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 an introduction to 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 the latest 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.

Thought Coconut syntax is primarily based on that of Python, other languages that inspired Coconut include Haskell, CoffeeScript, F#, and Julia.

Try It Out#

If you want to try Coconut in your browser, check out the online interpreter. Note, however, that it may be running an outdated version of Coconut.


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 adding --user or running

pip install --no-deps --upgrade coconut "pyparsing<3"

which will force Coconut to use the pure-Python pyparsing module instead of the faster cPyparsing module. If you are still getting errors, you may want to try using conda instead.

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: Coconut’s conda recipe uses pyparsing rather than cPyparsing, which may lead to degraded performance relative to installing Coconut via pip.

Using Homebrew#

If you prefer to use Homebrew, you can also install Coconut using brew:

brew install coconut

Note: Coconut’s Homebrew formula may not always be up-to-date with the latest version of Coconut.

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 coconut[opt_dep_1,opt_dep_2]

The full list of optional dependencies is:

  • all: alias for everything below (this is the recommended way to install a feature-complete version of Coconut).

  • jupyter/ipython: enables use of the --jupyter / --ipython flag.

  • kernel: lightweight subset of jupyter that only includes the dependencies that are strictly necessary for Coconut’s Jupyter kernel.

  • watch: enables use of the --watch flag.

  • mypy: enables use of the --mypy flag.

  • xonsh: enables use of Coconut’s xonsh support.

  • numpy: installs everything necessary for making use of Coconut’s numpy integration.

  • jupyterlab: installs everything necessary to use JupyterLab with Coconut.

  • jupytext: installs everything necessary to use Jupytext with Coconut.

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.

Note: if you have an existing release version of coconut installed, you’ll need to pip uninstall coconut before installing coconut-develop.



coconut [-h] [--and source [dest ...]] [-v] [-t version] [-i] [-p] [-a] [-l]
        [--no-line-numbers] [-k] [-w] [-r] [-n] [-d] [-q] [-s] [--no-tco]
        [--no-wrap-types] [-c code] [--incremental] [-j processes] [-f] [--minify]
        [--jupyter ...] [--mypy ...] [--argv ...] [--tutorial] [--docs] [--style name]
        [--vi-mode] [--recursion-limit limit] [--stack-size kbs] [--site-install]
        [--site-uninstall] [--verbose] [--trace] [--profile]
        [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
--and source [dest ...]
                      add an additional source/dest pair to compile (dest is optional)
-v, -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
-a, --standalone, --stand-alone
                      compile source as standalone files (defaults to only if source is a single
-l, --line-numbers, --linenumbers
                      force enable line number comments (--line-numbers are enabled by default
                      unless --minify is passed)
--no-line-numbers, --nolinenumbers
                      disable line number comments (opposite of --line-numbers)
-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
--no-wrap-types, --nowraptypes
                      disable wrapping type annotations in strings and turn off 'from __future__
                      import annotations' behavior
-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 'sys') (0 is no
                      additional processes; 'sys' uses machine default)
-f, --force           force re-compilation even when source code and compilation parameters
                      haven't changed
--minify              reduce size of compiled Python
--jupyter ..., --ipython ...
                      run Jupyter/IPython with Coconut as the kernel (remaining args passed to
--mypy ...            run MyPy on compiled Python (remaining args passed to MyPy) (implies
                      --package --line-numbers)
--argv ..., --args ...
                      set sys.argv to source plus remaining args for use in the Coconut script
                      being run
--tutorial            open Coconut's tutorial in the default web browser
--docs, --documentation
                      open Coconut's documentation in the default web browser
--style name          set Pygments syntax highlighting style (or 'list' to list styles)
                      (defaults to COCONUT_STYLE environment variable if it exists, otherwise
--vi-mode, --vimode   enable vi mode in the interpreter (currently set to False) (can be
                      modified by setting COCONUT_VI_MODE environment variable)
--recursion-limit limit, --recursionlimit limit
                      set maximum recursion depth in compiler (defaults to 1920) (when
                      increasing --recursion-limit, you may also need to increase --stack-size;
                      setting them to approximately equal values is recommended)
--stack-size kbs, --stacksize kbs
                      run the compiler in a separate thread with the given stack size in
--fail-fast           causes the compiler to fail immediately upon encountering a compilation
                      error rather than attempting to continue compiling other files
--no-cache            disables use of Coconut's incremental parsing cache (caches previous
                      parses to improve recompilation performance for slightly modified files)
--site-install, --siteinstall
                      set up coconut.api to be imported on Python start
--site-uninstall, --siteuninstall
                      revert the effects of --site-install
--verbose             print verbose debug output
--trace               print verbose parsing data (only available in coconut-develop)
--profile             collect and print timing info (only available in coconut-develop)

Coconut Scripts#

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

coconut-run <source> <args>

as an alias for

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

which will quietly compile and run <source>, passing any additional arguments to the script, mimicking how the python command works. To instead pass additional compilation arguments to Coconut itself (e.g. --no-tco), put them before the <source> file.

coconut-run can be used to compile and run directories rather than files, again mimicking how the python command works. Specifically, Coconut will compile the directory and then run the __main__.coco in that directory, which must exist.

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

coconut-run will always enable automatic compilation, such that Coconut source files can be directly imported from any Coconut files run via coconut-run. Additionally, compilation parameters (e.g. --no-tco) used in coconut-run will be passed along and used for any auto compilation.

On Python 3.4+, coconut-run will use a __coconut_cache__ directory to cache the compiled Python. Note that __coconut_cache__ will always be removed from __file__.

Naming Source Files#

Coconut source files should, so the compiler can recognize them, use the extension .coco.

When Coconut compiles a .coco 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, 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 will compile to

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 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 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 the latest Python 3, Coconut code compiled in universal mode (the default --target)—and the Coconut compiler itself—should run on any Python version >= 2.6 on the 2.x branch or >= 3.2 on the 3.x branch (and on either CPython or PyPy).

To make Coconut built-ins universal across Python versions, Coconut makes available on any Python version built-ins that only exist in later versions, including automatically overwriting 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_. Specifically, the overwritten built-ins are:

  • py_bytes

  • py_chr

  • py_dict

  • py_hex

  • py_input

  • py_int

  • py_map

  • py_object

  • py_oct

  • py_open

  • py_print

  • py_range

  • py_str

  • py_super

  • py_zip

  • py_filter

  • py_reversed

  • py_enumerate

  • py_raw_input

  • py_xrange

  • py_repr

  • py_breakpoint

  • py_min

  • py_max

Note: Coconut’s repr can be somewhat tricky, as it will attempt to remove the u before reprs of unicode strings on Python 2, but will not always be able to do so if the unicode string is nested.

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 packages 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:

  • the nonlocal keyword,

  • keyword-only function parameters (use pattern-matching function definition for universal code),

  • async and await statements (requires a specific target; Coconut will attempt different backports based on the targeted version),

  • := assignment expressions (requires --target 3.8),

  • positional-only function parameters (use pattern-matching function definition for universal code) (requires --target 3.8),

  • a[x, *y] variadic generic syntax (use type parameter syntax for universal code) (requires --target 3.11), and

  • except* multi-except statements (requires --target 3.11).

Note: Coconut also universalizes many magic methods, including making __bool__ and __set_name__ work on any Python version.

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 syntax differs across versions, Coconut syntax will always follow the latest Python 3 across all targets. The supported targets are:

  • universal, univ (the 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: will work on any Python >= 3.3

  • 3.4: will work on any Python >= 3.4

  • 3.5: will work on any Python >= 3.5

  • 3.6: will work on any Python >= 3.6

  • 3.7: will work on any Python >= 3.7

  • 3.8: will work on any Python >= 3.8

  • 3.9: will work on any Python >= 3.9

  • 3.10: will work on any Python >= 3.10

  • 3.11: will work on any Python >= 3.11

  • 3.12: will work on any Python >= 3.12

  • 3.13: will work on any Python >= 3.13

  • sys: chooses the target corresponding to the current Python version

  • psf: will work on any Python not considered end-of-life by the PSF (Python Software Foundation)

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

strict Mode#

If the --strict (-s for short) flag is enabled, Coconut will perform additional checks on the code being compiled. It is recommended that you use the --strict flag if you are starting a new Coconut project, as it will help you write cleaner code. Specifically, the extra checks done by --strict are:

  • disabling deprecated features (making them entirely unavailable to code compiled with --strict),

  • errors instead of warnings on unused imports (unless they have a # NOQA or # noqa comment),

  • errors instead of warnings when overwriting built-ins (unless a backslash is used to escape the built-in name),

  • warning on missing __init__.coco files when compiling in --package mode,

  • throwing errors on various style problems (see list below).

The style issues which will cause --strict to throw an error are:

  • mixing of tabs and spaces

  • use of from __future__ imports (Coconut does these automatically)

  • inheriting from object in classes (Coconut does this automatically)

  • semicolons at end of lines

  • use of u to denote Unicode strings (all Coconut strings are Unicode strings)

  • f-strings with no format expressions in them

  • commas after statement lambdas (not recommended as it can be unclear whether the comma is inside or outside the lambda)

  • missing new line at end of file

  • trailing whitespace at end of lines

  • use of the Python-style lambda statement (use Coconut’s lambda syntax instead)

  • use of backslash continuation (use parenthetical continuation instead)

  • Python-3.10/PEP-634-style dotted names in pattern-matching (Coconut style is to preface these with ==)

  • use of : instead of <: to specify upper bounds in Coconut’s type parameter syntax

Note that many of the above style issues will still show a warning if --strict is not present.


In addition to the newer Python features that Coconut can backport automatically itself to older Python versions, Coconut will also automatically compile code to make use of a variety of external backports as well. These backports are automatically installed with Coconut if needed and Coconut will automatically use them instead of the standard library if the standard library is not available. These backports are:

Note that, when distributing compiled Coconut code, if you use any of these backports, you’ll need to make sure that the requisite backport module is included as a dependency.


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 (e.g. for IntelliJ IDEA see registering file types).


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 use the .coco file extension (should be all you need to do for Spyder) and/or enter coconut as the language being highlighted 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 use IPython (the Python kernel for the Jupyter framework) notebooks or console, 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.

