Coconut Documentation


This documentation covers all the features of the Coconut Programming Language, and is intended as a reference/specification, not a tutorialized introduction. For a full introduction and tutorial of Coconut, see the tutorial.

Coconut is a variant of Python built for simple, elegant, Pythonic functional programming. Coconut syntax is a strict superset of Python 3 syntax. Thus, users familiar with Python will already be familiar with most of Coconut.

The Coconut compiler turns Coconut code into Python code. The primary method of accessing the Coconut compiler is through the Coconut command-line utility, which also features an interpreter for real-time compilation. In addition to the command-line utility, Coconut also supports the use of IPython/Jupyter notebooks.

While most of Coconut gets its inspiration simply from trying to make functional programming work in Python, additional inspiration came from Haskell, CoffeeScript, F#, and



Using Pip

Since Coconut is hosted on the Python Package Index, it can be installed easily using pip. Simply install Python, open up a command-line prompt, and enter

pip install coconut

which will install Coconut and its required dependencies.

Note: If you have an old version of Coconut installed and you want to upgrade, run pip install --upgrade coconut instead.

If you are encountering errors running pip install coconut, try re-running it with the --user option. If pip install coconut works, but you cannot access the coconut command, be sure that Coconut's installation location is in your PATH environment variable. On UNIX, that is /usr/local/bin (without --user) or ${HOME}/.local/bin/ (with --user).

Using Conda

If you prefer to use conda instead of pip to manage your Python packages, you can also install Coconut using conda. Just install conda, open up a command-line prompt, and enter

conda config --add channels conda-forge
conda install coconut

which will properly create and build a conda recipe out of Coconut's conda-forge feedstock.

Note: To use conda to install coconut-develop instead, just replace coconut with coconut-develop in the last three commands above.

Optional Dependencies

Coconut also has optional dependencies, which can be installed by entering

pip install coconut[name_of_optional_dependency]

or, to install multiple optional dependencies,

pip install coocnut[opt_dep_1,opt_dep_2]

The full list of optional dependencies is:

  • all: alias for jupyter,watch,jobs,mypy (this is the recommended way to install a feature-complete version of Coconut),
  • jupyter/ipython: enables use of the --jupyter / --ipython flag,
  • watch: enables use of the --watch flag,
  • jobs: improves use of the --jobs flag,
  • mypy: enables use of the --mypy flag,
  • cPyparsing: significantly speeds up compilation if your platform supports it by making use of cPyparsing,
  • tests: everything necessary to run Coconut's test suite,
  • docs: everything necessary to build Coconut's documentation, and
  • dev: everything necessary to develop on Coconut, including all of the dependencies above.

Develop Version

Alternatively, if you want to test out Coconut's latest and greatest, enter

pip install coconut-develop

which will install the most recent working version from Coconut's develop branch. Optional dependency installation is supported in the same manner as above. For more information on the current development build, check out the development version of this documentation. Be warned: coconut-develop is likely to be unstable—if you find a bug, please report it by creating a new issue.


coconut [-h] [-v] [-t version] [-i] [-p] [-a] [-l] [-k] [-w] [-r] [-n]
        [-d] [-q] [-s] [--no-tco] [-c code] [-j processes] [-f]
        [--minify] [--jupyter ...] [--mypy ...] [--argv ...]
        [--tutorial] [--documentation] [--style name]
        [--recursion-limit limit] [--verbose] [--trace]
        [source] [dest]

Positional Arguments

source              path to the Coconut file/folder to compile
dest                destination directory for compiled files (defaults to
                    the source directory)

Optional Arguments

-h, --help            show this help message and exit
-v, --version         print Coconut and Python version information
-t version, --target version
                      specify target Python version (defaults to universal)
-i, --interact        force the interpreter to start (otherwise starts if no
                      other command is given) (implies --run)
-p, --package         compile source as part of a package (defaults to only
                      if source is a directory)
-a, --standalone      compile source as standalone files (defaults to only
                      if source is a single file)
-l, --line-numbers, --linenumbers
                      add line number comments for ease of debugging
-k, --keep-lines, --keeplines
                      include source code in comments for ease of debugging
-w, --watch           watch a directory and recompile on changes
-r, --run             execute compiled Python
-n, --no-write, --nowrite
                      disable writing compiled Python
-d, --display         print compiled Python
-q, --quiet           suppress all informational output (combine with
                      --display to write runnable code to stdout)
-s, --strict          enforce code cleanliness standards
--no-tco, --notco     disable tail call optimization
-c code, --code code  run Coconut passed in as a string (can also be piped
                      into stdin)
-j processes, --jobs processes
                      number of additional processes to use (defaults to 0)
                      (pass 'sys' to use machine default)
-f, --force           force overwriting of compiled Python (otherwise only
                      overwrites when source code or compilation parameters
--minify              reduce size of compiled Python
--jupyter ..., --ipython ...
                      run Jupyter/IPython with Coconut as the kernel
                      (remaining args passed to Jupyter)
--mypy ...            run MyPy on compiled Python (remaining args passed to
                      MyPy) (implies --package --no-tco)
--argv ...            set sys.argv to source plus remaining args for use in
                      Coconut script being run
--tutorial            open Coconut's tutorial in the default web browser
--documentation       open Coconut's documentation in the default web
--style name          Pygments syntax highlighting style (or 'none' to
                      disable) (defaults to COCONUT_STYLE environment
                      variable, if it exists, otherwise 'default')
--recursion-limit limit, --recursionlimit limit
                      set maximum recursion depth in compiler (defaults to
--verbose             print verbose debug output
--trace               print verbose parsing data (only available in coconut-

Coconut Scripts

To run a Coconut file as a script, Coconut provides the command

coconut-run <source> <args>

as an alias for

coconut --run --quiet --target sys <source> --argv <args>

which will quietly compile and run <source>, passing any additional arguments to the script, mimicking how the python command works.

coconut-run can be used in a Unix shebang line to create a Coconut script by adding the following line to the start of your script:

#!/usr/bin/env coconut-run

Naming Source Files

Coconut source files should, so the compiler can recognize them, use the extension .coco (preferred), .coc, or .coconut. When Coconut compiles a .coco (or .coc/.coconut) file, it will compile to another file with the same name, except with .py instead of .coco, which will hold the compiled code. If an extension other than .py is desired for the compiled files, such as .pyde for Python Processing, then that extension can be put before .coco in the source file name, and it will be used instead of .py for the compiled files. For example, name.coco will compile to, whereas name.pyde.coco will compile to name.pyde.

Compilation Modes

Files compiled by the coconut command-line utility will vary based on compilation parameters. If an entire directory of files is compiled (which the compiler will search recursively for any folders containing .coco, .coc, or .coconut files), a file will be created to house necessary functions (package mode), whereas if only a single file is compiled, that information will be stored within a header inside the file (standalone mode). Standalone mode is better for single files because it gets rid of the overhead involved in importing, but package mode is better for large packages because it gets rid of the need to run the same Coconut header code again in every file, since it can just be imported from

By default, if the source argument to the command-line utility is a file, it will perform standalone compilation on it, whereas if it is a directory, it will recursively search for all .coco (or .coc / .coconut) files and perform package compilation on them. Thus, in most cases, the mode chosen by Coconut automatically will be the right one. But if it is very important that no additional files like be created, for example, then the command-line utility can also be forced to use a specific mode with the --package (-p) and --standalone (-a) flags.

Compatible Python Versions

While Coconut syntax is based off of Python 3, Coconut code compiled in universal mode (the default --target), and the Coconut compiler, should run on any Python version >= 2.6 on the 2.x branch or >= 3.2 on the 3.x branch.

Note: The tested against implementations are CPython 2.6, 2.7, 3.2, 3.3, 3.4, 3.5, 3.6 and PyPy 2.7, 3.2.

To make Coconut built-ins universal across Python versions, Coconut automatically overwrites Python 2 built-ins with their Python 3 counterparts. Additionally, Coconut also overwrites some Python 3 built-ins for optimization and enhancement purposes. If access to the original Python versions of any overwritten built-ins is desired, the old built-ins can be retrieved by prefixing them with py_.

For standard library compatibility, Coconut automatically maps imports under Python 3 names to imports under Python 2 names. Thus, Coconut will automatically take care of any standard library modules that were renamed from Python 2 to Python 3 if just the Python 3 name is used. For modules or objects that only exist in Python 3, however, Coconut has no way of maintaining compatibility.

Finally, while Coconut will try to compile Python-3-specific syntax to its universal equivalent, the following constructs have no equivalent in Python 2, and require the specification of a target of at least 3 to be used:

  • destructuring assignment with *s (use Coconut pattern-matching instead),
  • the nonlocal keyword,
  • exec used in a context where it must be a function,
  • keyword class definition,
  • tuples and lists with * unpacking or dicts with ** unpacking (requires --target 3.5),
  • @ as matrix multiplication (requires --target 3.5),
  • async and await statements (requires --target 3.5), and
  • formatting strings by prefixing them with f (requires --target 3.6).

