:

from markdown.extensions import Extension

class MyExtension(Extension):
    def extendMarkdown(self, md, md_globals):
        # Insert instance of 'mypattern' before 'references' pattern
        md.inlinePatterns.add('mypattern', MyPattern(md), '<references')


An OrderedDict is a dictionary like object that retains the order of it’s items. The items are ordered in the order in which they were appended to the OrderedDict. However, an item can also be inserted into the OrderedDict in a specific location in relation to the existing items.

Think of OrderedDict as a combination of a list and a dictionary as it has methods common to both. For example, you can get and set items using the od[key] = value syntax and the methods keys(), values(), and items() work as expected with the keys, values and items returned in the proper order. At the same time, you can use insert(), append(), and index() as you would with a list.

Generally speaking, within Markdown extensions you will be using the special helper method add() to add additional items to an existing OrderedDict.

The add() method accepts three arguments:

  • key: A string. The key is used for later reference to the item.

  • value: The object instance stored in this item.

  • location: Optional. The items location in relation to other items.

    Note that the location can consist of a few different values:

    • The special strings "_begin" and "_end" insert that item at the beginning or end of the OrderedDict respectively.

    • A less-than sign (<) followed by an existing key (i.e.: "<somekey") inserts that item before the existing key.

    • A greater-than sign (>) followed by an existing key (i.e.: ">somekey") inserts that item after the existing key.

Consider the following example:

>>> from markdown.odict import OrderedDict
>>> od = OrderedDict()
>>> od['one'] =  1           # The same as: od.add('one', 1, '_begin')
>>> od['three'] = 3          # The same as: od.add('three', 3, '>one')
>>> od['four'] = 4           # The same as: od.add('four', 4, '_end')
>>> od.items()
[("one", 1), ("three", 3), ("four", 4)]

Note that when building an OrderedDict in order, the extra features of the add method offer no real value and are not necessary. However, when manipulating an existing OrderedDict, add can be very helpful. So let’s insert another item into the OrderedDict.

>>> od.add('two', 2, '>one')         # Insert after 'one'
>>> od.values()
[1, 2, 3, 4]

Now let’s insert another item.

>>> od.add('twohalf', 2.5, '<three') # Insert before 'three'
>>> od.keys()
["one", "two", "twohalf", "three", "four"]

Note that we also could have set the location of “twohalf” to be ‘after two’ (i.e.: '>two'). However, it’s unlikely that you will have control over the order in which extensions will be loaded, and this could affect the final sorted order of an OrderedDict. For example, suppose an extension adding ‘twohalf’ in the above examples was loaded before a separate extension which adds ‘two’. You may need to take this into consideration when adding your extension components to the various markdown OrderedDicts.

Once an OrderedDict is created, the items are available via key:

MyNode = od['somekey']

Therefore, to delete an existing item:

del od['somekey']

To change the value of an existing item (leaving location unchanged):

od['somekey'] = MyNewObject()

To change the location of an existing item:

t.link('somekey', '<otherkey')


Some extensions may need to have their state reset between multiple runs of the Markdown class. For example, consider the following use of the Footnotes extension:

md = markdown.Markdown(extensions=['footnotes'])
html1 = md.convert(text_with_footnote)
html2 = md.convert(text_without_footnote)

Without calling reset, the footnote definitions from the first document will be inserted into the second document as they are still stored within the class instance. Therefore the Extension class needs to define a reset method that will reset the state of the extension (i.e.: self.footnotes = {}). However, as many extensions do not have a need for reset, reset is only called on extensions that are registered.

To register an extension, call md.registerExtension from within your extendMarkdown method:

def extendMarkdown(self, md, md_globals):
    # insert processors and patterns here

Then, each time reset is called on the Markdown instance, the reset method of each registered extension will be called as well. You should also note that reset will be called on each registered extension after it is initialized the first time. Keep that in mind when over-riding the extension’s reset method.

Config Settings

If an extension uses any parameters that the user may want to change, those parameters should be stored in self.config of your markdown.Extension class in the following format:

self.config = {parameter_1_name : [value1, description1],
               parameter_2_name : [value2, description2] }

When stored this way the config parameters can be over-ridden from the command line or at the time Markdown is initiated:

markdown.py -x myextension(SOME_PARAM=2) inputfile.txt > output.txt

Note that parameters should always be assumed to be set to string values, and should be converted at run time. For example:

i = int(self.getConfig("SOME_PARAM"))


As noted in the library reference an instance of an extension can be passed directly to Markdown. In fact, this is the prefered way to use third-party extensions.

For example:

import markdown
import myextension
configs = {...}
myext = myextension.MyExtension(configs=configs)
md = markdown.Markdown(extensions=[myext])

This is especially useful if you need to implement a large number of extensions with more than one residing in a module.

However, for historical reasons, Markdown also accepts “named” third party extensions. In that case, only one extension can be defined per module and that extension must define a module-level function called makeExtension that takes an optional parameter consisting of a dictionary of configuration over-rides and returns an instance of the extension. For example:

class MyExtension(markdown.extensions.Extension)
    # Define extension here...

def makeExtension(configs=None):
    return MyExtension(configs=configs)

When Markdown is passed the “name” of your extension as a string, it will import the module and call the makeExtension function to initiate your extension.

The “name” of your extension must be a string consisting of the importable path to your module using Python’s dot notation. Therefore, if you are providing a library to your users and would like to include a custom markdown extensions within your library, that extension would be named "mylib.mdext.myext" where mylib/mdext/myext.py contains the makeExtension function and the mylib directory is on the PYTHONPATH.

You may have noted that the extensions packaged with Python-Markdown do not use Python’s dot notation in their names. This is because they are all part of the markdown.extensions package. If a “name” contains any dots (.), then it will be imported as-is. Otherwise, Markdown will first try to import from markdown.extensions.extname and upon failure, mdx_extname where "extname" is the “name” passed to Markdown.


While the mdx_extname method of naming extensions is still supported, it remains solely for historical reasons to support the various existing third-party extensions. This method is discouraged going forward and support may be removed in a future release.