Simply installing Coconut should add a Coconut kernel to your Jupyter/IPython notebooks. If you are having issues accessing the Coconut kernel, however, the special command coconut --jupyter install will re-install the Coconut kernel to ensure it is using the current Python as well as add the additional kernels Coconut (Default Python), Coconut (Default Python 2), and Coconut (Default Python 3) which will use, respectively, the Python accessible as python, python2, and python3 (these kernels are accessible in the console as coconut_py, coconut_py2, and coconut_py3). Coconut also supports coconut --jupyter install --user for user installation. Furthermore, the Coconut kernel fully supports nb_conda_kernels to enable accessing the Coconut kernel in one Conda environment from another Conda environment.

The Coconut kernel will always compile using the parameters: --target sys --line-numbers --keep-lines --no-wrap-types.

Coconut also provides the following commands:

  • coconut --jupyter notebook will ensure that the Coconut kernel is available and launch a Jupyter/IPython notebook.

  • coconut --jupyter console will launch a Jupyter/IPython console using the Coconut kernel.

  • coconut --jupyter lab will ensure that the Coconut kernel is available and launch JupyterLab.

Additionally, Jupytext contains special support for the Coconut kernel and Coconut contains special support for Papermill.


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.

Note: Unlike the normal Coconut command-line, %%coconut defaults to the sys target rather than the universal target.

Type Checking#

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 coconut 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.

You can also run mypy—or any other static type checker—directly on the compiled Coconut. If the static type checker is unable to find the necessary stub files, however, then you may need to:

  1. run coconut --mypy install and

  2. tell your static type checker of choice to look in ~/.coconut_stubs for stub files (for mypy, this is done by adding it to your MYPYPATH).

To distribute your code with checkable type annotations, you’ll need to include coconut as a dependency (though a --no-deps install should be fine), as installing it is necessary to make the requisite stub files available. You’ll also probably want to include a py.typed file.


To explicitly annotate your code with types to be checked, Coconut supports (on all Python versions):

By default, all type annotations are compiled to Python-2-compatible type comments, which means they should all work on any Python version.

Sometimes, MyPy will not know how to handle certain Coconut constructs, such as addpattern. For the addpattern case, it is recommended to pass --allow-redefinition to MyPy (i.e. run coconut <args> --mypy --allow-redefinition), though in some cases --allow-redefinition may not be sufficient. In that case, either hide the offending code using TYPE_CHECKING or put a # type: ignore comment on the Coconut line which is generating the line MyPy is complaining about and the comment will be added to every generated line.


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")
>>> reveal_type(a)
<string>:19: note: Revealed type is 'builtins.unicode'

For more information on reveal_type, see reveal_type and reveal_locals.

numpy Integration#

To allow for better use of numpy objects in Coconut, all compiled Coconut code will do a number of special things to better integrate with numpy (if numpy is available to import when the code is run). Specifically:

  • Coconut’s multidimensional array literal and array concatenation syntax supports numpy objects, including using fast numpy concatenation methods if given numpy arrays rather than Coconut’s default much slower implementation built for Python lists of lists.

  • Many of Coconut’s built-ins include special numpy support, specifically:

    • fmap will use numpy.vectorize to map over numpy arrays.

    • multi_enumerate allows for easily looping over all the multidimensional indices in a numpy array.

    • cartesian_product can compute the Cartesian product of given numpy arrays as a numpy array.

    • all_equal allows for easily checking if all the elements in a numpy array are the same.

  • numpy.ndarray is registered as a, enabling it to be used in sequence patterns.

  • numpy objects are allowed seamlessly in Coconut’s implicit coefficient syntax, allowing the use of e.g. A B**2 shorthand for A * B**2 when A and B are numpy arrays (note: not A @ B**2).

  • Coconut supports @ for matrix multiplication of numpy arrays on all Python versions, as well as supplying the (@) operator function.

Additionally, Coconut provides the exact same support for pandas, xarray, pytorch, and jax.numpy objects.

xonsh Support#

Coconut integrates with xonsh to allow the use of Coconut code directly from your command line. To use Coconut in xonsh, simply pip install coconut should be all you need to enable the use of Coconut syntax in the xonsh shell. In some circumstances, in particular depending on the installed xonsh version, adding xontrib load coconut to your xonshrc file might be necessary.

For an example of using Coconut from xonsh:

user@computer ~ $ xontrib load coconut
user@computer ~ $ cd ./files
user@computer ~ $ $(ls -la) |> .splitlines() |> len

Compilation always uses the same parameters as in the Coconut Jupyter kernel.

Note that the way that Coconut integrates with xonsh, @(<code>) syntax and the execx command will only work with Python code, not Coconut code. Additionally, Coconut will only compile individual commands—Coconut will not touch the .xonshrc or any other .xsh files.



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

====================== ==========================
Symbol(s)              Associativity
====================== ==========================
await x                n/a
**                     right (allows unary)
f x                    n/a
+, -, ~                unary
*, /, //, %, @         left
+, -                   left
<<, >>                 left
&                      left
&:                     yes
^                      left
|                      left
::                     yes (lazy)
..                     yes
a `b` c,               left (captures lambda)
  all custom operators
??                     left (short-circuits)
..>, <.., ..*>, <*..,  n/a (captures lambda)
  ..**>, <**..
|>, <|, |*>, <*|,      left (captures lambda)
  |**>, <**|
==, !=, <, >,
  <=, >=,
  in, not in,
  is, is not           n/a
not                    unary
and                    left (short-circuits)
or                     left (short-circuits)
x if c else y,         ternary (short-circuits)
  if c then x else y
=>                     right
====================== ==========================

For example, since addition has a higher precedence than piping, expressions of the form x |> y + z are equivalent to x |> (y + z).


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. Statement lambdas support full statements rather than just expressions and allow for the use of pattern-matching function definition.

Deprecated: -> can be used as an alternative to =>, though ->-based lambdas are disabled inside type annotations to avoid conflicting with Coconut’s enhanced type annotation syntax.


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))

Implicit Lambdas#

Coconut also supports implicit lambdas, which allow a lambda to take either no arguments or a single argument. Implicit lambdas are formed with the usual Coconut lambda operator =>, in the form (=> expression). This is equivalent to ((_=None) => expression). When an argument is passed to an implicit lambda, it will be assigned to _, replacing the default value None.

Below are two examples of implicit lambdas. The first uses the implicit argument _, while the second does not.

Single Argument Example:

square = (=> _**2)

No-Argument Example:

import random

get_random_number = (=> random.random())

Note: Nesting implicit lambdas can lead to problems with the scope of the _ parameter to each lambda. It is recommended that nesting implicit lambdas be avoided.

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.

Additionally, ? can even be used as the value of keyword arguments to convert them into positional arguments. For example, f$(x=?) is effectively equivalent to

def new_f(x, *args, **kwargs):
    kwargs["x"] = x
    return f(*args, **kwargs)

Unlike functools.partial, Coconut’s partial application will preserve the __name__ of the wrapped function.


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
(|**>)  => keyword argument pipe forward
(<|)    => pipe backward
(<*|)   => multiple-argument pipe backward
(<**|)  => keyword argument pipe backward
(|?>)   => None-aware pipe forward
(|?*>)  => None-aware multi-arg pipe forward
(|?**>) => None-aware keyword arg pipe forward
(<?|)   => None-aware pipe backward
(<*?|)  => None-aware multi-arg pipe backward
(<**?|) => None-aware keyword arg pipe backward

The None-aware pipe operators here are equivalent to a monadic bind treating the object as a Maybe monad composed of either None or the given object. Thus, x |?> f is equivalent to None if x is None else f(x). Note that only the object being piped, not the function being piped into, may be None for None-aware pipes.

Additionally, some special syntax constructs are only available in pipes to enable doing as many operations as possible via pipes if so desired:

  • For working with async functions in pipes, all non-starred pipes support piping into await to await the awaitable piped into them, such that x |> await is equivalent to await x.

  • All non-starred pipes support piping into (<name> := .) (mirroring the syntax for operator implicit partials) to assign the piped in item to <name>.

  • All pipe operators support a lambda as the last argument, despite lambdas having a lower precedence. Thus, a |> x => b |> c is equivalent to a |> (x => b |> c), not a |> (x => b) |> c.

Note: To visually spread operations across several lines, just use parenthetical continuation.


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 significantly slower than the Python-style syntax. 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.

This applies even to in-place pipes such as |>=.



def sq(x) = x**2
(1, 2) |*> (+) |> sq |> print
async def do_stuff(some_data) = (
    |> async_func
    |> await
    |> post_proc


import operator
def sq(x): return x**2
print(sq(operator.add(1, 2)))
async def do_stuff(some_data):
    return post_proc(await async_func(some_data))

Function Composition#

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. 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 ..> and <.. function composition pipe operators also have multi-arg, keyword, and None variants as with normal pipes. The full list of function composition pipe operators is:

..>    => forwards function composition pipe
<..    => backwards function composition pipe
..*>   => forwards multi-arg function composition pipe
<*..   => backwards multi-arg function composition pipe
..**>  => forwards keyword arg function composition pipe
<**..  => backwards keyword arg function composition pipe
..?>   => forwards None-aware function composition pipe
<?..   => backwards None-aware function composition pipe
..?*>  => forwards None-aware multi-arg function composition pipe
<*?..  => backwards None-aware multi-arg function composition pipe
..?**> => forwards None-aware keyword arg function composition pipe
<**?.. => backwards None-aware keyword arg function composition pipe

Note that None-aware function composition pipes don’t allow either function to be None—rather, they allow the return of the first evaluated function to be None, in which case None is returned immediately rather than calling the next function.

The .. operator has lower precedence than :: but higher precedence than infix functions while the ..> pipe operators have a precedence directly higher than normal pipes.

All function composition operators also have in-place versions (e.g. ..=).

Since all forms of function composition always call the first function in the composition (f in f ..> g and g in f <.. g) with exactly the arguments passed into the composition, all forms of function composition will preserve all metadata attached to the first function in the composition, including the function’s signature and any of that function’s attributes.

Note: for composing async functions, see and_then and and_then_await.



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))

Iterator Slicing#

Coconut uses a $ sign right after an iterator before a slice to perform iterator slicing, as in it$[:5]. 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 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 __iter_getitem__ (always used if available) or __getitem__ (only used if the object is a 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.

Iterator Chaining#

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. Chains are reiterable (can be iterated over multiple times and get the same result) only when the iterators passed in are reiterable. The in-place operator is ::=.

Note that lazy lists and flatten are used under the hood to implement chaining such that a :: b is equivalent to flatten((|a, b|)).


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.

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.