Allowable Targets

If the version of Python that the compiled code will be running on is known ahead of time, a target should be specified with --target. The given target will only affect the compiled code and whether or not the Python-3-specific syntax detailed above is allowed. Where Python 3 and Python 2 syntax standards differ, Coconut syntax will always follow Python 3 across all targets. The supported targets are:

  • universal (default) (will work on any of the below),
  • 2, 2.6 (will work on any Python >= 2.6 but < 3),
  • 2.7 (will work on any Python >= 2.7 but < 3),
  • 3, 3.2 (will work on any Python >= 3.2),
  • 3.3, 3.4 (will work on any Python >= 3.3),
  • 3.5 (will work on any Python >= 3.5),
  • 3.6 (will work on any Python >= 3.6),
  • sys (chooses the specific target corresponding to the current version).

Note: Periods are ignored in target specifications, such that the target 27 is equivalent to the target 2.7.

strict Mode

If the --strict (or -s) flag is enabled, Coconut will throw errors on various style problems. These are

  • mixing of tabs and spaces (without --strict will show a Warning),
  • use of from __future__ imports (without --strict will show a Warning)
  • missing new line at end of file,
  • trailing whitespace at end of lines,
  • semicolons at end of lines,
  • use of the Python-style lambda statement,
  • inheriting from object in classes (Coconut does this automatically),
  • use of u to denote Unicode strings (all Coconut strings are Unicode strings), and
  • use of backslash continuations (use parenthetical continuation instead).

Additionally, --strict disables deprecated features, making them entirely unavailable to code compiled with --strict. It is recommended that you use the --strict (or -s) flag if you are starting a new Coconut project, as it will help you write cleaner code.


Syntax Highlighting

Text editors with support for Coconut syntax highlighting are:

Alternatively, if none of the above work for you, you can just treat Coconut as Python. Simply set up your editor so it interprets all .coco files as Python and that should highlight most of your code well enough.


Coconut syntax highlighting for SublimeText requires that Package Control, the standard package manager for SublimeText, be installed. Once that is done, simply:

  1. open the SublimeText command palette by pressing Ctrl+Shift+P (or Cmd+Shift+P on Mac),
  2. type and enter Package Control: Install Package, and
  3. finally type and enter Coconut.

To make sure everything is working properly, open a .coco file, and make sure Coconut appears in the bottom right-hand corner. If something else appears, like Plain Text, click on it, select Open all with current extension as... at the top of the resulting menu, and then select Coconut.

Note: Coconut syntax highlighting for SublimeText is provided by the sublime-coconut package.


The same pip install coconut command that installs the Coconut command-line utility will also install the coconut Pygments lexer. How to use this lexer depends on the Pygments-enabled application being used, but in general simply enter coconut as the language being highlighted and/or use a valid Coconut file extension (.coco, .coc, or .coconut) and Pygments should be able to figure it out. For example, this documentation is generated with Sphinx, with the syntax highlighting you see created by adding the line

highlight_language = "coconut"

to Coconut's

IPython/Jupyter Support

If you prefer IPython (the python kernel for the Jupyter framework) to the normal Python shell, Coconut can be used as a Jupyter kernel or IPython extension.


If Coconut is used as a kernel, all code in the console or notebook will be sent directly to Coconut instead of Python to be evaluated. Otherwise, the Coconut kernel behaves exactly like the IPython kernel, including support for %magic commands.

The command coconut --jupyter notebook (or coconut --ipython notebook) will launch an IPython/Jupyter notebook using Coconut as the kernel and the command coconut --jupyter console (or coconut --ipython console) will launch an IPython/Jupyter console using Coconut as the kernel. Additionally, the command coconut --jupyter (or coconut --ipython) will add Coconut as a language option inside of all IPython/Jupyter notebooks, even those not launched with Coconut. This command may need to be re-run when a new version of Coconut is installed.


If Coconut is used as an extension, a special magic command will send snippets of code to be evaluated using Coconut instead of IPython, but IPython will still be used as the default.

The line magic %load_ext coconut will load Coconut as an extension, providing the %coconut and %%coconut magics and adding Coconut built-ins. The %coconut line magic will run a line of Coconut with default parameters, and the %%coconut block magic will take command-line arguments on the first line, and run any Coconut code provided in the rest of the cell with those parameters.

MyPy Integration

Coconut has the ability to integrate with MyPy to provide optional static type-checking, including for all Coconut built-ins. Simply pass --mypy to enable MyPy integration, though be careful to pass it only as the last argument, since all arguments after --mypy are passed to mypy, not Coconut.

Note: Since tail call optimization prevents proper type-checking, --mypy implicitly disables it.

To explicitly annotate your code with types for MyPy to check, Coconut supports Python 3 function type annotations, Python 3.6 variable type annotations, and even Coconut's own enhanced type annotation syntax. By default, all type annotations are compiled to Python-2-compatible type comments, which means all of the above works on any Python version.

Coconut even supports --mypy in the interpreter, which will intelligently scan each new line of code, in the context of previous lines, for newly-introduced MyPy errors. For example:

>>> a: str = count()[0]
<string>:14: error: Incompatible types in assignment (expression has type "int", variable has type "str")

Note: Sometimes, MyPy will not know how to handle certain Coconut constructs, such as addpattern. In that case, simply put a # type: ignore comment on the Coconut line which is generating the line MyPy is complaining about (you can figure out what line this is using --line-numbers) and the comment will be added to every generated line.


In order of precedence, highest first, the operators supported in Coconut are:

===================== ==========================
Symbol(s)             Associativity
===================== ==========================
..                    n/a won't capture call
**                    right
+, -, ~               unary
*, /, //, %, @        left
+, -                  left
<<, >>                left
&                     left
^                     left
|                     left
::                    n/a lazy
a `b` c               left captures lambda
??                    left short-circuit
..>, <.., ..*>, <*..  n/a captures lambda
|>, <|, |*>, <*|      left captures lambda
==, !=, <, >,
  <=, >=,
  in, not in,
  is, is not          n/a
not                   unary
and                   left short-circuit
or                    left short-circuit
a if b else c         ternary left short-circuit
->                    right
===================== ==========================


Coconut provides the simple, clean -> operator as an alternative to Python's lambda statements. The syntax for the -> operator is (parameters) -> expression (or parameter -> expression for one-argument lambdas). The operator has the same precedence as the old statement, which means it will often be necessary to surround the lambda in parentheses, and is right-associative.

Additionally, Coconut also supports an implicit usage of the -> operator of the form (-> expression), which is equivalent to ((_=None) -> expression), which allows an implicit lambda to be used both when no arguments are required, and when one argument (assigned to _) is required.

Note: If normal lambda syntax is insufficient, Coconut also supports an extended lambda syntax in the form of statement lambdas.


In Python, lambdas are ugly and bulky, requiring the entire word lambda to be written out every time one is constructed. This is fine if in-line functions are very rarely needed, but in functional programming in-line functions are an essential tool.

Python Docs

Lambda forms (lambda expressions) have the same syntactic position as expressions. They are a shorthand to create anonymous functions; the expression (arguments) -> expression yields a function object. The unnamed object behaves like a function object defined with:

def <lambda>(arguments):
    return expression

Note that functions created with lambda forms cannot contain statements or annotations.



dubsums = map((x, y) -> 2*(x+y), range(0, 10), range(10, 20))
dubsums |> list |> print


dubsums = map(lambda x, y: 2*(x+y), range(0, 10), range(10, 20))

Partial Application

Coconut uses a $ sign right after a function's name but before the open parenthesis used to call the function to denote partial application.

Coconut's partial application also supports the use of a ? to skip partially applying an argument, deferring filling in that argument until the partially-applied function is called. This is useful if you want to partially apply arguments that aren't first in the argument order.


Partial application, or currying, is a mainstay of functional programming, and for good reason: it allows the dynamic customization of functions to fit the needs of where they are being used. Partial application allows a new function to be created out of an old function with some of its arguments pre-specified.

