Skip to main content

A generator of PEG/Packrat parsers from EBNF grammars.

Reason this release was yanked:


Project description

At least for the people who send me mail about a new language that they’re designing, the general advice is: do it to learn about how to write a compiler. Don’t have any expectations that anyone will use it, unless you hook up with some sort of organization in a position to push it hard. It’s a lottery, and some can buy a lot of the tickets. There are plenty of beautiful languages (more beautiful than C) that didn’t catch on. But someone does win the lottery, and doing a language at least teaches you something.

Dennis Ritchie (1941-2011) Creator of the C programming language and of UNIX


Grako (for grammar compiler) is a tool that takes grammars in a variation of EBNF as input, and outputs memoizing (Packrat) PEG parsers in Python.

Grako is different from other PEG parser generators in that the generated parsers use Python’s very efficient exception-handling system to backtrack. Grako generated parsers simply assert what must be parsed; there are no complicated if-then-else sequences for decision making or backtracking. Positive and negative lookaheads, and the cut element allow for additional, hand-crafted optimizations at the grammar level, and delegation to Python’s re module for lexemes allows for (Perl-like) powerful and efficient lexical analysis. The use of Python’s context managers considerably reduces the size of the generated parsers for enhanced CPU-cache hits.

Grako, the runtime support, and the generated parsers have measurably low Cyclomatic complexity. At around 3000 lines of Python, it is possible to study all its source code in a single session. Grako’s only dependencies are on the Python 2.7, 3.x, or PyPy standard libraries.

Grako is feature complete and currently being used over very large grammars to parse and translate hundreds of thousands of lines of legacy code.

Table of Contents


Grako was created to address recurring problems encountered over decades of working with parser generation tools:

  • To deal with programming languages in which important statement words (can’t call them keywords) may be used as identifiers in programs, the parser must be able to lead the lexer. The parser must also lead the lexer to parse languages in which the meaning of input symbols may change with context, like Ruby.

  • When ambiguity is the norm in the parsed language (as is the case in several legacy ones), an LL or LR grammar becomes contaminated with myriads of lookaheads. PEG parsers address ambiguity from the onset. Memoization, and relying on the exception-handling system makes backtracking very efficient.

  • Semantic actions, like Abstract Syntax Tree (AST) transformation, do not belong in the grammar. Semantic actions in the grammar create yet another programming language to deal with when doing parsing and translation: the source language, the grammar language, the semantics language, the generated parser’s language, and translation’s target language. Most grammar parsers do not check that the embedded semantic actions have correct syntax, so errors get reported at awkward moments, and against the generated code, not against the source.

  • Pre-processing (like dealing with includes, fixed column formats, or Python’s structure through indentation) belong in well-designed program code, and not in the grammar.

  • It is easy to recruit help with knowledge about a mainstream programming language (Python in this case); it is not so for grammar description languages. As a grammar language becomes more complex, it becomes increasingly difficult to find who can maintain a parser. Grako grammars are in the spirit of a Translators and Interpreters 101 course (if something is hard to explain to an university student, it’s probably too complicated).

  • Generated parsers should be humanly-readable and debuggable. Looking at the generated source code is sometimes the only way to find problems in a grammar, the semantic actions, or in the parser generator itself. It’s inconvenient to trust generated code that you cannot understand.

  • Python is a great language for working in language parsing and translation.

The Generated Parsers

A Grako generated parser consists of the following classes:

  • A parser class derived from Parser which implements the parser using one method for each grammar rule:

    def myrulename(self):
  • A semantics delegate class with one semantic method per grammar rule. Each method receives as its single parameter the Abstract Syntax Tree (AST) built from the rule invocation:

    def myrulename(self, ast):
        return ast

The methods in the delegate class return the same AST received as parameter, but custom semantic classes can override the methods to have them return anything (for example, a Semantic Graph). The semantics class can be used as a template for the final semantics implementation, which can omit methods for the rules it is not interested in.

Using the Tool

Grako is run from the command line:

$ python -m grako


$ scripts/grako

or just:

$ grako

if Grako was installed using easy_install or pip.