The allowable notations for infix calls are:

x `f` y  =>  f(x, y)
`f` x    =>  f(x)
x `f`    =>  f(x)
`f`      =>  f()

Additionally, infix notation supports a lambda as the last argument, despite lambdas having a lower precedence. Thus, a `func` b => c is equivalent to func(a, b => c).

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))

Custom Operators#

Coconut allows you to declare your own custom operators with the syntax

operator <op>

where <op> is whatever sequence of Unicode characters you want to use as a custom operator. The operator statement must appear at the top level and only affects code that comes after it.

Once declared, you can use your custom operator anywhere where you would be able to use an infix function as well as refer to the actual operator itself with the same (<op>) syntax as in other operator functions. Since custom operators work like infix functions, they always have the same precedence as infix functions and are always left-associative. Custom operators can be used as binary, unary, or none-ary operators, and both prefix and postfix notation for unary operators is supported.

Some example syntaxes for defining custom operators once declared:

def x <op> y: ...
def <op> x = ...
match def (x) <op> (y): ...
(<op>) = ...
from module import name as (<op>)

And some example syntaxes for using custom operators:

x <op> y
x <op> y <op> z
<op> x
x <op>
x = (<op>)
(x <op> .)
(. <op> y)
match x <op> in ...: ...
match x <op> y in ...: ...

Additionally, to import custom operators from other modules, Coconut supports the special syntax:

from <module> import operator <op>

Custom operators will often need to be surrounded by whitespace (or parentheses when used as an operator function) to be parsed correctly.

If a custom operator that is also a valid name is desired, you can use a backslash before the name to get back the name instead using Coconut’s keyword/variable disambiguation syntax.

Note: redefining existing Coconut operators using custom operator definition syntax is forbidden, including Coconut’s built-in Unicode operator alternatives.



operator %%
(%%) = math.remainder
10 %% 3 |> print

operator !!
(!!) = bool
!! 0 |> print

operator log10
from math import \log10 as (log10)
100 log10 |> print


print(math.remainder(10, 3))



None Coalescing#

Coconut provides ?? as a None-coalescing operator, similar to the ?? null-coalescing operator in C# and Swift. Additionally, Coconut implements all of the None-aware operators proposed in PEP 505.

Coconut’s ?? operator evaluates to its left operand if that operand is not None, otherwise its right operand. The expression foo ?? bar evaluates to foo as long as it isn’t None, and to bar if it is. The None-coalescing operator is short-circuiting, such that if the left operand is not None, the right operand won’t be evaluated. This allows the right operand to be a potentially expensive operation without incurring any unnecessary cost.

The None-coalescing operator has a precedence in-between infix function calls and composition pipes, and is left-associative.



could_be_none() ?? calculate_default_value()


(lambda result: result if result is not None else calculate_default_value())(could_be_none())

Coalescing Assignment Operator#

The in-place assignment operator is ??=, which allows conditionally setting a variable if it is currently None.

foo = 1
bar = None
foo ??= 10  # foo is still 1
bar ??= 10  # bar is now 10

As described with the standard ?? operator, the None-coalescing assignment operator will not evaluate the right hand side unless the left hand side is None.

baz = 0
baz ??= expensive_task()  # right hand side isn't evaluated

Other None-Aware Operators#

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. This is sometimes known as a “safe navigation” operator.

When using a None-aware operator for member access, either for a method or an attribute, the syntax is obj?.method() or obj?.attr respectively. obj?.attr is equivalent to obj.attr if obj is not None else obj. This does not prevent an AttributeError if attr is not an attribute or method of obj.

The None-aware indexing operator is used identically to normal indexing, using ?[] instead of []. seq?[index] is equivalent to the expression seq[index] is seq is not None else seq. Using this operator will not prevent an IndexError if index is outside the bounds of seq.

Coconut also supports None-aware pipe operators and function composition pipes.



could_be_none?.attr     # attribute access
could_be_none?(arg)     # function calling
could_be_none?.method() # method calling
could_be_none?$(arg)    # partial application
could_be_none()?[0]     # indexing


import functools
(lambda result: None if result is None else result.attr)(could_be_none())
(lambda result: None if result is None else result(arg))(could_be_none())
(lambda result: None if result is None else result.method())(could_be_none())
(lambda result: None if result is None else functools.partial(result, arg))(could_be_none())
(lambda result: None if result is None else result[0])(could_be_none())
(lambda result: None if result is None else result.attr[index].method())(could_be_none())

Protocol Intersection#

Coconut uses the &: operator to indicate protocol intersection. That is, for two typing.Protocols Protocol1 and Protocol1, Protocol1 &: Protocol2 is equivalent to a Protocol that combines the requirements of both Protocol1 and Protocol2.

The recommended way to use Coconut’s protocol intersection operator is in combination with Coconut’s operator Protocols. Note, however, that while &: will work anywhere, operator Protocols will only work inside type annotations (which means, for example, you’ll need to do type HasAdd = (+) instead of just HasAdd = (+)).

See Coconut’s enhanced type annotation for more information on how Coconut handles type annotations more generally.



from typing import Protocol

class X(Protocol):
    x: str

class Y(Protocol):
    y: str

def foo(xy: X &: Y) -> None:
    print(xy.x, xy.y)

type CanAddAndSub = (+) &: (-)


from typing import Protocol, TypeVar, Generic

class X(Protocol):
    x: str

class Y(Protocol):
    y: str

class XY(X, Y, Protocol):

def foo(xy: XY) -> None:
    print(xy.x, xy.y)

T = TypeVar("T", infer_variance=True)
U = TypeVar("U", infer_variance=True)
V = TypeVar("V", infer_variance=True)

class CanAddAndSub(Protocol, Generic[T, U, V]):
    def __add__(self: T, other: U) -> V:
        raise NotImplementedError
    def __sub__(self: T, other: U) -> V:
        raise NotImplementedError
    def __neg__(self: T) -> V:
        raise NotImplementedError

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.

Note: these are only the default, built-in unicode operators. Coconut supports custom operator definition to define your own.

Full List#

 (\u21d2)                 => "=>"
 (\u2192)                  => "->"
× (\xd7)                    => "*" (only multiplication)
 (\u2191)                  => "**" (only exponentiation)
÷ (\xf7)                    => "/" (only division)
÷/ (\xf7/)                  => "//"
 (\u207b)                  => "-" (only negation)
 (\u2260) or ¬= (\xac=)    => "!="
 (\u2264) or  (\u2286)    => "<="
 (\u2265) or  (\u2287)    => ">="
 (\u228a)                  => "<"
 (\u228b)                  => ">"
 (\u2229)                  => "&"
 (\u222a)                  => "|"
« (\xab)                    => "<<"
» (\xbb)                    => ">>"
 (\u2026)                  => "..."
λ (\u03bb)                  => "lambda"
 (\u21a6)                 => "|>"
 (\u21a4)                 => "<|"
*↦ (*\u21a6)               => "|*>"
↤* (\u21a4*)               => "<*|"
**↦ (**\u21a6)             => "|**>"
↤** (\u21a4**)             => "<**|"
?↦ (?\u21a6)               => "|?>"
↤? (?\u21a4)               => "<?|"
?*↦ (?*\u21a6)             => "|?*>"
↤*? (\u21a4*?)             => "<*?|"
?**↦ (?**\u21a6)           => "|?**>"
↤**? (\u21a4**?)           => "<**?|"
 (\u2218)                  => ".."
∘> (\u2218>)                => "..>"
<∘ (<\u2218)                => "<.."
∘*> (\u2218*>)              => "..*>"
<*∘ (<*\u2218)              => "<*.."
∘**> (\u2218**>)            => "..**>"
<**∘ (<**\u2218)            => "<**.."
∘?> (\u2218?>)              => "..?>"
<?∘ (<?\u2218)              => "<?.."
∘?*> (\u2218?*>)            => "..?*>"
<*?∘ (<*?\u2218)            => "<*?.."
∘?**> (\u2218?**>)          => "..?**>"
<**?∘ (<**?\u2218)          => "<**?.."
 (\u23e8)                  => "e" (in scientific notation)



Coconut provides fully-featured, functional pattern-matching through its match statements. Coconut match syntax is a strict superset of Python’s match syntax.

Note: In describing Coconut’s pattern-matching syntax, this section focuses on match statements, but Coconut’s pattern-matching can also be used in many other places, such as pattern-matching function definition, case statements, destructuring assignment, match data, and match for.


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.

All pattern-matching in Coconut is atomic, such that no assignments will be executed unless the whole match succeeds.

Syntax Specification#

Coconut match statement syntax is

match <pattern> [not] 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 as below. In the syntax specification below, brackets denote optional syntax and parentheses followed by a * denote that the syntax may appear zero or more times.

pattern ::= and_pattern ("or" and_pattern)*  # match any

and_pattern ::= as_pattern ("and" as_pattern)*  # match all

as_pattern ::= infix_pattern ("as" name)*  # explicit binding

infix_pattern ::= bar_or_pattern ("`" EXPR "`" [EXPR])*  # infix check

bar_or_pattern ::= pattern ("|" pattern)*  # match any

base_pattern ::= (
    "(" pattern ")"                  # parentheses
    | "None" | "True" | "False"      # constants
    | NUMBER                         # numbers
    | STRING                         # strings
    | ["as"] NAME                    # variable binding
    | "==" EXPR                      # equality check
    | "is" EXPR                      # identity check
    | DOTTED_NAME                    # implicit equality check (disabled in destructuring assignment)
    | NAME "(" patterns ")"          # classes or data types
    | "data" NAME "(" patterns ")"   # data types
    | "class" NAME "(" patterns ")"  # classes
    | "{" pattern_pairs              # dictionaries
        ["," "**" (NAME | "{}")] "}" #  (keys must be constants or equality checks)
    | ["s" | "f" | "m"] "{"
        ["," ("*_" | "*()")]
      "}"                            # sets
    | (EXPR) -> pattern              # view patterns
    | "(" patterns ")"               # sequences can be in tuple form
    | "[" patterns "]"               #  or in list form
    | "(|" patterns "|)"             # lazy lists
    | ("(" | "[")                    # star splits
        [patterns ","]
        "*" pattern
        ["," patterns]
        ["*" pattern
        ["," patterns]]
      (")" | "]")
    | [(                             # sequence splits
        "(" patterns ")"
        | "[" patterns "]"
      ) "+"] NAME ["+" (
        "(" patterns ")"                 # this match must be the same
        | "[" patterns "]"               #  construct as the first match
      )] ["+" NAME ["+" (
        "(" patterns ")"                 # and same here
        | "[" patterns "]"
    | [(                             # iterable splits
        "(" patterns ")"
        | "[" patterns "]"
        | "(|" patterns "|)"
      ) "::"] NAME ["::" (
        "(" patterns ")"
        | "[" patterns "]"
        | "(|" patterns "|)"
      )] [ "::" NAME [
        "(" patterns ")"
        | "[" patterns "]"
        | "(|" patterns "|)"
    | [STRING "+"] NAME              # complex string matching
        ["+" STRING]
        ["+" NAME ["+" 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:

  • Variable Bindings:

    • Implicit Bindings (<var>): 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 matches to the same value.