Python Docs

Return a new partial object which when called will behave like func called with the positional arguments args and keyword arguments keywords. If more arguments are supplied to the call, they are appended to args. If additional keyword arguments are supplied, they extend and override keywords. Roughly equivalent to:

def partial(func, *args, **keywords):
    def newfunc(*fargs, **fkeywords):
        newkeywords = keywords.copy()
        return func(*(args + fargs), **newkeywords)
    newfunc.func = func
    newfunc.args = args
    newfunc.keywords = keywords
    return newfunc

The partial object is used for partial function application which “freezes” some portion of a function's arguments and/or keywords resulting in a new object with a simplified signature.



expnums = range(5) |> map$(pow$(?, 2))
expnums |> list |> print


# unlike this simple lambda, $ produces a pickleable object
expnums = map(lambda x: pow(x, 2), range(5))


Coconut uses pipe operators for pipeline-style function application. All the operators have a precedence in-between function composition pipes and comparisons, and are left-associative. All operators also support in-place versions. The different operators are:

(|>)    => pipe forward
(|*>)   => multiple-argument pipe forward
(<|)    => pipe backward
(<*|)   => multiple-argument pipe backward

Additionally, all pipe operators support a lambda as the last argument, despite lambdas having a lower precedence. Thus, 10 |> x -> x**2 is valid, though the body of the lambda will still capture all following pipe operators.


It is common in Coconut to write code that uses pipes to pass an object through a series of partials and/or implicit partials, as in

obj |> .attribute |> .method(args) |> func$(args) |> .[index]

which is often much more readable, as it allows the operations to be written in the order in which they are performed, instead of as in

func(args, obj.attribute.method(args))[index]

where func has to go at the beginning.

If Coconut compiled each of the partials in the pipe syntax as an actual partial application object, it would make the Coconut-style syntax here so much slower than the Python-style syntax as to make it near-useless. Thus, Coconut does not do that. If any of the above styles of partials or implicit partials are used in pipes, they will whenever possible be compiled to the Python-style syntax, producing no intermediate partial application objects.



def sq(x) = x**2
(1, 2) |*> (+) |> sq |> print


import operator
def sq(x): return x**2
print(sq(operator.add(1, 2)))


Coconut has three basic function composition operators: .., ..>, and <... Both .. and <.. use math-style "backwards" function composition, where the first function is called last, while ..> uses "forwards" function composition, where the first function is called first.

The ..> and <.. function composition pipe operators also have ..*> and <*.. forms which are, respectively, the equivalents of |*> and <*|. Forwards and backwards function composition pipes cannot be used together in the same expression (unlike normal pipes) and have precedence in-between None-coalescing and normal pipes.

The .. operator has lower precedence than attribute access (.), slicing ([]), etc., except for function calling, which it has higher precedence than. Thus, a.b..c.d is equivalent to (a.b)..(c.d), while f..g(x) is equivalent to (f..g)(x).

The in-place function composition operators are ..=, ..>=, <..=, ..*>=, and <*..=.



fog = f..g
f_into_g = f ..> g


# unlike these simple lambdas, Coconut produces pickleable objects
fog = lambda *args, **kwargs: f(g(*args, **kwargs))
f_into_g = lambda *args, **kwargs: g(f(*args, **kwargs))


Coconut uses the :: operator for iterator chaining. Coconut's iterator chaining is done lazily, in that the arguments are not evaluated until they are needed. It has a precedence in-between bitwise or and infix calls. The in-place operator is ::=.


A useful tool to make working with iterators as easy as working with sequences is the ability to lazily combine multiple iterators together. This operation is called chain, and is equivalent to addition with sequences, except that nothing gets evaluated until it is needed.

Python Docs

Make an iterator that returns elements from the first iterable until it is exhausted, then proceeds to the next iterable, until all of the iterables are exhausted. Used for treating consecutive sequences as a single sequence. Chained inputs are evaluated lazily. Roughly equivalent to:

def chain(*iterables):
    # chain('ABC', 'DEF') --> A B C D E F
    for it in iterables:
        for element in it:
            yield element



def N(n=0) = (n,) :: N(n+1)  # no infinite loop because :: is lazy

(range(-10, 0) :: N())$[5:15] |> list |> print

Python: Can't be done without a complicated iterator comprehension in place of the lazy chaining. See the compiled code for the Python syntax.

Iterator Slicing

Coconut uses a $ sign right after an iterator before a slice to perform iterator slicing. Coconut's iterator slicing works much the same as Python's sequence slicing, and looks much the same as Coconut's partial application, but with brackets instead of parentheses.

Iterator slicing works just like sequence slicing, including support for negative indices and slices, and support for slice objects in the same way as can be done with normal slicing. Iterator slicing makes no guarantee, however, that the original iterator passed to it be preserved (to preserve the iterator, use Coconut's tee or reiterable built-in).

Coconut's iterator slicing is very similar to Python's itertools.islice, but unlike itertools.islice, Coconut's iterator slicing supports negative indices, and will preferentially call an object's __getitem__, if it exists. Coconut's iterator slicing is also optimized to work well with all of Coconut's built-in objects, only computing the elements of each that are actually necessary to extract the desired slice.



map(x -> x*2, range(10**100))$[-1] |> print

Python: Can't be done without a complicated iterator slicing function and inspection of custom objects. The necessary definitions in Python can be found in the Coconut header.

None Coalescing

Coconut provides ?? as a None-coalescing operator, similar to the ?? null-coalescing operator in C#. The None-coalescing operator evaluates to its left operand if that operand is not None, otherwise its right operand. The None-coalescing operator is short-circuiting, such that if the left operand is not None, it will not evaluate the right operand. The None-coalescing operator has a precedence in-between infix function calls and composition pipes, and is left-associative. The in-place operator is ??=.

Coconut also allows a single ? before attribute access, function calling, partial application, and (iterator) indexing to short-circuit the rest of the evaluation if everything so far evaluates to None. Thus, a?.b is equivalent to a.b if a is not None else a.



could_be_none() ?? calculate_default_value()


(lambda result: result if result is not None else calculate_default_value())(could_be_none())
(lambda result: None if result is None else result.attr[index].method())(could_be_none())

Unicode Alternatives

Coconut supports Unicode alternatives to many different operator symbols. The Unicode alternatives are relatively straightforward, and chosen to reflect the look and/or meaning of the original symbol.

Full List

 (\u2192)                  => "->"
 (\u21a6)                  => "|>"
*↦ (*\u21a6)                => "|*>"
 (\u21a4)                  => "<|"
↤* (\u21a4*)                => "<*|"
 (\u22c5)                  => "*"
 (\u2191)                  => "**"
÷ (\xf7)                    => "/"
÷/ (\xf7/)                  => "//"
 (\u2218)                  => ".."
∘> (\u2218>)                => "..>"
<∘ (<\u2218)                => "<.."
∘*> (\u2218*>)              => "..*>"
<*∘ (<*\u2218)              => "<*.."
 (\u2212)                  => "-" (only subtraction)
 (\u207b)                  => "-" (only negation)
¬ (\xac)                    => "~"
 (\u2260) or ¬= (\xac=)    => "!="
 (\u2264)                  => "<="
 (\u2265)                  => ">="
 (\u2227) or  (\u2229)    => "&"
 (\u2228) or  (\u222a)    => "|"
 (\u22bb) or  (\u2295)    => "^"
« (\xab)                    => "<<"
» (\xbb)                    => ">>"
 (\u2026)                  => "..."
× (\xd7)                    => "@" (only matrix multiplication)



Coconut's data keyword is used to create immutable, algebraic data types with built-in support for destructuring pattern-matching and fmap.

The syntax for data blocks is a cross between the syntax for functions and the syntax for classes. The first line looks like a function definition, but the rest of the body looks like a class, usually containing method definitions. This is because while data blocks actually end up as classes in Python, Coconut automatically creates a special, immutable constructor based on the given arguments.

Coconut data statement syntax looks like:

data <name>(<args>) [from <inherits>]:

<name> is the name of the new data type, <args> are the arguments to its constructor as well as the names of its attributes, <body> contains the data type's methods, and <inherits> optionally contains any desired base classes.

Coconut allows data fields in <args> to have defaults and/or type annotations attached to them, and supports a starred parameter at the end to collect extra arguments.

Writing constructors for data types must be done using the __new__ method instead of the __init__ method. For helping to easily write __new__ methods, Coconut provides the makedata built-in.

Subclassing data types can be done easily by inheriting from them either in another data statement or a normal Python class. If a normal class statement is used, making the new subclass immutable will require adding the line

__slots__ = ()

which will need to be put in the subclass body before any method or attribute definitions.


A mainstay of functional programming that Coconut improves in Python is the use of values, or immutable data types. Immutable data can be very useful because it guarantees that once you have some data it won't change, but in Python creating custom immutable data types is difficult. Coconut makes it very easy by providing data blocks.

Python Docs

Returns a new tuple subclass. The new subclass is used to create tuple-like objects that have fields accessible by attribute lookup as well as being indexable and iterable. Instances of the subclass also have a helpful docstring (with type names and field names) and a helpful __repr__() method which lists the tuple contents in a name=value format.

Any valid Python identifier may be used for a field name except for names starting with an underscore. Valid identifiers consist of letters, digits, and underscores but do not start with a digit or underscore and cannot be a keyword such as class, for, return, global, pass, or raise.