The -h and –help parameters provide full usage information:

$ python -m grako -h
usage: grako [-h] [-m name] [-o outfile] [-v] grammar

Grako (for grammar compiler) takes grammars in a variation of EBNF as input,
and outputs a memoizing PEG parser in Python.

positional arguments:
  grammar               The file name of the grammar to generate a parser for

optional arguments:
  -h, --help            show this help message and exit
  -m name, --name name  An optional name for the grammar. It defaults to the
                        basename of the grammar file's name
  -o outfile, --outfile outfile
                        specify where the output should go (default is stdout)
  -t, --trace           produce verbose parsing output
  -b, --binary          generate a pickled grammar model instead of a parser
  -d, --draw            generate a diagram of the grammar


Using the Generated Parser

To use the generated parser, just subclass the base or the abstract parser, create an instance of it, and invoke its parse() method passing the grammar to parse and the starting rule’s name as parameter:

parser = MyParser()
ast = parser.parse('text to parse', rule_name='start')
print(json.dumps(ast, indent=2)) # ASTs are JSON-friendy

This is more or less what happens if you invoke the generated parser directly:

python inputfile startrule

The generated parsers’ constructors accept named arguments to specify whitespace characters, the regular expression for comments, case sensitivity, verbosity, and more (see below).

To add semantic actions, just pass a semantic delegate to the parse method:

model = parser.parse(text, rule_name='start', semantics=MySemantics())

The EBNF Grammar Syntax

Grako uses a variant of the standard EBNF syntax. A grammar consists of a sequence of one or more rules of the form:

name = expre ;


name = expre .

Both the semicolon (;) and the period (.) are accepted as rule definition terminators.

If a name collides with a Python keyword, an underscore (_) will be appended to it on the generated parser.

If you define more than one rule with the same name:

name = expre1 ;
name = expre2 ;

The result will be equivalent to applying the choice operator to the right-hand-side expressions:

name = expre1 | expre2 ;

Rule names that start with an uppercase character:

FRAGMENT = ?/[a-z]+/?

do not advance over whitespace before beginning to parse. This feature becomes handy when defining complex lexical elements, as it allows breaking them into several rules.

The expressions, in reverse order of operator precedence, can be:

e1 | e2

Match either e1 or e2.

e1 e2

Match e1 and then match e2.

( e )

Grouping. Match e. Note that the AST for the group will be a list if more than one element is matched.

[ e ]

Optionally match e.

{ e } or { e }*

Closure. Match e zero or more times. Note that the AST returned for a closure is always a list.

{ e }+ or { e }-

Closure+1. Match e one or more times.


Positive lookahead. Try parsing e, but do not consume any input.


Negative lookahead. Try parsing e and fail if there’s a match. Do not consume any input whichever the outcome.

'text' or "text"

Match the token text within the quotation marks.

Note that if text is alphanumeric, then Grako will check that the character following the token is not alphanumeric. This is done to prevent tokens like IN matching when the text ahead is INITIALIZE. This feature can be turned off by passing nameguard=False to the Parser or the Buffer, or by using a pattern expression (see below) instead of a token expression.


The pattern expression. Match the Python regular expression regexp at the current text position. Unlike other expressions, this one does not advance over whitespace or comments. For that, place the regexp as the only term in its own rule.

The regexp is passed as-is to the Python re module, using re.match() at the current position in the text. The matched text is the AST for the expression.


Invoke the rule named rulename. To help with lexical aspects of grammars, rules with names that begin with an uppercase letter will not advance the input over whitespace or comments.


The empty expression. Succeed without advancing over input.


The fail expression. This is actually ! applied to (), which always fails.


The cut expression. After this point, prevent other options from being considered even if the current option fails to parse.


Add the result of e to the AST using name as key. If more than one item is added with the same name, the entry is converted to a list.


Add the result of e to the AST using name as key. Force the entry to be a list even if only one element is added.


The override operator. Make the AST for the complete rule be the AST for e.

The override operator is useful to recover only part of the right hand side of a rule without the need to name it, and then add a semantic action to recover the interesting part.

This is a typical use of the override operator:

subexp = '(' @expre ')' .

The AST returned for the subexp rule will be the AST recovered from invoking expre, without having to write a semantic action.

When there are no named items in a rule, the AST consists of the elements parsed by the rule, either a single item or a list. This default behavior makes it easier to write simple rules:

number = ?/[0-9]+/? .

without having to write:

number = number:?/[0-9]+/?

When a rule has named elements, the unnamed ones are excluded from the AST (they are ignored).