      • If the variable name _ is used, nothing will be bound and everything will always match to it (_ is Coconut’s “wildcard”).

    • Explicit Bindings (<pattern> as <var>): will bind <var> to <pattern>.

  • Basic Checks:

    • Constants, Numbers, and Strings: will only match to the same constant, number, or string in the same position in the arguments.

    • Equality Checks (==<expr>): will check that whatever is in that position is == to the expression <expr>.

    • Identity Checks (is <expr>): will check that whatever is in that position is the expression <expr>.

  • Arbitrary Function Patterns:

    • Infix Checks (<pattern> `<op>` <expr>): will check that the operator <op>$(?, <expr>) returns a truthy value when called on whatever is in that position, then matches <pattern>. For example, x `isinstance` int will check that whatever is in that position isinstance$(?, int) and bind it to x. If <expr> is not given, will simply check <op> directly rather than <op>$(<expr>). Additionally, `<op>` can instead be a custom operator (in that case, no backticks should be used).

    • View Patterns ((<expression>) -> <pattern>): calls <expression> on the item being matched and matches the result to <pattern>. The match fails if a MatchError is raised. <expression> may be unparenthesized only when it is a single atom.

  • Class and Data Type Matching:

    • Classes or Data Types (<name>(<args>)): will match as a data type if given a Coconut data type (or a tuple of Coconut data types) and a class otherwise.

    • Data Types (data <name>(<args>)): will check that whatever is in that position is of data type <name> and will match the attributes to <args>. Generally, data <name>(<args>) will match any data type that could have been constructed with makedata(<name>, <args>). Includes support for positional arguments, named arguments, default arguments, and starred arguments. Also supports strict attributes by prepending a dot to the attribute name that raises AttributError if the attribute is not present rather than failing the match (e.g. data MyData(.my_attr=<some_pattern>)).

    • Classes (class <name>(<args>)): does PEP-634-style class matching. Also supports strict attribute matching as above.

  • Mapping Destructuring:

    • Dicts ({<key>: <value>, ...}): will match any mapping ( with the given keys and values that match the value patterns. Keys must be constants or equality checks.

    • Dicts With Rest ({<pairs>, **<rest>}): will match a mapping ( containing all the <pairs>, and will put a dict of everything else into <rest>. If <rest> is {}, will enforce that the mapping is exactly the same length as <pairs>.

  • Set Destructuring:

    • Sets (s{<constants>, *_}): will match a set ( that contains the given <constants>, though it may also contain other items. The s prefix and the *_ are optional.

    • Fixed-length Sets (s{<constants>, *()}): will match a set ( that contains the given <constants>, and nothing else.

    • Frozensets (f{<constants>}): will match a frozenset (frozenset) that contains the given <constants>. May use either normal or fixed-length syntax.

    • Multisets (m{<constants>}): will match a multiset (collections.Counter) that contains at least the given <constants>. May use either normal or fixed-length syntax.

  • Sequence Destructuring:

    • Lists ([<patterns>]), Tuples ((<patterns>)): will only match a sequence ( of the same length, and will check the contents against <patterns> (Coconut automatically registers numpy arrays and collections.deque objects as sequences).

    • Lazy lists ((|<patterns>|)): same as list or tuple matching, but checks for an Iterable ( instead of a Sequence.

    • Head-Tail Splits (<list/tuple> + <var> or (<patterns>, *<var>)): will match the beginning of the sequence against the <list/tuple>/<patterns>, then bind the rest to <var>, and make it the type of the construct used.

    • Init-Last Splits (<var> + <list/tuple> or (*<var>, <patterns>)): 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> or (<patterns>, *<var>, <patterns>)): the combination of a head-tail and an init-last split.

    • Search Splits (<var1> + <list/tuple> + <var2> or (*<var1>, <patterns>, *<var2>)): searches for the first occurrence of the <list/tuple>/<patterns> in the sequence, then puts everything before into <var1> and everything after into <var2>.

    • Head-Last Search Splits (<list/tuple> + <var> + <list/tuple> + <var> + <list/tuple> or (<patterns>, *<var>, <patterns>, *<var>, <patterns>)): the combination of a head-tail split and a search split.

    • Iterable Splits (<list/tuple/lazy list> :: <var> :: <list/tuple/lazy list> :: <var> :: <list/tuple/lazy list>): same as other sequence destructuring, but works on any iterable (, including infinite iterators (note that if an iterator is matched against it will be modified unless it is reiterable).

    • Complex String Matching (<string> + <var> + <string> + <var> + <string>): string matching supports the same destructuring options as above.

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 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 ensure proper matching for a custom object, it’s recommended to register it with the proper abstract base classes.



def factorial(value):
    match 0 in value:
        return 1
    else: match n `isinstance` int in value if n > 0:  # Coconut allows nesting of statements on the same line
        return n * factorial(n-1)
        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")

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

Showcases matching to data types and the default equality operator. 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.

class Tree
data Empty() from Tree
data Leaf(n) from Tree
data Node(l, r) from Tree

def depth(Tree()) = 0

addpattern def depth(Tree(n)) = 1

addpattern 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)

addpattern 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.

def odd_primes(p=3) =
    (p,) :: filter(=> _ % p != 0, odd_primes(p + 2))

def primes() =
    (2,) :: odd_primes()

def twin_primes(_ :: [p, (.-2) -> p] :: ps) =
    [(p, p+2)] :: twin_primes([p + 2] :: ps)

addpattern def twin_primes() =  # type: ignore

twin_primes()$[:5] |> list |> print

Showcases the use of an iterable search pattern and a view pattern to construct an iterator of all twin primes.

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 blocks serve as 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.

Coconut’s case blocks are an extension of Python 3.10’s case blocks to support additional pattern-matching constructs added by Coconut (and Coconut will ensure that they work on all Python versions, not just 3.10+).

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

match <value>:
    case <pattern> [if <cond>]:
    case <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. If no else is present and no match succeeds, then the case statement is simply skipped over as with match statements (though unlike destructuring assignments).

Additionally, cases can be used as the top-level keyword instead of match, and in such a case block match is allowed for each case rather than case. Deprecated: Coconut also supports case instead of cases as the top-level keyword for backwards-compatibility purposes.



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

Example of using Coconut’s case syntax.

cases {"a": 1, "b": 2}:
    match {"a": a}:
    match _:
        assert False
assert a == 1

Example of the cases keyword instead.

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

match for#

Coconut supports pattern-matching in for loops, where the pattern is matched against each item in the iterable. The syntax is

[match] for <pattern> in <iterable>:

which is equivalent to the destructuring assignment

for elem in <iterable>:
    match <pattern> = elem

Pattern-matching can also be used in async for loops, with both async match for and match async for allowed as explicit syntaxes.



for {"user": uid, **_} in get_data():


for user_data in get_data():
    uid = user_data["user"]


Coconut’s data keyword is used to create immutable, algebraic data types, including 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. Additionally, Coconut allows type parameters to be specified in brackets after <name> using Coconut’s type parameter syntax.

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. If you need to inherit magic methods from a base class in your data type, such subclassing is the recommended method, as the data ... from ... syntax will overwrite any magic methods in the base class with magic methods built for the new data type.

Compared to namedtuples, from which data types are derived, data types:

  • use typed equality,

  • don’t support tuple addition or multiplication (unless explicitly defined via special methods in the data body),

  • support starred, typed, and pattern-matching arguments, and

  • have special pattern-matching behavior.

Like namedtuples, data types also support a variety of extra methods, such as ._asdict() and ._replace(**kwargs).


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.



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

addpattern def size(Leaf(n)) = 1

addpattern 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.

Python: Can’t be done without a series of method definitions for each data type. See the compiled code for the Python syntax.

match data#

In addition to normal data statements, Coconut also supports pattern-matching data statements that enable the use of Coconut’s pattern-matching syntax to define the data type’s constructor. Pattern-matching data types look like

[match] data <name>(<patterns>) [from <base class>]:

where <patterns> are exactly as in pattern-matching functions.

It is important to keep in mind that pattern-matching data types vary from normal data types in a variety of ways. First, like pattern-matching functions, they raise MatchError instead of TypeError when passed the wrong arguments. Second, pattern-matching data types will not do any special handling of starred arguments. Thus,

data vec(*xs)

when iterated over will iterate over all the elements of xs, but

match data vec(*xs)

when iterated over will only give the single element xs.



data namedpt(name `isinstance` str, x `isinstance` int, y `isinstance` int):
    def mag(self) = (self.x**2 + self.y**2)**0.5

Python: Can’t be done without a series of method definitions for each data type. See the compiled code for the Python syntax.


Coconut’s where statement is fairly straightforward. The syntax for a where statement is just

<stmt> where:

which executes <body> followed by <stmt>, with the exception that any new variables defined in <body> are available only in <stmt> (though they are only mangled, not deleted, such that e.g. lambdas can still capture them).



result = a + b where:
    a = 1
    b = 2


_a = 1
_b = 2
result = _a + _b

async with for#

In modern Python async code, such as when using contextlib.aclosing, it is often recommended to use a pattern like

async with aclosing(my_generator()) as values:
    async for value in values:

since it is substantially safer than the more syntactically straightforward

async for value in my_generator():

This is especially true when using trio, which completely disallows iterating over async generators with async for, instead requiring the above async with ... async for pattern using utilities such as trio_util.trio_async_generator.

Since this pattern can often be quite syntactically cumbersome, Coconut provides the shortcut syntax

async with for value in aclosing(my_generator()):

which compiles to exactly the pattern above.

async with for also supports pattern-matching, just like normal Coconut for loops.



from trio_util import trio_async_generator

async def my_generator():
    # yield values, possibly from a nursery or cancel scope
    # ...

async with for value in my_generator():


from trio_util import trio_async_generator

async def my_generator():
    # yield values, possibly from a nursery or cancel scope
    # ...

async with my_generator() as agen:
    async for value in agen:

Handling Keyword/Variable Name Overlap#

In Coconut, the following keywords are also valid variable names:

While Coconut can usually disambiguate these two use cases, special syntax is available for disambiguating them if necessary. Note that, if what you’re writing can be interpreted as valid Python 3, Coconut will always prefer that interpretation by default.