Named tuple instances do not have per-instance dictionaries, so they are lightweight and require no more memory than regular tuples.



data vector2(x:int=0, y:int=0):
    def __abs__(self):
        return (self.x**2 + self.y**2)**.5

v = vector2(3, 4)
v |> print  # all data types come with a built-in __repr__
v |> abs |> print
v.x = 2  # this will fail because data objects are immutable
vector2() |> print

Showcases the syntax, features, and immutable nature of data types, as well as the use of default arguments and type annotations.

data Empty()
data Leaf(n)
data Node(l, r)

def size(Empty()) = 0

def size(Leaf(n)) = 1

def size(Node(l, r)) = size(l) + size(r)

size(Node(Empty(), Leaf(10))) == 1

Showcases the algebraic nature of data types when combined with pattern-matching.

data vector(*pts):
    """Immutable arbitrary-length vector."""

    def __abs__(self) =
        self.pts |> map$(pow$(?, 2)) |> sum |> pow$(?, 0.5)

    def __add__(self, other) =
        vector(*other_pts) = other
        assert len(other_pts) == len(self.pts)
        map((+), self.pts, other_pts) |*> vector

    def __neg__(self) =
        self.pts |> map$((-)) |*> vector

    def __sub__(self, other) =
        self + -other

Showcases starred data declaration.


import typing
class vector2(typing.NamedTuple("vector2", [("x", int), ("y", int)]), object):
    __slots__ = ()
    def __new__(cls, x=0, y=0):
        return super(vector2, cls).__new__((x, y))
    def __abs__(self):
        return (self.x**2 + self.y**2)**.5

v = vector2(3, 4)
v.x = 2
import collections
class Empty(collections.namedtuple("Empty", ""), object):
    __slots__ = ()
class Leaf(collections.namedtuple("Leaf", "n"), object):
    __slots__ = ()
class Node(collections.namedtuple("Node", "l, r"), object):
    __slots__ = ()

def size(tree):
    if isinstance(tree, Empty):
        return 0
    elif isinstance(tree, Leaf):
        return 1
    elif isinstance(tree, Node):
        return size(tree[0]) + size(tree[1])
        raise MatchError()

size(Node(Empty(), Leaf(10))) == 1

Starred data declarations can't be done without a long sequence of method definitions. See the compiled code for the Python syntax.


Coconut provides fully-featured, functional pattern-matching through its match statements.


Match statements follow the basic syntax match <pattern> in <value>. The match statement will attempt to match the value against the pattern, and if successful, bind any variables in the pattern to whatever is in the same position in the value, and execute the code below the match statement. Match statements also support, in their basic syntax, an if <cond> that will check the condition after executing the match before executing the code below, and an else statement afterwards that will only be executed if the match statement is not. What is allowed in the match statement's pattern has no equivalent in Python, and thus the specifications below are provided to explain it.

Syntax Specification

Coconut match statement syntax is

match <pattern> in <value> [if <cond>]:

where <value> is the item to match against, <cond> is an optional additional check, and <body> is simply code that is executed if the header above it succeeds. <pattern> follows its own, special syntax, defined roughly like so:

pattern ::= (
    "(" pattern ")"                 # parentheses
    | "None" | "True" | "False"     # constants
    | "=" NAME                      # check
    | NUMBER                        # numbers
    | STRING                        # stringsh
    | [pattern "as"] NAME           # capture
    | NAME "(" patterns ")"         # data types
    | pattern "is" exprs            # type-checking
    | pattern "and" pattern         # match all
    | pattern "or" pattern          # match any
    | "{" pattern_pairs             # dictionaries
        ["," **" NAME] "}"
    | ["s"] "{" pattern_consts "}"  # sets
    | "(" patterns ")"              # sequences can be in tuple form
    | "[" patterns "]"              #  or in list form
    | "(|" patterns "|)"            # lazy lists
    | ("(" | "[")                   # star splits
        "*" middle
      (")" | "]")
    | (                             # head-tail splits
        "(" patterns ")"
        | "[" patterns "]"
      ) "+" pattern
    | pattern "+" (                 # init-last splits
        "(" patterns ")"
        | "[" patterns "]"
    | (                             # head-last splits
        "(" patterns ")"
        | "[" patterns "]"
      ) "+" pattern "+" (
        "(" patterns ")"                # this match must be the same
        | "[" patterns "]"              #  construct as the first match
    | (                             # iterator splits
        "(" patterns ")"
        | "[" patterns "]"
        | "(|" patterns "|)"
      ) "::" pattern
    | ([STRING "+"] NAME            # complex string matching
        ["+" STRING])

Semantics Specification

match statements will take their pattern and attempt to "match" against it, performing the checks and deconstructions on the arguments as specified by the pattern. The different constructs that can be specified in a pattern, and their function, are:

  • Constants, Numbers, and Strings: will only match to the same constant, number, or string in the same position in the arguments.
  • Variables: will match to anything, and will be bound to whatever they match to, with some exceptions:
    • If the same variable is used multiple times, a check will be performed that each use match to the same value.
    • If the variable name _ is used, nothing will be bound and everything will always match to it.
  • Explicit Bindings (<pattern> as <var>): will bind <var> to <pattern>.
  • Checks (=<var>): will check that whatever is in that position is equal to the previously defined variable <var>.
  • Type Checks (<var> is <types>): will check that whatever is in that position is of type(s) <types> before binding the <var>.
  • Data Types (<name>(<args>)): will check that whatever is in that position is of data type <name> and will match the attributes to <args>.
  • Lists ([<patterns>]), Tuples ((<patterns>)): will only match a sequence ( of the same length, and will check the contents against <patterns>.
  • Lazy lists ((|<patterns>|)): same as list or tuple matching, but checks iterable ( instead of sequence.
  • Fixed-Length Dicts ({<pairs>}): will only match a mapping ( of the same length, and will check the contents against <pairs>.
  • Dicts With Rest ({<pairs>, **<rest>}): will match a mapping ( containing all the <pairs>, and will put a dict of everything else into <rest>.
  • Sets ({<constants>}): will only match a set ( of the same length and contents.
  • Head-Tail Splits (<list/tuple> + <var>): will match the beginning of the sequence against the <list/tuple>, then bind the rest to <var>, and make it the type of the construct used.
  • Init-Last Splits (<var> + <list/tuple>): exactly the same as head-tail splits, but on the end instead of the beginning of the sequence.
  • Head-Last Splits (<list/tuple> + <var> + <list/tuple>): the combination of a head-tail and an init-last split.
  • Iterator Splits (<list/tuple/lazy list> :: <var>): will match the beginning of an iterable ( against the <list/tuple/lazy list>, then bind the rest to <var> or check that the iterable is done.
  • Complex String Matching (<string> + <var> + <string>): matches strings that start and end with the given substrings, binding the middle to <var>.

Note: Like iterator slicing, iterator and lazy list matching make no guarantee that the original iterator matched against be preserved (to preserve the iterator, use Coconut's tee or reiterable built-in).

When checking whether or not an object can be matched against in a particular fashion, Coconut makes use of Python's abstract base classes. Therefore, to enable proper matching for a custom object, register it with the proper abstract base classes.



def factorial(value):
    match 0 in value:
        return 1
    else: match n is int in value if n > 0:  # possible because of Coconut's
        return n * factorial(n-1)            #   enhanced else statements
        raise TypeError("invalid argument to factorial of: "+repr(value))

3 |> factorial |> print

Showcases else statements, which work much like else statements in Python: the code under an else statement is only executed if the corresponding match fails.

data point(x, y):
    def transform(self, other):
        match point(x, y) in other:
            return point(self.x + x, self.y + y)
            raise TypeError("arg to transform must be a point")
    def __eq__(self, other):
        match point(=self.x, =self.y) in other:
            return True
            return False

point(1,2) |> point(3,4).transform |> print
point(1,2) |> (==)$(point(1,2)) |> print

Showcases matching to data types. Values defined by the user with the data statement can be matched against and their contents accessed by specifically referencing arguments to the data type's constructor.

data Empty()
data Leaf(n)
data Node(l, r)
Tree = (Empty, Leaf, Node)  # type union

def depth(Tree()) = 0

def depth(Tree(n)) = 1

def depth(Tree(l, r)) = 1 + max([depth(l), depth(r)])

Empty() |> depth |> print
Leaf(5) |> depth |> print
Node(Leaf(2), Node(Empty(), Leaf(3))) |> depth |> print

Showcases how the combination of data types and match statements can be used to powerful effect to replicate the usage of algebraic data types in other functional programming languages.

def duplicate_first([x] + xs as l) =
    [x] + l

[1,2,3] |> duplicate_first |> print

Showcases head-tail splitting, one of the most common uses of pattern-matching, where a + <var> (or :: <var> for any iterable) at the end of a list or tuple literal can be used to match the rest of the sequence.

def sieve([head] :: tail) =
    [head] :: sieve(n for n in tail if n % head)

def sieve((||)) = []

Showcases how to match against iterators, namely that the empty iterator case ((||)) must come last, otherwise that case will exhaust the whole iterator before any other pattern has a chance to match against it.