Abstract Syntax Trees (ASTs)

By default, and AST is either a list (for closures and rules without named elements), or dict-derived object that contains one item for every named element in the grammar rule. Items can be accessed through the standard dict syntax, ast['key'], or as attributes, ast.key.

AST entries are single values if only one item was associated with a name, or lists if more than one item was matched. There’s a provision in the grammar syntax (the +: operator) to force an AST entry to be a list even if only one element was matched. The value for named elements that were not found during the parse (perhaps because they are optional) is None.

When the parseinfo=True keyword argument has been passed to the Parser constructor, a parseinfo element is added to AST nodes that are dict-like. The element contains a namedtuple with the parse information for the node:

ParseInfo = namedtuple('ParseInfo', ['buffer', 'rule', 'pos', 'endpos'])

With the help of the Buffer.line_info() method, it is possible to recover the line, column, and original text parsed for the node. Note that when parseinfo is generated, the buffer used during parsing is kept in memory with the AST.


By default, Grako generated parsers skip the usual whitespace characters (whatever Python defines as string.whitespace), but you can change that behaviour by passing a whitespace parameter to your parser. For example:

parser = MyParser(text, whitespace='\t ')

will not consider end-of-line characters as whitespace.

If you don’t define any whitespace characters:

parser = MyParser(text, whitespace='')

then you will have to handle whitespace in your grammar rules (as it’s often done in PEG parsers).

Case Sensitivity

If the source language is case insensitive, you can tell your parser by using the ignorecase parameter:

parser = MyParser(text, ignorecase=True)

The change will affect both token and pattern matching.


Parsers will skip over comments specified as a regular expression using the comments_re parameter:

parser = MyParser(text, comments_re="\(\*.*?\*\)")

For more complex comment handling, you can override the Parser._eatcomments() method.

Semantic Actions

There are no constructs for semantic actions in Grako grammars. This is on purpose, as we believe that semantic actions obscure the declarative nature of grammars and provide for poor modularization from the parser execution perspective.

The per-rule methods in classes implementing the semantics provide enough opportunity to do rule post-processing operations, like verifications (for inadequate use of keywords as identifiers), or AST transformation.

For finer-grained control it is enough to declare more rules, as the impact on the parsing times will be minimal.

If pre-processing is required at some point, it is enough to place invocations of empty rules where appropriate:

myrule = first_part preproc {second_part} ;

preproc = () ;

The abstract parser will honor as a semantic action a method declared as:

def preproc(self, ast):

Templates and Translation

Grako doesn’t impose a way to create translators with it, but it exposes the facilities it uses to generate the Python source code for parsers.

Translation in Grako is template-based, but instead of defining or using a complex templating engine (yet another language), it relies on the simple but powerful string.Formatter of the Python standard library. The templates are simple strings that, in Grako’s style, are inlined with the code.

To generate a parser, Grako constructs an object model of the parsed grammar. Each node in the model is a descendant of rendering.Renderer, and knows how to render itself. Templates are left-trimmed on whitespace, like Python doc-comments are. This is an example taken from Grako’s source code:

class LookaheadGrammar(_DecoratorGrammar):


    template = '''\
                with self._if():

Every attribute of the object that doesn’t start with an underscore (_) may be used as a template field, and fields can be added or modified by overriding the render_fields() method. Fields themselves are lazily rendered before being expanded by the template, so a field may be an instance of a Renderer descendant.

The rendering module uses a Formatter enhanced to support the rendering of items in an iterable one by one. The syntax to achieve that is:


All of ind, sep, and fmt are optional, but the three colons are not. Such a field will be rendered using:

indent(sep.join(fmt % render(v) for v in value), ind)

The extended format can also be used with non-iterables, in which case the rendering will be:

indent(fmt % render(value), ind)

The default multiplier for ind is 4, but that can be overridden using n*m (for example 3*1) in the format.


Using a newline (\n) as separator will interfere with left trimming and indentation of templates. To use newline as separator, specify it as \\n, and the renderer will understand the intention.



The file etc/grako.ebnf contains a grammar for the Grako EBNF language written in the same language. It is used in the bootstrap test suite to prove that Grako can generate a parser to parse its own language.


The project examples/regexp contains a regexp-to-EBNF translator and parser generator. The project has no practical use, but it’s a complete, end-to-end example of how to implement a translator using Grako.