To specify that you want a variable, prefix the name with a backslash as in \data, and to specify that you want a keyword, prefix the name with a colon as in :match.

Additionally, backslash syntax for escaping variable names can also be used to distinguish between variable names and custom operators as well as explicitly signify that an assignment to a built-in is desirable to dismiss --strict warnings.

Finally, such disambiguation syntax can also be helpful for letting syntax highlighters know what you’re doing.



\data = 5
# without the colon, Coconut will interpret this as the valid Python match[x, y] = input_list
:match [x, y] = input_list


data = 5
x, y = input_list


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

[async|match|copyclosure] 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. Note that the async, match, and copyclosure keywords can be combined and can be in any order.

If the last statement (not followed by a semicolon) in a statement lambda is an expression, it will automatically be returned.

Statement lambdas also support implicit lambda syntax such that def => _ is equivalent to def (_=None) => _ as well as explicitly marking them as pattern-matching such that match def (x) => x will be a pattern-matching function.

Additionally, statement lambdas have slightly different scoping rules than normal lambdas. When a statement lambda is inside of an expression with an expression-local variable, such as a normal lambda or comprehension, the statement lambda will capture the value of the variable at the time that the statement lambda is defined (rather than a reference to the overall namespace as with normal lambdas). As a result, while [=> y for y in range(2)] |> map$(call) |> list is [1, 1], [def => y for y in range(2)] |> map$(call) |> list is [0, 1]. Note that this only works for expression-local variables: to copy the entire namespace at the time of function definition, use copyclosure (which can be used with statement lambdas).

Note that statement lambdas have a lower precedence than normal lambdas and thus capture things like trailing commas. To avoid confusion, statement lambdas should always be wrapped in their own set of parentheses.

Deprecated: Statement lambdas also support -> instead of =>. Note that when using ->, any lambdas in the body of the statement lambda must also use -> rather than =>.



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


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

Type annotations#

Another case where statement lambdas would be used over standard lambdas is when the parameters to the lambda are typed with type annotations. Statement lambdas use the standard Python syntax for adding type annotations to their parameters:

f = def (c: str) -> None => print(c)

g = def (a: int, b: int) -> int => a ** b

Deprecated: if the deprecated -> is used in place of =>, then return type annotations will not be available.

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.

All operator functions also support implicit partial application, e.g. (. + 1) is equivalent to (=> _ + 1).


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#

(::)        => (itertools.chain)  # will not evaluate its arguments lazily
($)         => (functools.partial)
(.)         => (getattr)
(,)         => (*args) => args  # (but pickleable)
(+)         => (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)
(|>)        => # pipe forward
(|*>)       => # multi-arg pipe forward
(|**>)      => # keyword arg pipe forward
(<|)        => # pipe backward
(<*|)       => # multi-arg pipe backward
(<**|)      => # keyword arg pipe backward
(|?>)       => # None-aware pipe forward
(|?*>)      => # None-aware multi-arg pipe forward
(|?**>)     => # None-aware keyword arg pipe forward
(<?|)       => # None-aware pipe backward
(<*?|)      => # None-aware multi-arg pipe backward
(<**?|)     => # None-aware keyword arg pipe backward
(..), (<..) => # backward function composition
(..>)       => # forward function composition
(<*..)      => # multi-arg backward function composition
(..*>)      => # multi-arg forward function composition
(<**..)     => # keyword arg backward function composition
(..**>)     => # keyword arg forward function composition
(not)       => (operator.not_)
(and)       => # boolean and
(or)        => # boolean or
(is)        => (operator.is_)
(is not)    => (operator.is_not)
(in)        => (operator.contains)
(not in)    => # negative containment
(assert)    => def (cond, msg=None) => assert cond, msg  # (but a better msg if msg is None)
(raise)     => def (exc=None, from_exc=None) => raise exc from from_exc  # or just raise if exc is None
# operator functions for multidimensional array concatenation use brackets:
[;]         => def (x, y) => [x; y]
[;;]        => def (x, y) => [x;; y]
...  # and so on for any number of semicolons
# there are two operator functions that don't require parentheses:
.[]         => (operator.getitem)
.$[]        => # iterator slicing operator

For an operator function for function application, see call.

Though no operator function is available for await, an equivalent syntax is available for pipes in the form of awaitable |> await.



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


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

Implicit Partial Application#

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

# attribute access and method calling
.attr1.attr2        =>  operator.attrgetter("attr1.attr2")
.method(args)       =>  operator.methodcaller("method", args)
.attr.method(args)  =>  .attr ..> .method(args)

# indexing
.[a:b:c]            =>  operator.itemgetter(slice(a, b, c))
.[x][y]             => .[x] ..> .[y]
.method[x]          => .method ..> .[x]
seq[]               =>  operator.getitem$(seq)

# iterator indexing
.$[a:b:c]           =>  # the equivalent of .[a:b:c] for iterators
.$[x]$[y]           => .$[x] ..> .$[y]
iter$[]             =>  # the equivalent of seq[] for iterators

# currying
func$               =>  ($)$(func)

In addition, for every Coconut operator function, Coconut supports syntax for implicitly partially applying that operator function as

(. <op> <arg>)
(<arg> <op> .)

where <op> is the operator function and <arg> is any expression. Note that, as with operator functions themselves, the parentheses are necessary for this type of implicit partial application. This syntax is slightly different for multidimensional array concatenation operator functions, which use brackets instead of parentheses.

Furthermore, Coconut also supports implicit operator function partials for arbitrary functions as

(. `<name>` <arg>)
(<arg> `<name>` .)

based on Coconut’s infix notation where <name> is the name of the function. Additionally, `<name>` can instead be a custom operator (in that case, no backticks should be used).

Deprecated: Coconut also supports obj. as an implicit partial for getattr$(obj), but its usage is deprecated and will show a warning to switch to getattr$(obj) instead.



1 |> "123"[]
mod$ <| 5 <| 3
3 |> (.*2) |> (.+1)


mod(5, 3)
(3 * 2) + 1

Enhanced Type Annotation#

Since Coconut syntax is a superset of the latest 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.

Since not all supported Python versions support the typing module, Coconut provides the TYPE_CHECKING built-in for hiding your typing imports and TypeVar definitions from being executed at runtime. Coconut will also automatically use typing_extensions over typing objects at runtime when importing them from typing, even when they aren’t natively supported on the current Python version (this works even if you just do import typing and then typing.<Object>).

Furthermore, when compiling type annotations to Python 3 versions without PEP 563 support, Coconut wraps annotation in strings to prevent them from being evaluated at runtime (to avoid this, e.g. if you want to use annotations at runtime, --no-wrap-types will disable all wrapping, including via PEP 563 support). Only on --target 3.13 does --no-wrap-types do nothing, since there PEP 649 support is used instead.

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:

A | B
    => typing.Union[A, B]
(A; B)
    => typing.Tuple[A, B]
    => typing.Optional[A]
    => typing.Sequence[A]
    => typing.Iterable[A]
() -> <ret>
    => typing.Callable[[], <ret>]
<arg> -> <ret>
    => typing.Callable[[<arg>], <ret>]
(<args>) -> <ret>
    => typing.Callable[[<args>], <ret>]
-> <ret>
    => typing.Callable[..., <ret>]
(<args>, **<ParamSpec>) -> <ret>
    => typing.Callable[typing.Concatenate[<args>, <ParamSpec>], <ret>]
async (<args>) -> <ret>
    => typing.Callable[[<args>], typing.Awaitable[<ret>]]

where typing is the Python 3.5 built-in typing module. For more information on the Callable syntax, see PEP 677, which Coconut fully supports.

Additionally, many of Coconut’s operator functions will compile into equivalent Protocols instead when inside a type annotation. See below for the full list and specification.

Note: The transformation to Union is not done on Python 3.10 as Python 3.10 has native PEP 604 support.

To use these transformations in a type alias, use the syntax

type <name> = <type>

which will allow <type> to include Coconut’s special type annotation syntax and type <name> as a typing.TypeAlias. If you try to instead just do a naked <name> = <type> type alias, Coconut won’t be able to tell you’re attempting a type alias and thus won’t apply any of the above transformations.

Such type alias statements—as well as all class, data, and function definitions in Coconut—also support Coconut’s type parameter syntax, allowing you to do things like type OrStr[T] = T | str.

Supported Protocols#

Using Coconut’s operator function syntax inside of a type annotation will instead produce a Protocol corresponding to that operator (or raise a syntax error if no such Protocol is available). All available Protocols are listed below.

For the operator functions


the resulting Protocol is

class SupportsOp[T, U, V](Protocol):
    def __op__(self: T, other: U) -> V:
        raise NotImplementedError(...)

where __op__ is the magic method corresponding to that operator.

For the operator function (-), the resulting Protocol is:

class SupportsMinus[T, U, V](Protocol):
    def __sub__(self: T, other: U) -> V:
        raise NotImplementedError
    def __neg__(self: T) -> V:
        raise NotImplementedError

For the operator function (~), the resulting Protocol is:

class SupportsInv[T, V](Protocol):
    def __invert__(self: T) -> V:
        raise NotImplementedError(...)

List vs. Sequence#

Importantly, note that T[] does not map onto typing.List[T] but onto typing.Sequence[T]. This allows the resulting type to be covariant, such that if U is a subtype of T, then U[] is a subtype of T[]. Additionally, Sequence[T] allows for tuples, and when writing in an idiomatic functional style, assignment should be rare and tuples should be common. Using Sequence covers both cases, accommodating tuples and lists and preventing indexed assignment. When an indexed assignment is attempted into a variable typed with Sequence, MyPy will generate an error:

foo: int[] = [0, 1, 2, 3, 4, 5]
foo[0] = 1   # MyPy error: "Unsupported target for indexed assignment"

If you want to use List instead (e.g. if you want to support indexed assignment), use the standard Python 3.5 variable type annotation syntax: foo: List[<type>].

Note: To easily view your defined types, see reveal_type and reveal_locals.