Python: Can't be done without a long series of checks for each match statement. See the compiled code for the Python syntax.


Coconut's case statement is an extension of Coconut's match statement for performing multiple match statements against the same value, where only one of them should succeed. Unlike lone match statements, only one match statement inside of a case block will ever succeed, and thus more general matches should be put below more specific ones.

Each pattern in a case block is checked until a match is found, and then the corresponding body is executed, and the case block terminated. The syntax for case blocks is

case <value>:
    match <pattern> [if <cond>]:
    match <pattern> [if <cond>]:

where <pattern> is any match pattern, <value> is the item to match against, <cond> is an optional additional check, and <body> is simply code that is executed if the header above it succeeds. Note the absence of an in in the match statements: that's because the <value> in case <value> is taking its place.



def classify_sequence(value):
    out = ""        # unlike with normal matches, only one of the patterns
    case value:     #  will match, and out will only get appended to once
        match ():
            out += "empty"
        match (_,):
            out += "singleton"
        match (x,x):
            out += "duplicate pair of "+str(x)
        match (_,_):
            out += "pair"
        match _ is (tuple, list):
            out += "sequence"
        raise TypeError()
    return out

[] |> classify_sequence |> print
() |> classify_sequence |> print
[1] |> classify_sequence |> print
(1,1) |> classify_sequence |> print
(1,2) |> classify_sequence |> print
(1,1,1) |> classify_sequence |> print

Python: Can't be done without a long series of checks for each match statement. See the compiled code for the Python syntax.


In Coconut, the keywords data, match, case, async (keyword in Python 3.5), and await (keyword in Python 3.5) are also valid variable names. While Coconut can disambiguate these two use cases, when using one of these keywords as a variable name, a backslash is allowed in front to be explicit about using a keyword as a variable name.



\data = 5


data = 5


Statement Lambdas

The statement lambda syntax is an extension of the normal lambda syntax to support statements, not just expressions.

The syntax for a statement lambda is

def (arguments) -> statement; statement; ...

where arguments can be standard function arguments or pattern-matching function definition arguments and statement can be an assignment statement or a keyword statement. If the last statement (not followed by a semicolon) is an expression, it will automatically be returned.

Statement lambdas also support implicit lambda syntax, where when the arguments are omitted, as in def -> _, def (_=None) -> _ is assumed.



L |> map$(def (x) -> y = 1/x; y*(1 - y))


def _lambda(x):
    y = 1/x
    return y*(1 - y)
map(_lambda, L)

Lazy Lists

Coconut supports the creation of lazy lists, where the contents in the list will be treated as an iterator and not evaluated until they are needed. Lazy lists can be created in Coconut simply by simply surrounding a comma-seperated list of items with (| and |) (so-called "banana brackets") instead of [ and ] for a list or ( and ) for a tuple.

Lazy lists use the same machinery as iterator chaining to make themselves lazy, and thus the lazy list (| x, y |) is equivalent to the iterator chaining expression (x,) :: (y,), although the lazy list won't construct the intermediate tuples.


Lazy lists, where sequences are only evaluated when their contents are requested, are a mainstay of functional programming, allowing for dynamic evaluation of the list's contents.



(| print("hello,"), print("world!") |) |> consume

Python: Can't be done without a complicated iterator comprehension in place of the lazy list. See the compiled code for the Python syntax.

Implicit Partial Application

Coconut supports a number of different syntactical aliases for common partial application use cases. These are:

.attr           =>      operator.attrgetter("attr")
.method(args)   =>      operator.methodcaller("method", args)
obj.            =>      getattr$(obj)
func$           =>      ($)$(func)
seq[]           =>      operator.getitem$(seq)
iter$[]         =>      # the equivalent of seq[] for iterators
.[a:b:c]        =>      operator.itemgetter(slice(a, b, c))
.$[a:b:c]       =>      # the equivalent of .[a:b:c] for iterators



1 |> "123"[]
mod$ <| 5 <| 3


mod(5, 3)

Enhanced Type Annotations

Since Coconut syntax is a superset of Python 3 syntax, it supports Python 3 function type annotation syntax and Python 3.6 variable type annotation syntax. By default, Coconut compiles all type annotations into Python-2-compatible type comments. If you want to keep the type annotations instead, simply pass a --target that supports them.

Note: When compiling type annotations to Python 3 syntax, Coconut will wrap every annotation in a string when in a position where Python would otherwise evaluate it (Python 3 function annotation), so that all type annotations are only ever evaluated at compile time, never at run time.

Additionally, Coconut adds special syntax for making type annotations easier and simpler to write. When inside of a type annotation, Coconut treats certain syntax constructs differently, compiling them to type annotations instead of what they would normally represent. Specifically, Coconut applies the following transformations

    => typing.Optional[<type>]
    => typing.Sequence[<type>]
    => typing.Iterable[<type>]
() -> <ret>
    => typing.Callable[[], <ret>]
<arg> -> <ret>
    => typing.Callable[[<arg>], <ret>]
(<args>) -> <ret>
    => typing.Callable[[<args>], <ret>]
-> <ret>
    => typing.Callable[..., <ret>]

where typing is the Python 3.5 built-in typing module.



def int_map(
    f: int -> int,
    xs: int[],
) -> int[] =
    xs |> map$(f) |> list


import typing  # unlike this typing import, Coconut produces universal code
def int_map(
    f,  # type: typing.Callable[[int], int]
    xs,  # type: typing.Sequence[int]
    # type: (...) -> typing.Sequence[int]
    return list(map(f, xs))

Operator Functions

Coconut uses a simple operator function short-hand: surround an operator with parentheses to retrieve its function. Similarly to iterator comprehensions, if the operator function is the only argument to a function, the parentheses of the function call can also serve as the parentheses for the operator function.


A very common thing to do in functional programming is to make use of function versions of built-in operators: currying them, composing them, and piping them. To make this easy, Coconut provides a short-hand syntax to access operator functions.

Full List

(|>)        => # pipe forward
(|*>)       => # multi-arg pipe forward
(<|)        => # pipe backward
(<*|)       => # multi-arg pipe backward
(..), (<..) => # backward function composition
(..>)       => # forward function composition
(<*..)      => # multi-arg backward function composition
(..*>)      => # multi-arg forward function composition
(.)         => (getattr)
(::)        => (itertools.chain)  # will not evaluate its arguments lazily
($)         => (functools.partial)
($[])       => # iterator slicing operator
(+)         => (operator.add)
(-)         => # 1 arg: operator.neg, 2 args: operator.sub
(*)         => (operator.mul)
(**)        => (operator.pow)
(/)         => (operator.truediv)
(//)        => (operator.floordiv)
(%)         => (operator.mod)
(&)         => (operator.and_)
(^)         => (operator.xor)
(|)         => (operator.or_)
(<<)        => (operator.lshift)
(>>)        => (operator.rshift)
(<)         => (
(>)         => (
(==)        => (operator.eq)
(<=)        => (operator.le)
(>=)        => (
(!=)        => (
(~)         => (operator.inv)
(@)         => (operator.matmul)
(not)       => (operator.not_)
(and)       => # boolean and
(or)        => # boolean or
(is)        => (operator.is_)
(in)        => (operator.contains)



(range(0, 5), range(5, 10)) |*> map$(+) |> list |> print


import operator
print(list(map(operator.add, range(0, 5), range(5, 10))))

Set Literals

Coconut allows an optional s to be prepended in front of Python set literals. While in most cases this does nothing, in the case of the empty set it lets Coconut know that it is an empty set and not an empty dictionary. Additionally, an f is also supported, in which case a Python frozenset will be generated instead of a normal set.



empty_frozen_set = f{}


empty_frozen_set = frozenset()

Imaginary Literals

In addition to Python's <num>j or <num>J notation for imaginary literals, Coconut also supports <num>i or <num>I, to make imaginary literals more readable if used in a mathematical context.

Python Docs

Imaginary literals are described by the following lexical definitions:

imagnumber ::= (floatnumber | intpart) ("j" | "J" | "i" | "I")

An imaginary literal yields a complex number with a real part of 0.0. Complex numbers are represented as a pair of floating point numbers and have the same restrictions on their range. To create a complex number with a nonzero real part, add a floating point number to it, e.g., (3+4i). Some examples of imaginary literals:

3.14i   10.i    10i     .001i   1e100i  3.14e-10i



3 + 4i |> abs |> print


print(abs(3 + 4j))

Function Definition

Tail Call Optimization

Coconut will perform automatic tail call optimization and tail recursion elimination on any function that meets the following criteria:

  1. it must directly return (using either return or assignment function notation) a call to itself (tail recursion elimination, the most powerful optimization) or another function (tail call optimization).
  2. it must not be a generator (uses yield) or an asynchronous function (uses async).

