markdown-it design principles

Data flow

Input data is parsed via nested chains of rules. There are 3 nested chains - core, block & inline:

    core.rule1 (normalize)

        block.rule1 (blockquote)

    core.ruleX1 (intermediate rule that applies on block tokens, nothing yet)

    inline (applied to each block token with "inline" type)
        inline.rule1 (text)

    core.ruleYY (applies to all tokens)
    ... (abbreviation, footnote, typographer, linkifier)

The result of the parsing is a list of tokens, that will be passed to the renderer to generate the html content.

These tokens can be themselves parsed again to generate more tokens (ex: a list token can be divided into multiple inline tokens).

An env sandbox can be used alongside tokens to inject external variables for your parsers and renderers.

Each chain (core / block / inline) uses an independent state object when parsing data, so that each parsing operation is independent and can be disabled on the fly.

Token stream

Instead of traditional AST we use more low-level data representation - tokens. The difference is simple:

  • Tokens are a simple sequence (Array).

  • Opening and closing tags are separate.

  • There are special token objects, “inline containers”, having nested tokens. sequences with inline markup (bold, italic, text, …).

See token class for details about each token content.

In total, a token stream is:

  • On the top level - array of paired or single “block” tokens:

    • open/close for headers, lists, blockquotes, paragraphs, …

    • codes, fenced blocks, horizontal rules, html blocks, inlines containers

  • Each inline token have a .children property with a nested token stream for inline content:

    • open/close for strong, em, link, code, …

    • text, line breaks

Why not AST? Because it’s not needed for our tasks. We follow KISS principle. If you wish - you can call a parser without a renderer and convert the token stream to an AST.

More details about tokens:


Rules are functions, doing “magic” with parser state objects. A rule is associated with one or more chains and is unique. For instance, a blockquote token is associated with blockquote, paragraph, heading and list chains.

Rules are managed by names via Ruler instances and can be enabled / disabled from the MarkdownIt methods.

You can note, that some rules have a validation mode - in this mode rules do not modify the token stream, and only look ahead for the end of a token. It’s one important design principle - a token stream is “write only” on block & inline parse stages.

Parsers are designed to keep rules independent of each other. You can safely enable/disable them, or add new ones. There are no universal recipes for how to create new rules - design of distributed state machines with good data isolation is a tricky business. But you can investigate existing rules & plugins to see possible approaches.

Also, in complex cases you can try to ask for help in tracker. Condition is very simple - it should be clear from your ticket, that you studied docs, sources, and tried to do something yourself. We never reject with help to real developers.


After the token stream is generated, it’s passed to a renderer. It then plays all the tokens, passing each to a rule with the same name as token type.

Renderer rules are located in md.renderer.rules[name] and are simple functions with the same signature:

def function(renderer, tokens, idx, options, env):
  return htmlResult

In many cases that allows easy output change even without parser intrusion. For example, let’s replace images with vimeo links to player’s iframe:

import re
md = MarkdownIt("commonmark")

vimeoRE = re.compile(r'^https?:\/\/(www\.)?\/(\d+)($|\/)')

def render_vimeo(self, tokens, idx, options, env):
    token = tokens[idx]

    if vimeoRE.match(token.attrs["src"]):

        ident = vimeoRE.match(token.attrs["src"])[2]

        return ('<div class="embed-responsive embed-responsive-16by9">\n' +
               '  <iframe class="embed-responsive-item" src="//' +
                ident + '"></iframe>\n' +
    return self.image(tokens, idx, options, env)

md = MarkdownIt("commonmark")
md.add_render_rule("image", render_vimeo)

Here is another example, how to add target="_blank" to all links:

from markdown_it import MarkdownIt

def render_blank_link(self, tokens, idx, options, env):
    tokens[idx].attrSet("target", "_blank")

    # pass token to default renderer.
    return self.renderToken(tokens, idx, options, env)

md = MarkdownIt("commonmark")
md.add_render_rule("link_open", render_blank_link)
print(md.render("[a]\n\n[a]: b"))

Note, if you need to add attributes, you can do things without renderer override. For example, you can update tokens in core chain. That is slower, than direct renderer override, but can be more simple.

You also can write your own renderer to generate other formats than HTML, such as JSON/XML… You can even use it to generate AST.


This was mentioned in Data flow, but let’s repeat sequence again:

  1. Blocks are parsed, and top level of token stream filled with block tokens.

  2. Content on inline containers is parsed, filling .children properties.

  3. Rendering happens.

And somewhere between you can apply additional transformations :) . Full content of each chain can be seen on the top of, and files.

Also you can change output directly in renderer for many simple cases.