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

type CanAddAndSub = (+) &: (-)


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))

T = typing.TypeVar("T", infer_variance=True)
U = typing.TypeVar("U", infer_variance=True)
V = typing.TypeVar("V", infer_variance=True)
class CanAddAndSub(typing.Protocol, typing.Generic[T, U, V]):
    def __add__(self: T, other: U) -> V:
        raise NotImplementedError
    def __sub__(self: T, other: U) -> V:
        raise NotImplementedError
    def __neg__(self: T) -> V:
        raise NotImplementedError

Multidimensional Array Literal/Concatenation Syntax#

Coconut supports multidimensional array literal and array concatenation/stack syntax.

By default, all multidimensional array syntax will simply operate on Python lists of lists (or any non-str Sequence). However, if numpy objects are used, the appropriate numpy calls will be made instead. To give custom objects multidimensional array concatenation support, define type(obj).__matconcat__ (should behave as np.concat), obj.ndim (should behave as np.ndarray.ndim), and obj.reshape (should behave as np.ndarray.reshape).

As a simple example, 2D matrices can be constructed by separating the rows with ;; inside of a list literal:

>>> [1, 2 ;;
     3, 4]

[[1, 2], [3, 4]]
>>> import numpy as np
>>> np.array([1, 2 ;; 3, 4])
array([[1, 2],
       [3, 4]])

As can be seen, np.array (or equivalent) can be used to turn the resulting list of lists into an actual array. This syntax works because ;; inside of a list literal functions as a concatenation/stack along the -2 axis (with the inner arrays being broadcast to (1, 2) arrays before concatenation). Note that this concatenation is done entirely in Python lists of lists here, since the np.array call comes only at the end.

In general, the number of semicolons indicates the dimension from the end on which to concatenate. Thus, ; indicates conatenation along the -1 axis, ;; along the -2 axis, and so on. Before concatenation, arrays are always broadcast to a shape which is large enough to allow the concatenation.

Thus, if a is a numpy array, [a; a] is equivalent to np.concatenate((a, a), axis=-1), while [a ;; a] would be equivalent to a version of np.concatenate((a, a), axis=-2) that also ensures that a is at least two dimensional. For normal lists of lists, the behavior is the same, but is implemented without any numpy calls.

If multiple different concatenation operators are used, the operators with the least number of semicolons will bind most tightly. Thus, you can write a 3D array literal as:

>>> [1, 2 ;;
     3, 4
     5, 6 ;;
     7, 8]

[[[1, 2], [3, 4]], [[5, 6], [7, 8]]]

Note: the operator functions for multidimensional array concatenation are spelled [;], [;;], etc. (with any number of parentheses). The implicit partials are similarly spelled [. ; x], [x ; .], etc.

Comparison to Julia#

Coconut’s multidimensional array syntax is based on that of Julia. The primary difference between Coconut’s syntax and Julia’s syntax is that multidimensional arrays are row-first in Coconut (following numpy), but column-first in Julia. Thus, ; is vertical concatenation in Julia but horizontal concatenation in Coconut and ;; is horizontal concatenation in Julia but vertical concatenation in Coconut.



>>> [[1;;2] ; [3;;4]]
[[1, 3], [2, 4]]

Array literals can be written in column-first order if the columns are first created via vertical concatenation (;;) and then joined via horizontal concatenation (;).

>>> [range(3) |> list ;; x+1 for x in range(3)]
[[0, 1, 2], [1, 2, 3]]

Arbitrary expressions, including comprehensions, are allowed in multidimensional array literals.

>>> import numpy as np
>>> a = np.array([1, 2 ;; 3, 4])
>>> [a ; a]
array([[1, 2, 1, 2],
       [3, 4, 3, 4]])
>>> [a ;; a]
array([[1, 2],
       [3, 4],
       [1, 2],
       [3, 4]])
>>> [a ;;; a]
array([[[1, 2],
        [3, 4]],

       [[1, 2],
        [3, 4]]])

General showcase of how the different concatenation operators work using numpy arrays.

Python: The equivalent Python array literals can be seen in the printed representations in each example.

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. Unlike normal iterators, however, lazy lists can be iterated over multiple times and still return the same result. Lazy lists can be created in Coconut simply by surrounding a comma-separated list of items with (| and |) (so-called “banana brackets”) instead of [ and ] for a list or ( and ) for a tuple.

Lazy lists use reiterable under the hood to enable them to be iterated over multiple times. Lazy lists will even continue to be reiterable when combined with lazy chaining.


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 Function Application and Coefficients#

Coconut supports implicit function application of the form f x y, which is compiled to f(x, y) (note: not f(x)(y) as is common in many languages with automatic currying).

Additionally, if the first argument is not callable, and is instead an int, float, complex, or numpy object, then the result is multiplication rather than function application, such that 2 x is equivalent to 2*x.

Though the first item may be any atom, following arguments are highly restricted, and must be:

  • variables/attributes (e.g. a.b),

  • literal constants (e.g. True),

  • number literals (e.g. 1.5) (and no binary, hex, or octal), or

  • one of the above followed by an exponent (e.g. a**-5).

For example, (f .. g) x 1 will work, but f x [1], f x (1+2), and f "abc" will not.

Implicit function application and coefficient syntax is only intended for simple use cases. For more complex cases, use the standard multiplication operator *, standard function application, or pipes.

Implicit function application and coefficient syntax has a lower precedence than ** but a higher precedence than unary operators. As a result, 2 x**2 + 3 x is equivalent to 2 * x**2 + 3 * x.

Due to potential confusion, some syntactic constructs are explicitly disallowed in implicit function application and coefficient syntax. Specifically:

  • Strings are always disallowed everywhere in implicit function application / coefficient syntax due to conflicting with Python’s implicit string concatenation.

  • Multiplying two or more numeric literals with implicit coefficient syntax is prohibited, so 10 20 is not allowed.

  • await is not allowed in front of implicit function application and coefficient syntax. To use await, simply parenthesize the expression, as in await (f x).

Note: implicit function application and coefficient syntax is disabled when using Coconut in xonsh due to conflicting with console commands.



def f(x, y) = (x, y)
print(f 5 10)
def p1(x) = x + 1
print <| p1 5
quad = 5 x**2 + 3 x + 1


def f(x, y): return (x, y)
print(f(100, 5+6))
def p1(x): return x + 1
quad = 5 * x**2 + 3 * x + 1

Keyword Argument Name Elision#

When passing in long variable names as keyword arguments of the same name, Coconut supports the syntax


as a shorthand for


Such syntax is also supported in partial application and anonymous namedtuples.



really_long_variable_name_1 = get_1()
really_long_variable_name_2 = get_2()


really_long_variable_name_1 = get_1()
really_long_variable_name_2 = get_2()

Anonymous Namedtuples#

Coconut supports anonymous namedtuple literals, such that (a=1, b=2) can be used just as (1, 2), but with added names. Anonymous namedtuples are always pickleable.

The syntax for anonymous namedtuple literals is:

(<name> [: <type>] = <value>, ...)

where, if <type> is given for any field, typing.NamedTuple is used instead of collections.namedtuple.

Anonymous namedtuples also support keyword argument name elision.


On Python versions >=3.6, _namedtuple_of is provided as a built-in that can mimic the behavior of anonymous namedtuple literals such that _namedtuple_of(a=1, b=2) is equivalent to (a=1, b=2). Since _namedtuple_of is only available on Python 3.6 and above, however, it is generally recommended to use anonymous namedtuple literals instead, as they work on any Python version.

_namedtuple_of is just provided to give namedtuple literals a representation that corresponds to an expression that can be used to recreate them.



users = [
    (id=1, name="Alice"),
    (id=2, name="Bob"),


from collections import namedtuple

users = [
    namedtuple("_", "id, name")(1, "Alice"),
    namedtuple("_", "id, name")(2, "Bob"),

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. Set literals also support unpacking syntax (e.g. s{*xs}).

Additionally, Coconut also supports replacing the s with an f to generate a frozenset or an m to generate a Coconut multiset.



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))

Alternative Ternary Operator#

Python supports the ternary operator syntax

result = if_true if condition else if_false

which, since Coconut is a superset of Python, Coconut also supports.

However, Coconut also provides an alternative syntax that uses the more conventional argument ordering as

result = if condition then if_true else if_false

making use of the Coconut-specific then keyword (though Coconut still allows then as a variable name).



value = (
    if should_use_a() then a
    else if should_use_b() then b
    else if should_use_c() then c
    else fallback


value = (
    a if should_use_a() else
    b if should_use_b() else
    c if should_use_c() else

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).

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.



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

def is_odd(0) = False
addpattern def is_odd(n `isinstance` int if n > 0) = is_even(n-1)

Showcases tail call optimization.

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

--no-tco flag#

Tail call optimization 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.

--no-tco does not disable tail recursion elimination. This is because tail recursion elimination is usually faster than doing nothing, while other types of tail call optimization are usually slower than doing nothing. Tail recursion elimination results in a big performance win because Python has a fairly large function call overhead. By unwinding a recursive function, far fewer function calls need to be made. When the --no-tco flag is disabled, Coconut will attempt to do all types of tail call optimizations, handling non-recursive tail calls, split pattern-matching functions, mutual recursion, and tail recursion. When the --no-tco flag is enabled, Coconut will no longer perform any tail call optimizations other than tail recursion elimination.

Tail Recursion Elimination and Python lambdas#

Coconut does not perform tail recursion elimination in functions that utilize lambdas or inner functions. This is because of the way that Python handles lambdas.

Each lambda stores a pointer to the namespace enclosing it, rather than a copy of the namespace. Thus, if the Coconut compiler tries to recycle anything in the namespace that produced the lambda, which needs to be done for TRE, the lambda can be changed retroactively.

A simple example demonstrating this behavior in Python:

x = 1
foo = lambda: x
print(foo())  # 1
x = 2         # Directly alter the values in the namespace enclosing foo
print(foo())  # 2 (!)

Because this could have unintended and potentially damaging consequences, Coconut opts to not perform TRE on any function with a lambda or inner function.

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

[async] def <name>(<args>) = <expr>

for one-liners or

[async] 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 pattern-matching functions are just normal functions, except where the arguments are patterns to be matched against instead of variables to be assigned to. The syntax for pattern-matching function definition is

[match] def <name>(<arg>, <arg>, ... [if <cond>]) [-> <return_type>]:

where <arg> is defined as

[*|**] <pattern> [= <default>]

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, <default> is the optional default if no argument is passed, and <return_type> is the optional return type annotation (note that argument type annotations are not supported for pattern-matching functions). The match keyword at the beginning is optional, but is sometimes necessary to disambiguate pattern-matching function definition from normal function definition, since Python function definition will always take precedence. Note that the async and match keywords can be in any order.