Note: Tail call optimization (though not tail recursion elimination) will work even for 1) mutual recursion and 2) pattern-matching functions split across multiple definitions using addpattern.

If you are encountering a RuntimeError due to maximum recursion depth, it is highly recommended that you rewrite your function to meet either the criteria above for tail call optimization, or the corresponding criteria for recursive_iterator, either of which should prevent such errors.

Note: Tail call optimization (though not tail recursion elimination) will be turned off if you pass the --no-tco command-line option, which is useful if you are having trouble reading your tracebacks and/or need maximum performance.



# unlike in Python, this function will never hit a maximum recursion depth error
def factorial(n, acc=1):
    case n:
        match 0:
            return acc
        match _ is int if n > 0:
            return factorial(n-1, acc*n)

Showcases tail recursion elimination.

# unlike in Python, neither of these functions will ever hit a maximum recursion depth error
def is_even(0) = True
def is_even(n is int if n > 0) = is_odd(n-1)

def is_odd(0) = False
def is_odd(n is int if n > 0) = is_even(n-1)

Showcases tail call optimization.

Python: Can't be done without rewriting the function(s).

Assignment Functions

Coconut allows for assignment function definition that automatically returns the last line of the function body. An assignment function is constructed by substituting = for : after the function definition line. Thus, the syntax for assignment function definition is either

def <name>(<args>) = <expr>

for one-liners or

def <name>(<args>) =

for full functions, where <name> is the name of the function, <args> are the functions arguments, <stmts> are any statements that the function should execute, and <expr> is the value that the function should return.

Note: Assignment function definition can be combined with infix and/or pattern-matching function definition.


Coconut's Assignment function definition is as easy to write as assignment to a lambda, but will appear named in tracebacks, as it compiles to normal Python function definition.



def binexp(x) = 2**x
5 |> binexp |> print


def binexp(x): return 2**x

Pattern-Matching Functions

Coconut supports pattern-matching on the arguments to a function in that function's definition. The syntax for pattern-matching function definition is

[match] def <name>(<pattern> [= <default>], ... [if <cond>]):

where <name> is the name of the function, <cond> is an optional additional check, <body> is the body of the function, <pattern> is defined by Coconut's match statement, and <default> is the optional default if no argument is passed. The match keyword at the beginning is optional, but is sometimes necessary to disambiguate pattern-matching function definition from normal function definition, which will always take precedence.

If <pattern> has a variable name (either directly or with as), the resulting pattern-matching function will support keyword arguments using that variable name. If pattern-matching function definition fails, it will raise a MatchError object just like destructuring assignment.

Note: Pattern-matching function definition can be combined with assignment and/or infix function definition.



def last_two(_ + [a, b]):
    return a, b
def xydict_to_xytuple({"x":x is int, "y":y is int}):
    return x, y

range(5) |> last_two |> print
{"x":1, "y":2} |> xydict_to_xytuple |> print

Python: Can't be done without a long series of checks at the top of the function. See the compiled code for the Python syntax.

Infix Functions

Coconut allows for infix function calling, where an expression that evaluates to a function is surrounded by backticks and then can have arguments placed in front of or behind it. Infix calling has a precedence in-between chaining and None-coalescing, and is left-associative.

Coconut also supports infix function definition to make defining functions that are intended for infix usage simpler. The syntax for infix function definition is

def <arg> `<name>` <arg>:

where <name> is the name of the function, the <arg>s are the function arguments, and <body> is the body of the function. If an <arg> includes a default, the <arg> must be surrounded in parentheses.

Note: Infix function definition can be combined with assignment and/or pattern-matching function definition.


A common idiom in functional programming is to write functions that are intended to behave somewhat like operators, and to call and define them by placing them between their arguments. Coconut's infix syntax makes this possible.



def a `mod` b = a % b
(x `mod` 2) `print`


def mod(a, b): return a % b
print(mod(x, 2))

Dotted Function Definition

Coconut allows for function definition using a dotted name to assign a function as a method of an object as specified in PEP 542.



def MyClass.my_method(self):


def my_method(self):
MyClass.my_method = my_method


Destructuring Assignment

Coconut supports significantly enhanced destructuring assignment, similar to Python's tuple/list destructuring, but much more powerful. The syntax for Coconut's destructuring assignment is

[match] <pattern> = <value>

where <value> is any expression and <pattern> is defined by Coconut's match statement. The match keyword at the beginning is optional, but is sometimes necessary to disambiguate destructuring assignment from normal assignment, which will always take precedence. Coconut's destructuring assignment is equivalent to a match statement that follows the syntax:

match <pattern> in <value>:
    err = MatchError(<error message>)
    err.pattern = "<pattern>"
    err.value = <value>
    raise err

If a destructuring assignment statement fails, then instead of continuing on as if a match block had failed, a MatchError object will be raised describing the failure.



_ + [a, b] = [0, 1, 2, 3]
print(a, b)

Python: Can't be done without a long series of checks in place of the destructuring assignment statement. See the compiled code for the Python syntax.


Unlike Python, which only supports a single variable or function call in a decorator, Coconut supports any expression.



@ wrapper1 .. wrapper2 $(arg)
def func(x) = x**2


def wrapper(func):
    return wrapper1(wrapper2(arg, func))
def func(x):
    return x**2

else Statements

Coconut supports the compound statements try, if, and match on the end of an else statement like any simple statement would be. This is most useful for mixing match and if statements together, but also allows for compound try statements.



if invalid(input_list):
    raise Exception()
else: match [head] + tail in input_list:
    print(head, tail)


from import Sequence
if invalid(input_list):
    raise Exception()
elif isinstance(input_list, Sequence):
    head, tail = inputlist[0], inputlist[1:]
    print(head, tail)

except Statements

Python 3 requires that if multiple exceptions are to be caught, they must be placed inside of parentheses, so as to disallow Python 2's use of a comma instead of as. Coconut allows commas in except statements to translate to catching multiple exceptions without the need for parentheses, since, as in Python 3, as is always required to bind the exception to a name.



except SyntaxError, ValueError as err:


except (SyntaxError, ValueError) as err:

Implicit pass

Coconut supports the simple class name(base) and data name(args) as aliases for class name(base): pass and data name(args): pass.



data Empty
data Leaf(item)
data Node(left, right)


import collections
class Empty(collections.namedtuple("Empty", ""), object):
    __slots__ = ()
class Leaf(collections.namedtuple("Leaf", "n"), object):
    __slots__ = ()
class Node(collections.namedtuple("Node", "l, r"), object):
    __slots__ = ()

In-line global And nonlocal Assignment

Coconut allows for global or nonlocal to precede assignment to a variable or list of variables to make that assignment global or nonlocal, respectively.



global state_a, state_b = 10, 100


global state_a, state_b; state_a, state_b = 10, 100

Code Passthrough

Coconut supports the ability to pass arbitrary code through the compiler without being touched, for compatibility with other variants of Python, such as Cython or Mython. Anything placed between \( and the corresponding close parenthesis will be passed through, as well as any line starting with \\, which will have the additional effect of allowing indentation under it.



\\cdef f(x):
    return x |> g


cdef f(x):
    return g(x)

Enhanced Parenthetical Continuation

Since Coconut syntax is a superset of Python 3 syntax, Coconut supports the same line continuation syntax as Python. That means both backslash line continuation and implied line continuation inside of parentheses, brackets, or braces will all work.

In Python, however, there are some cases (such as multiple with statements) where only backslash continuation, and not parenthetical continuation, is supported. Coconut adds support for parenthetical continuation in all these cases.

Supporting parenthetical continuation everywhere allows the PEP 8 convention, which avoids backslash continuation in favor of implied parenthetical continuation, to always be possible to follow. From PEP 8:

The preferred way of wrapping long lines is by using Python's implied line continuation inside parentheses, brackets and braces. Long lines can be broken over multiple lines by wrapping expressions in parentheses. These should be used in preference to using a backslash for line continuation.

Note: Passing --strict will enforce the PEP 8 convention by disallowing backslash continuations.



with (open('/path/to/some/file/you/want/to/read') as file_1,
      open('/path/to/some/file/being/written', 'w') as file_2):


# split into two with statements for Python 2.6 compatibility
with open('/path/to/some/file/you/want/to/read') as file_1:
    with open('/path/to/some/file/being/written', 'w') as file_2:


Enhanced Built-Ins

Coconut's map, zip, filter, reversed, and enumerate objects are enhanced versions of their Python equivalents that support reversed, repr, optimized normal (and iterator) slicing (all but filter), len (all but filter), and have added attributes which subclasses can make use of to get at the original arguments to the object:

  • map: _func, _iters
  • zip: _iters
  • filter: _func, _iter
  • reversed: _iter
  • enumerate: _iter, _start



map((+), range(5), range(6)) |> len |> print
range(10) |> filter$((x) -> x < 5) |> reversed |> tuple |> print

Python: Can't be done without defining a custom map type. The full definition of map can be found in the Coconut header.