The project examples/antlr2grako contains a ANTLR to Grako grammar tanslator. The project is a good example of the use of models and templates in translation. The program, generates the Grako gramar on standard ouput, but because the model used is Grako’s own, the same code can be used to directly generate a parser from an ANTLR grammar. Please take a look at the examples README to know about limitations.


Grako is Copyright 2012-2013 by ResQSoft Inc. and Juancarlo Añez

You may use the tool under the terms of the BSD-style license described in the enclosed LICENSE.txt file.

If your project requires different licensing please contact

Contact and Updates

To discuss Grako and to receive notifications about new releases, please join the low-volume Grako Forum at Google Groups.


The following must be mentioned as contributors of thoughts, ideas, code, and funding to the Grako project:

  • Niklaus Wirth was the chief designer of the programming languages Euler, Algol W, Pascal, Modula, Modula-2, Oberon, Oberon-2, and Oberon-07. In the last chapter of his 1976 book Algorithms + Data Structures = Programs, Wirth creates a top-down, descent parser with recovery for the Pascal-like, LL(1) programming language PL/0. The structure of the program is that of a PEG parser, though the concept of PEG wasn’t formalized until 2004.

  • Bryan Ford introduced PEG (parsing expression grammars) in 2004.

  • Other parser generators like PEG.js by David Majda inspired the work in Grako.

  • William Thompson inspired the use of context managers with his blog post that I knew about through the invaluable Python Weekly newsletter, curated by Rahul Chaudhary

  • Jeff Knupp explains why Grako’s use of exceptions is sound, so I don’t have to.

  • Terence Parr created ANTLR, probably the most solid and professional parser generator out there. Ter, ANTLR, and the folks on the ANLTR forums helped me shape my ideas about Grako.

  • JavaCC (originally Jack) looks like an abandoned project. It was the first parser generator I used while teaching.

  • Grako is very fast. But dealing with millions of lines of legacy source code in a matter of minutes would be impossible without PyPy, the work of Armin Rigo and the PyPy team.

  • Guido van Rossum created and has lead the development of the Python programming environment for over a decade. A tool like Grako, at under three thousand lines of code, would not have been possible without Python.

  • Kota Mizushima welcomed me to the CSAIL at MIT PEG and Packrat parsing mailing list, and immediately offered ideas and pointed me to documentation about the implementation of cut in modern parsers. The optimization of memoization information is thanks to one of his papers.

  • My students at UCAB inspired me to think about how grammar-based parser generation could be made more approachable.

  • Gustavo Lau was my professor of Language Theory at USB, and he was kind enough to be my tutor in a thesis project on programming languages that was more than I could chew. My peers, and then teaching advisers Alberto Torres, and Enzo Chiariotti formed a team with Gustavo to challenge us with programming languages like LATORTA and term exams that went well into the eight hours. And, of course, there was also the pirate patch that should be worn on the left or right eye depending on the LL or LR challenge.

  • Manuel Rey led me through another, unfinished thesis project that taught me about what languages (spoken languages in general, and programming languages in particular) are about. I learned why languages use declensions, and why, although the underlying words are in English, the structure of the programs we write is more like Japanese.

  • Grako would not have been possible without the vision, the funding, and the trust provided by Thomas Bragg through ResQSoft.


  • 2.0.2
    • BUG! Trance information off by one character.

    • BUG! The AST for a closure might fold repeated symbols.

    • BUG! It was not possible to pass buffering parameters such as whitespace to the parser’s constructor.

  • 2.0.1
    • Republished to solve problems with md5 checksums on PyPi.

  • 2.0.0
    • Grako no longer assumes that parsers implement the semantics. A separate semantics implementation must be provided. This allows for less poluted namespaces and smaller classes.

    • A last_node protocol allowed the removal of all mentions of variable _e from generated parsers, which are thus more readable.

    • Refactored closures to be more pythonic (there are no anonymous blocks in Python!).

    • Fixes to the antlr2grako example to let it convert over 6000 lines of an ANTLR gramar to Grako.

    • Improved rendering of grammars by grammar models.

    • Now tokens accept Python escape sequences.

    • Added a simple Visitor Pattern for Renderer nodes. Used it to implement diagramming.

    • Create a basic diagram of a grammar if pygraphviz is available. Added the --draw option to the command-line tool.

For the complete history, consult Bitbucket or PyPi

Project details

Release history Release notifications | RSS feed

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

grako-2.0.2.tar.gz (41.5 kB view hashes)

Uploaded Source

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page