If <pattern> has a variable name (via any variable binding that binds the entire pattern, e.g. x in int(x) or [a, b] as x), the resulting pattern-matching function will support keyword arguments using that variable name.

In addition to supporting pattern-matching in their arguments, pattern-matching function definitions also have a couple of notable differences compared to Python functions. Specifically:

  • If pattern-matching function definition fails, it will raise a MatchError (just like destructuring assignment) instead of a TypeError.

  • All defaults in pattern-matching function definition are late-bound rather than early-bound. Thus, match def f(xs=[]) = xs will instantiate a new list for each call where xs is not given, unlike def f(xs=[]) = xs, which will use the same list for all calls where xs is unspecified. This also allows defaults for later arguments to be specified in terms of matched values from earlier arguments, as in match def f(x, y=x) = (x, y).

Pattern-matching function definition can also be combined with async functions, copyclosure functions, yield functions, infix function definition, and assignment function syntax. The various keywords in front of the def can be put in any order.



def last_two(_ + [a, b]):
    return a, b
def xydict_to_xytuple({"x": x `isinstance` int, "y": y `isinstance` 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.

addpattern Functions#

Coconut provides the addpattern def syntax as a shortcut for the full

match def func(...):

syntax using the addpattern decorator.

Additionally, addpattern def will act just like a normal match def if the function has not previously been defined, allowing for addpattern def to be used for each case rather than requiring match def for the first case and addpattern def for future cases.

If you want to put a decorator on an addpattern def function, make sure to put it on the last pattern function.



addpattern def factorial(0) = 1
addpattern 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.

copyclosure Functions#

Coconut supports the syntax

copyclosure def <name>(<args>):

to define a function that uses as its closure a shallow copy of its enclosing scopes at the time that the function is defined, rather than a reference to those scopes (as with normal Python functions).

For example,`in

def outer_func():
    funcs = []
    for x in range(10):
        copyclosure def inner_func():
            return x
    return funcs

the resulting inner_funcs will each return a different x value rather than all the same x value, since they look at what x was bound to at function definition time rather than during function execution.

copyclosure functions can also be combined with async functions, yield functions, pattern-matching functions, infix function definition, and assignment function syntax. The various keywords in front of the def can be put in any order.

If global or nonlocal are used in a copyclosure function, they will not be able to modify variables in enclosing scopes. However, they will allow state to be preserved accross multiple calls to the copyclosure function.



def outer_func():
    funcs = []
    for x in range(10):
        copyclosure def inner_func():
            return x
    return funcs


from functools import partial

def outer_func():
    funcs = []
    for x in range(10):
        def inner_func(_x):
            return _x
        funcs.append(partial(inner_func, x))
    return funcs

Explicit Generators#

Coconut supports the syntax

yield def <name>(<args>):

to denote that you are explicitly defining a generator function. This is useful to ensure that, even if all the yields in your function are removed, it’ll always be a generator function.

Explicit generator functions can also be combined with async functions, copyclosure functions, pattern-matching functions, infix function definition, and assignment function syntax (though note that assignment function syntax here creates a generator return). The various keywords in front of the def can be put in any order.



yield def empty_it(): pass


def empty_it():
    if False:

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. Dotted function definition can be combined with all other types of function definition above.



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.

Type Parameter Syntax#

Coconut fully supports Python 3.12 PEP 695 type parameter syntax on all Python versions.

That includes type parameters for classes, data types, and all types of function definition. For different types of function definition, the type parameters always come in brackets right after the function name. Coconut’s enhanced type annotation syntax is supported for all type parameter bounds.

Warning: until mypy adds support for infer_variance=True in TypeVar, TypeVars created this way will always be invariant.

Additionally, Coconut supports the alternative bounds syntax of type NewType[T <: bound] = ... rather than type NewType[T: bound] = ..., to make it more clear that it is an upper bound rather than a type. In --strict mode, <: is required over : for all type parameter bounds. Deprecated: <= can also be used as an alternative to <:.

Note that the <: syntax should only be used for type bounds, not type constraints—for type constraints, Coconut style prefers the vanilla Python : syntax, which helps to disambiguate between the two cases, as they are functionally different but otherwise hard to tell apart at a glance. This is enforced in --strict mode.

Note that, by default, all type declarations are wrapped in strings to enable forward references and improve runtime performance. If you don’t want that—e.g. because you want to use type annotations at runtime—simply pass the --no-wrap-types flag.

PEP 695 Docs#

Defining a generic class prior to this PEP looks something like this.

from typing import Generic, TypeVar

_T_co = TypeVar("_T_co", covariant=True, bound=str)

class ClassA(Generic[_T_co]):
    def method1(self) -> _T_co:

With the new syntax, it looks like this.

class ClassA[T: str]:
    def method1(self) -> T:

Here is an example of a generic function today.

from typing import TypeVar

_T = TypeVar("_T")

def func(a: _T, b: _T) -> _T:

And the new syntax.

def func[T](a: T, b: T) -> T:

Here is an example of a generic type alias today.

from typing import TypeAlias

_T = TypeVar("_T")

ListOrSet: TypeAlias = list[_T] | set[_T]

And with the new syntax.

type ListOrSet[T] = list[T] | set[T]



data D[T](x: T, y: T)

def my_ident[T](x: T) -> T = x

Python: Can’t be done without a complex definition for the data type. See the compiled code for the Python syntax.

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.



class Tree
data Empty from Tree
data Leaf(item) from Tree
data Node(left, right) from Tree

Python: Can’t be done without a series of method definitions for each data type. See the compiled code for the Python syntax.

Statement Nesting#

Coconut supports the nesting of compound statements on the same line. This allows the mixing of match and if statements together, as well as 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) and len(input_list) >= 1:
    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:

In-line global And nonlocal Assignment#

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



global state_a, state_b = 10, 100
global state_c += 1


global state_a, state_b; state_a, state_b = 10, 100
global state_c; state_c += 1

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. When using Coconut to compile to another variant of Python, make sure you name your source file properly to ensure the resulting compiled code has the right file extension for the intended usage.

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 the latest 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. This also includes support as per PEP 679 for parenthesized assert statements.

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:

Assignment Expression Chaining#

Unlike Python, Coconut allows assignment expressions to be chained, as in a := b := c. Note, however, that assignment expressions in general are currently only supported on --target 3.8 or higher.



(a := b := 1)


(a := (b := 1))


Built-In Function Decorators#


addpattern(base_func, *add_funcs, allow_any_func=False)

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. addpattern also supports a shortcut syntax where the new patterns can be passed in directly.

Roughly equivalent to:

def _pattern_adder(base_func, add_func):
    def add_pattern_func(*args, **kwargs):
            return base_func(*args, **kwargs)
        except MatchError:
            return add_func(*args, **kwargs)
    return add_pattern_func
def addpattern(base_func, *add_funcs, allow_any_func=False):
    """Decorator to add a new case to a pattern-matching function (where the new case is checked last).

    Pass allow_any_func=True to allow any object as the base_func rather than just pattern-matching functions.
    If add_func is passed, addpattern(base_func, add_func) is equivalent to addpattern(base_func)(add_func).
    if not add_funcs:
        return addpattern$(base_func)
    for add_func in add_funcs:
        base_func = pattern_adder(base_func, add_func)
    return base_func

If you want to give an addpattern function a docstring, make sure to put it on the last function.

Note that the function taken by addpattern must be a pattern-matching function. If addpattern receives a non pattern-matching function, the function with not raise MatchError, and addpattern won’t be able to detect the failed match. Thus, if a later function was meant to be called, addpattern will not know that the first match failed and the correct path will never be reached.

For example, the following code raises a TypeError:

def print_type():
    print("Received no arguments.")

def print_type(int()):
    print("Received an int.")

print_type()  # appears to work
print_type(1) # TypeError: print_type() takes 0 positional arguments but 1 was given

This can be fixed by using either the match or addpattern keyword. For example:

match def print_type():
    print("Received no arguments.")

addpattern def print_type(int()):
    print("Received an int.")

print_type(1)  # Works as expected
print_type("This is a string.") # Raises MatchError

The last case in an addpattern function, however, doesn’t have to be a pattern-matching function if it is intended to catch all remaining cases.

To catch this mistake, addpattern will emit a warning if passed what it believes to be a non-pattern-matching function. However, this warning can sometimes be erroneous if the original pattern-matching function has been wrapped in some way, in which case you can pass allow_any_func=True to dismiss the warning.



def factorial(0) = 1

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

Simple example of adding a new pattern to a pattern-matching function.

"[A], [B]" |> windowsof$(3) |> map$(addpattern(
    (def (("[","A","]")) => "A"),
    (def (("[","B","]")) => "B"),
    (def ((_,_,_)) => None),
)) |> filter$((.is None) ..> (not)) |> list |> print

An example of a case where using the addpattern function is necessary over the addpattern keyword due to the use of in-line pattern-matching statement lambdas.

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.


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.


memoize(maxsize=None, typed=False)


Coconut provides functools.lru_cache as a built-in under the name memoize with the modification that the maxsize parameter is set to None by default. memoize makes the use case of optimizing recursive functions easier, as a maxsize of None is usually what is desired in that case.

Use of memoize requires functools.lru_cache, which exists in the Python 3 standard library, but under Python 2 will require pip install backports.functools_lru_cache to function. Additionally, if on Python 2 and backports.functools_lru_cache is present, Coconut will patch functools such that functools.lru_cache = backports.functools_lru_cache.lru_cache.

Note that, if the function to be memoized is a generator or otherwise returns an iterator, recursive_generator can also be used to achieve a similar effect, the use of which is required for recursive generators.

Python Docs#


@memoize(maxsize=None, typed=False)

Decorator to wrap a function with a memoizing callable that saves up to the maxsize most recent calls. It can save time when an expensive or I/O bound function is periodically called with the same arguments.

Since a dictionary is used to cache results, the positional and keyword arguments to the function must be hashable.

Distinct argument patterns may be considered to be distinct calls with separate cache entries. For example, f(a=1, b=2) and f(b=2, a=1) differ in their keyword argument order and may have two separate cache entries.

If user_function is specified, it must be a callable. This allows the memoize decorator to be applied directly to a user function, leaving the maxsize at its default value of None:

def count_vowels(sentence):
    return sum(sentence.count(vowel) for vowel in 'AEIOUaeiou')

If maxsize is set to None, the LRU feature is disabled and the cache can grow without bound.