Takes one argument that is a pattern-matching function, and returns a decorator that adds the patterns in the existing function to the new function being decorated, where the existing patterns are checked first, then the new. Roughly equivalent to:

def addpattern(base_func):
    """Decorator to add a new case to a pattern-matching function, where the new case is checked last."""
    def pattern_adder(func):
        def add_pattern_func(*args, **kwargs):
                return base_func(*args, **kwargs)
            except MatchError:
                return func(*args, **kwargs)
        return add_pattern_func
    return pattern_adder

DEPRECATED: Coconut also has a prepattern built-in, which adds patterns in the opposite order of addpattern; prepattern is defined as:

def prepattern(base_func):
    """Decorator to add a new case to a pattern-matching function,
    where the new case is checked first."""
    def pattern_prepender(func):
        return addpattern(func)(base_func)
    return pattern_prepender

Note: Passing --strict disables deprecated features.



def factorial(0) = 1

def factorial(n) = n * factorial(n - 1)

Python: Can't be done without a complicated decorator definition and a long series of checks for each pattern-matching. See the compiled code for the Python syntax.


Coconut re-introduces Python 2's reduce built-in, using the functools.reduce version.

Python Docs

reduce(function, iterable[, initializer])

Apply function of two arguments cumulatively to the items of sequence, from left to right, so as to reduce the sequence to a single value. For example, reduce((x, y) -> x+y, [1, 2, 3, 4, 5]) calculates ((((1+2)+3)+4)+5). The left argument, x, is the accumulated value and the right argument, y, is the update value from the sequence. If the optional initializer is present, it is placed before the items of the sequence in the calculation, and serves as a default when the sequence is empty. If initializer is not given and sequence contains only one item, the first item is returned.



product = reduce$(*)
range(1, 10) |> product |> print


import operator
import functools
product = functools.partial(functools.reduce, operator.mul)
print(product(range(1, 10)))


Coconut provides itertools.takewhile as a built-in under the name takewhile.

Python Docs

takewhile(predicate, iterable)

Make an iterator that returns elements from the iterable as long as the predicate is true. Equivalent to:

def takewhile(predicate, iterable):
    # takewhile(lambda x: x<5, [1,4,6,4,1]) --> 1 4
    for x in iterable:
        if predicate(x):
            yield x



negatives = takewhile(numiter, (x) -> x<0)


import itertools
negatives = itertools.takewhile(numiter, lambda x: x<0)


Coconut provides itertools.dropwhile as a built-in under the name dropwhile.

Python Docs

dropwhile(predicate, iterable)

Make an iterator that drops elements from the iterable as long as the predicate is true; afterwards, returns every element. Note: the iterator does not produce any output until the predicate first becomes false, so it may have a lengthy start-up time. Equivalent to:

def dropwhile(predicate, iterable):
    # dropwhile(lambda x: x<5, [1,4,6,4,1]) --> 6 4 1
    iterable = iter(iterable)
    for x in iterable:
        if not predicate(x):
            yield x
    for x in iterable:
        yield x



positives = dropwhile(numiter, (x) -> x<0)


import itertools
positives = itertools.dropwhile(numiter, lambda x: x<0)


Coconut provides the groupsof built-in to split an iterable into groups of a specific length. Specifically, groupsof(n, iterable) will split iterable into tuples of length n, with only the last tuple potentially of size < n if the length of iterable is not divisible by n.



pairs = range(1, 11) |> groupsof$(2)


pairs = []
group = []
for item in range(1, 11):
    if len(group) == 2:
        group = []
if group:


Coconut provides an optimized version of itertools.tee as a built-in under the name tee.

Python Docs

tee(iterable, n=2)

Return n independent iterators from a single iterable. Equivalent to:

def tee(iterable, n=2):
    it = iter(iterable)
    deques = [collections.deque() for i in range(n)]
    def gen(mydeque):
        while True:
            if not mydeque:             # when the local deque is empty
                newval = next(it)       # fetch a new value and
                for d in deques:        # load it to all the deques
            yield mydeque.popleft()
    return tuple(gen(d) for d in deques)

Once tee() has made a split, the original iterable should not be used anywhere else; otherwise, the iterable could get advanced without the tee objects being informed.

This itertool may require significant auxiliary storage (depending on how much temporary data needs to be stored). In general, if one iterator uses most or all of the data before another iterator starts, it is faster to use list() instead of tee().



original, temp = tee(original)
sliced = temp$[5:]


import itertools
original, temp = itertools.tee(original)
sliced = itertools.islice(temp, 5, None)


Sometimes, when an iterator may need to be iterated over an arbitrary number of times, tee can be cumbersome to use. For such cases, Coconut provides reiterable, which wraps the given iterator such that whenever an attempt to iterate over it is made, it iterates over a tee instead of the original.



def list_type(xs):
    case reiterable(xs):
        match [fst, snd] :: tail:
            return "at least 2"
        match [fst] :: tail:
            return "at least 1"
        match (| |):
            return "empty"

Python: Can't be done without a long series of checks for each match statement. See the compiled code for the Python syntax.


Coconut provides the consume function to efficiently exhaust an iterator and thus perform any lazy evaluation contained within it. consume takes one optional argument, keep_last, that defaults to 0 and specifies how many, if any, items from the end to return as an iterable (None will keep all elements).

Equivalent to:

def consume(iterable, keep_last=0):
    """Fully exhaust iterable and return the last keep_last elements."""
    return collections.deque(iterable, maxlen=keep_last)  # fastest way to exhaust an iterator


In the process of lazily applying operations to iterators, eventually a point is reached where evaluation of the iterator is necessary. To do this efficiently, Coconut provides the consume function, which will fully exhaust the iterator given to it.



range(10) |> map$((x) -> x**2) |> map$(print) |> consume


collections.deque(map(print, map(lambda x: x**2, range(10))), maxlen=0)


Coconut provides a modified version of itertools.count that supports in, normal slicing, optimized iterator slicing, count and index sequence methods, repr, and start and step attributes as a built-in under the name count.

Python Docs

count(start=0, step=1)

Make an iterator that returns evenly spaced values starting with number start. Often used as an argument to map() to generate consecutive data points. Also, used with zip() to add sequence numbers. Roughly equivalent to:

def count(start=0, step=1):
    # count(10) --> 10 11 12 13 14 ...
    # count(2.5, 0.5) -> 2.5 3.0 3.5 ...
    n = start
    while True:
        yield n
        n += step



count()$[10**100] |> print

Python: Can't be done quickly without Coconut's iterator slicing, which requires many complicated pieces. The necessary definitions in Python can be found in the Coconut header.


Coconut provides the makedata function to allow direct access to the base constructor of data types created with the Coconut data statement. This is particularly useful when writing alternative constructors for data types by overwriting __new__.

makedata takes the data type to call as the first argument, and the arguments for constructing that data type as the rest of the arguments. For data objects, makedata behaves as the underlying data type's original constructor, exactly as the data type was declared. For non-data objects, makedata is equivalent to:

def makedata(data_type, *args, **kwargs):
    """Returns base data constructor of data_type."""
    return super(data_type, data_type).__new__(data_type, *args, **kwargs)

DEPRECATED: Coconut also has a datamaker built-in, which partially applies makedata; datamaker is defined as:

def datamaker(data_type):
    """Get the original constructor of the given data type or class."""
    return makedata$(data_type)

Note: Passing --strict disables deprecated features.



data Tuple(elems):
    def __new__(cls, *elems):
        return elems |> makedata$(cls)


import collections
class Tuple(collections.namedtuple("Tuple", "elems"), object):
    __slots__ = ()
    def __new__(cls, *elems):
        return super(cls, cls).__new__(cls, elems)


In functional programming, fmap(func, obj) takes a data type obj and returns a new data type with func mapped over the contents. Coconut's fmap function does the exact same thing in Coconut.

fmap can also be used on built-ins such as str, list, set, and dict as a variant of map that returns back an object of the same type. The behavior of fmap for a given object can be overridden by defining an __fmap__(self, func) method that will be called whenever fmap is invoked on that object.

For dict, or any other, fmap will be called on the mapping's .items() instead of the default iteration through its .keys().



[1, 2, 3] |> fmap$(x -> x+1) == [2, 3, 4]

data Nothing()
data Just(n)

Just(3) |> fmap$(x -> x*2) == Just(6)
Nothing() |> fmap$(x -> x*2) == Nothing()


list(map(lambda x: x+1, [1, 2, 3])) == [2, 3, 4]

import collections
class Nothing(collections.namedtuple("Nothing", ""), object):
    __slots__ = ()
class Just(collections.namedtuple("Just", "n"), object):
    __slots__ = ()

Just(*map(lambda x: x*2, Just(3))) == Just(6)
Nothing(*map(lambda x: x*2, Nothing())) == Nothing()


Coconut provides a modified version of itertools.starmap that supports reversed, repr, optimized normal (and iterator) slicing, len, and _func and _iter attributes.

Python Docs

starmap(function, iterable)

Make an iterator that computes the function using arguments obtained from the iterable. Used instead of map() when argument parameters are already grouped in tuples from a single iterable (the data has been "pre-zipped"). The difference between map() and starmap() parallels the distinction between function(a,b) and function(*c). Roughly equivalent to:

def starmap(function, iterable):
    # starmap(pow, [(2,5), (3,2), (10,3)]) --> 32 9 1000
    for args in iterable:
        yield function(*args)



range(1, 5) |> map$(range) |> starmap$(print) |> consume


import itertools, collections
collections.deque(itertools.starmap(print, map(range, range(1, 5))), maxlen=0)


Coconut provides a modified version of itertools.accumulate with opposite argument order as scan that also supports repr, len, and func and iter attributes. scan works exactly like reduce, except that instead of only returning the last accumulated value, it returns an iterator of all the intermediate values.

Python Docs

scan(func, iterable)

Make an iterator that returns accumulated results of some function of two arguments. Elements of the input iterable may be any type that can be accepted as arguments to func. (For example, with the operation of addition, elements may be any addable type including Decimal or Fraction.) If the input iterable is empty, the output iterable will also be empty.

Roughly equivalent to:

def scan(func, iterable):
    'Return running totals'
    # scan(operator.add, [1,2,3,4,5]) --> 1 3 6 10 15
    # scan(operator.mul, [1,2,3,4,5]) --> 1 2 6 24 120
    it = iter(iterable)
        total = next(it)
    except StopIteration:
    yield total
    for element in it:
        total = func(total, element)
        yield total



input_data = [3, 4, 6, 2, 1, 9, 0, 7, 5, 8]
running_max = input_data |> scan$(max) |> list


input_data = [3, 4, 6, 2, 1, 9, 0, 7, 5, 8]
running_max = []
max_so_far = input_data[0]
for x in input_data:
    if x > max_so_far:
        max_so_far = x


Coconut provides a recursive_iterator decorator that provides significant optimizations for any stateless, recursive function that returns an iterator. To use recursive_iterator on a function, it must meet the following criteria:

  1. your function either always returns an iterator or generates an iterator using yield,
  2. when called multiple times with the same arguments, your function produces the same iterator (your function is stateless), and
  3. your function gets called (usually calls itself) multiple times with the same arguments.

If you are encountering a RuntimeError due to maximum recursion depth, it is highly recommended that you rewrite your function to meet either the criteria above for recursive_iterator, or the corresponding criteria for Coconut's tail call optimization, either of which should prevent such errors.

Furthermore, recursive_iterator also allows the resolution of a nasty segmentation fault in Python's iterator logic that has never been fixed. Specifically, instead of writing

seq = get_elem() :: seq

which will crash due to the aforementioned Python issue, write

def seq() = get_elem() :: seq()

which will work just fine.



def fib() = (1, 1) :: map((+), fib(), fib()$[1:])

Python: Can't be done without a long decorator definition. The full definition of the decorator in Python can be found in the Coconut header.


Coconut provides a parallel version of map under the name parallel_map. parallel_map makes use of multiple processes, and is therefore much faster than map for CPU-bound tasks. Use of parallel_map requires concurrent.futures, which exists in the Python 3 standard library, but under Python 2 will require pip install futures to function.

Because parallel_map uses multiple processes for its execution, it is necessary that all of its arguments be pickleable. Only objects defined at the module level, and not lambdas, objects defined inside of a function, or objects defined inside of the interpreter, are pickleable. Furthermore, on Windows, it is necessary that all calls to parallel_map occur inside of an if __name__ == "__main__" guard.

Python Docs

parallel_map(func, *iterables)

Equivalent to map(func, *iterables) except func is executed asynchronously and several calls to func may be made concurrently. If a call raises an exception, then that exception will be raised when its value is retrieved from the iterator.



parallel_map(pow$(2), range(100)) |> list |> print


import functools
import concurrent.futures
with concurrent.futures.ProcessPoolExecutor() as executor:
    print(list(, 2), range(100))))


Coconut provides a concurrent version of map under the name concurrent_map. concurrent_map makes use of multiple threads, and is therefore much faster than map for IO-bound tasks. Use of concurrent_map requires concurrent.futures, which exists in the Python 3 standard library, but under Python 2 will require pip install futures to function.

Python Docs

concurrent_map(func, *iterables)

Equivalent to map(func, *iterables) except func is executed asynchronously and several calls to func may be made concurrently. If a call raises an exception, then that exception will be raised when its value is retrieved from the iterator.



concurrent_map(get_data_for_user, get_all_users()) |> list |> print


import functools
import concurrent.futures
with concurrent.futures.ThreadPoolExecutor() as executor:
    print(list(, get_all_users())))


A MatchError is raised when a destructuring assignment statement fails, and thus MatchError is provided as a built-in for catching those errors. MatchError objects support two attributes, pattern, which is a string describing the failed pattern, and value, which is the object that failed to match that pattern.

Coconut Modules


It is sometimes useful to be able to access Coconut built-ins from pure Python. To accomplish this, Coconut provides coconut.__coconut__, which behaves exactly like the header file included when Coconut is compiled in package mode.

All Coconut built-ins are accessible from coconut.__coconut__. The recommended way to import them is to use from coconut.__coconut__ import and import whatever built-ins you'll be using.


from coconut.__coconut__ import parallel_map


It is sometimes useful to be able to use the Coconut compiler from code, instead of from the command line. The recommended way to do this is to use from coconut.convenience import and import whatever convenience functions you'll be using. Specifications of the different convenience functions are as follows.


coconut.convenience.parse([code, [mode]])

Likely the most useful of the convenience functions, parse takes Coconut code as input and outputs the equivalent compiled Python code. The second argument, mode, is used to indicate the context for the parsing.

If code is not passed, parse will output just the given mode's header, which can be executed to set up an execution environment in which future code can be parsed and executed without a header.

Each mode has two components: what parser it uses, and what header it prepends. The parser determines what Coconut code is allowed as input, and the header determines how the compiled Python can be used. Possible values of mode are:

  • "sys": (the default)
    • parser: file
      • The file parser can parse any Coconut code.
    • header: sys
      • This header imports coconut.__coconut__ to access the necessary Coconut objects.
  • "exec":
    • parser: file
    • header: exec
      • When passed to exec at the global level, this header will create all the necessary Coconut objects itself instead of importing them.
  • "file":
    • parser: file
    • header: file
      • This header is meant to be written to a --standalone file and should not be passed to exec.
  • "package":
    • parser: file
    • header: package
      • This header is meant to be written to a --package file and should not be passed to exec.
  • "block":
    • parser: file
    • header: none
      • No header is included, thus this can only be passed to exec if code with a header has already been executed at the global level.
  • "single":
    • parser: single
      • Can only parse one line of Coconut code.
    • header: none
  • "eval":
    • parser: eval
      • Can only parse a Coconut expression, not a statement.
    • header: none
  • "debug":
    • parser: debug
      • Can parse any Coconut code, allows leading whitespace, and has no trailing newline.
    • header: none
from coconut.convenience import parse
while True:
    exec(parse(input(), mode="block"))


coconut.convenience.setup(target, strict, minify, line_numbers, keep_lines, no_tco)

setup can be used to pass command line flags for use in parse. The possible values for each flag argument are:

  • target: None (default), or any allowable target
  • strict: False (default) or True
  • minify: False (default) or True
  • line_numbers: False (default) or True
  • keep_lines: False (default) or True
  • no_tco: False (default) or True


coconut.convenience.cmd(args, [interact])

Executes the given args as if they were fed to coconut on the command-line, with the exception that unless interact is true or -i is passed, the interpreter will not be started. Additionally, since parse and cmd share the same convenience parsing object, any changes made to the parsing with cmd will work just as if they were made with setup.



Retrieves a string containing information about the Coconut version. The optional argument which is the type of version information desired. Possible values of which are:

  • "num": the numerical version (the default)
  • "name": the version codename
  • "spec": the numerical version with the codename attached
  • "tag": the version tag used in GitHub and documentation URLs
  • "-v": the full string printed by coconut -v


If an error is encountered in a convenience function, a CoconutException instance may be raised. coconut.convenience.CoconutException is provided to allow catching such errors.