If typed is set to true, function arguments of different types will be cached separately. If typed is false, the implementation will usually regard them as equivalent calls and only cache a single result. (Some types such as str and int may be cached separately even when typed is false.)

Note, type specificity applies only to the function’s immediate arguments rather than their contents. The scalar arguments, Decimal(42) and Fraction(42) are be treated as distinct calls with distinct results. In contrast, the tuple arguments ('answer', Decimal(42)) and ('answer', Fraction(42)) are treated as equivalent.

The decorator also provides a cache_clear() function for clearing or invalidating the cache.

The original underlying function is accessible through the __wrapped__ attribute. This is useful for introspection, for bypassing the cache, or for rewrapping the function with a different cache.

The cache keeps references to the arguments and return values until they age out of the cache or until the cache is cleared.

If a method is cached, the self instance argument is included in the cache. See How do I cache method calls?

An LRU (least recently used) cache works best when the most recent calls are the best predictors of upcoming calls (for example, the most popular articles on a news server tend to change each day). The cache’s size limit assures that the cache does not grow without bound on long-running processes such as web servers.

In general, the LRU cache should only be used when you want to reuse previously computed values. Accordingly, it doesn’t make sense to cache functions with side-effects, functions that need to create distinct mutable objects on each call, or impure functions such as time() or random().

Example of efficiently computing Fibonacci numbers using a cache to implement a dynamic programming technique:

def fib(n):
    if n < 2:
        return n
    return fib(n-1) + fib(n-2)

>>> [fib(n) for n in range(16)]
[0, 1, 1, 2, 3, 5, 8, 13, 21, 34, 55, 89, 144, 233, 377, 610]

>>> fib.cache_info()
CacheInfo(hits=28, misses=16, maxsize=None, currsize=16)


def fib(n if n < 2) = n

def fib(n) = fib(n-1) + fib(n-2)


    from functools import lru_cache
except ImportError:
    from backports.functools_lru_cache import lru_cache
def fib(n):
    if n < 2:
        return n
    return fib(n-1) + fib(n-2)



Coconut provides the @override decorator to allow declaring a method definition in a subclass as an override of some parent class method. When @override is used on a method, if a method of the same name does not exist on some parent class, the class definition will raise a RuntimeError. @override works with other decorators such as @classmethod and @staticmethod, but only if @override is the outer-most decorator.

Additionally, override will present to type checkers as typing_extensions.override.



class A:
    x = 1
    def f(self, y) = self.x + y

class B:
    def f(self, y) = self.x + y + 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 recursive_generator decorator that memoizes and makes reiterable any generator or other stateless function that returns an iterator. To use recursive_generator 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 arguments that are equal, 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.

Importantly, recursive_generator 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.

One pitfall to keep in mind working with recursive_generator is that it shouldn’t be used in contexts where the function can potentially be called multiple times with the same iterator object as an input, but with that object not actually corresponding to the same items (e.g. because the first time the object hasn’t been iterated over yet and the second time it has been).

Deprecated: recursive_iterator is available as a deprecated alias for recursive_generator. Note that deprecated features are disabled in --strict mode.



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.

Built-In Types#


multiset(iterable=None, /, **kwds)

Coconut provides multiset as a built-in subclass of collections.Counter that additionally implements the full Set and MutableSet interfaces.

For easily constructing multisets, Coconut also provides multiset literals.

The new methods provided by multiset on top of collections.Counter are:

  • multiset.add(item): Add an element to a multiset.

  • multiset.discard(item): Remove an element from a multiset if it is a member.

  • multiset.remove(item): Remove an element from a multiset; it must be a member.

  • multiset.isdisjoint(other): Return True if two multisets have a null intersection.

  • multiset.__xor__(other): Return the symmetric difference of two multisets as a new multiset. Specifically: a ^ b = (a - b) | (b - a)

  • multiset.count(item): Return the number of times an element occurs in a multiset. Equivalent to multiset[item], but additionally verifies the count is non-negative.

  • multiset.__fmap__(func): Apply a function to the contents of the multiset, preserving counts; magic method for fmap.

Coconut also ensures that multiset supports rich comparisons and on all Python versions.



my_multiset = m{1, 1, 2}


from collections import Counter
my_counter = Counter((1, 1, 2))
my_counter[3] += 1
my_counter[2] -= 1
if my_counter[2] <= 0:
    del my_counter[2]


Expected(result=None, error=None)

Coconut’s Expected built-in is a Coconut data type that represents a value that may or may not be an error, similar to Haskell’s Either.

Expected is effectively equivalent to the following:

data Expected[T](result: T? = None, error: BaseException? = None):
    def __bool__(self) -> bool:
        return self.error is None
    def __fmap__[U](self, func: T -> U) -> Expected[U]:
        """Maps func over the result if it exists.

        __fmap__ should be used directly only when fmap is not available (e.g. when consuming an Expected in vanilla Python).
        return self.__class__(func(self.result)) if self else self
    def and_then[U](self, func: T -> Expected[U]) -> Expected[U]:
        """Maps a T -> Expected[U] over an Expected[T] to produce an Expected[U].
        Implements a monadic bind. Equivalent to fmap ..> .join()."""
        return self |> fmap$(func) |> .join()
    def join(self: Expected[Expected[T]]) -> Expected[T]:
        """Monadic join. Converts Expected[Expected[T]] to Expected[T]."""
        if not self:
            return self
        if not self.result `isinstance` Expected:
            raise TypeError("Expected.join() requires an Expected[Expected[_]]")
        return self.result
    def map_error(self, func: BaseException -> BaseException) -> Expected[T]:
        """Maps func over the error if it exists."""
        return self if self else self.__class__(error=func(self.error))
    def handle(self, err_type, handler: BaseException -> T) -> Expected[T]:
        """Recover from the given err_type by calling handler on the error to determine the result."""
        if not self and isinstance(self.error, err_type):
            return self.__class__(handler(self.error))
        return self
    def expect_error(self, *err_types: BaseException) -> Expected[T]:
        """Raise any errors that do not match the given error types."""
        if not self and not isinstance(self.error, err_types):
            raise self.error
        return self
    def unwrap(self) -> T:
        """Unwrap the result or raise the error."""
        if not self:
            raise self.error
        return self.result
    def or_else[U](self, func: BaseException -> Expected[U]) -> Expected[T | U]:
        """Return self if no error, otherwise return the result of evaluating func on the error."""
        return self if self else func(self.error)
    def result_or_else[U](self, func: BaseException -> U) -> T | U:
        """Return the result if it exists, otherwise return the result of evaluating func on the error."""
        return self.result if self else func(self.error)
    def result_or[U](self, default: U) -> T | U:
        """Return the result if it exists, otherwise return the default.

        Since .result_or() completely silences errors, it is highly recommended that you
        call .expect_error() first to explicitly declare what errors you are okay silencing.
        return self.result if self else default

Expected is primarily used as the return type for safe_call.

Generally, the best way to use Expected is with fmap, which will apply a function to the result if it exists, or otherwise retain the error. If you want to sequence multiple Expected-returning operations, .and_then should be used instead of fmap. To handle specific errors, the following patterns are equivalent:

safe_call(might_raise_IOError).handle(IOError, const 10).unwrap()

To match against an Expected, just:

Expected(res) = Expected("result")
Expected(error=err) = Expected(error=TypeError())


def try_divide(x: float, y: float) -> Expected[float]:
        return Expected(x / y)
    except Exception as err:
        return Expected(error=err)

try_divide(1, 2) |> fmap$(.+1) |> print
try_divide(1, 0) |> fmap$(.+1) |> print

Python: Can’t be done without a complex Expected definition. See the compiled code for the Python syntax.


A MatchError is raised when a destructuring assignment or pattern-matching function fails, and thus MatchError is provided as a built-in for catching those errors. MatchError objects support three attributes: pattern, which is a string describing the failed pattern; value, which is the object that failed to match that pattern; and message which is the full error message. To avoid unnecessary repr calls, MatchError only computes the message once it is actually requested.

Additionally, if you are using view patterns, you might need to raise your own MatchError (though you can also just use a destructuring assignment or pattern-matching function definition to do so). To raise your own MatchError, just raise MatchError(pattern, value) (both arguments are optional).

In some cases where there are multiple Coconut packages installed at the same time, there may be multiple MatchErrors defined in different packages. Coconut can perform some magic under the hood to make sure that all these MatchErrors will seamlessly interoperate, but only if all such packages are compiled in --package mode rather than --standalone mode.

Generic Built-In Functions#


makedata(data_type, *args)

Coconut provides the makedata function to construct a container given the desired type and contents. This is particularly useful when writing alternative constructors for data types by overwriting __new__, since it allows direct access to the base constructor of the data type created with the Coconut data statement.

makedata takes the data type to construct as the first argument, and the objects to put in that container as the rest of the arguments.

makedata can also be used to extract the underlying constructor for match data types that bypasses the normal pattern-matching constructor.

Additionally, makedata can also be called with non-data type as the first argument, in which case it will do its best to construct the given type of object with the given arguments. This functionality is used internally by fmap.


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)

Python: Can’t be done without a series of method definitions for each data type. See the compiled code for the Python syntax.


fmap(func, obj)

In Haskell, 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 for Coconut’s data types.

fmap can also be used on the built-in objects str, dict, list, tuple, set, frozenset, bytes, bytearray, and dict as a variant of map that returns back an object of the same type.

For dict, or any other, fmap will map over the mapping’s .items() instead of the default iteration through its .keys(), with the new mapping reconstructed from the mapped over items. Deprecated: fmap$(starmap_over_mappings=True) will starmap over the .items() instead of map over them.

For asynchronous iterables, fmap will map asynchronously, making fmap equivalent in that case to

async def fmap_over_async_iters(func, async_iter):
    async for item in async_iter:
        yield func(item)

such that fmap can effectively be used as an async map.

Some objects from external libraries are also given special support:

  • For numpy objects, fmap will use np.vectorize to produce the result.

  • For pandas objects, fmap will use .apply along the last axis (so row-wise for DataFrame’s, element-wise for Series’s).

  • For xarray objects, fmap will first convert them into pandas objects, apply fmap, then convert them back.

The behavior of fmap for a given object can be overridden by defining an __fmap__(self, func) magic method that will be called whenever fmap is invoked on that object. Note that __fmap__ implementations should always satisfy the Functor Laws.

Deprecated: fmap(func, obj, fallback_to_init=True) will fall back to obj.__class__(map(func, obj)) if no fmap implementation is available rather than raise TypeError.



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

class Maybe
data Nothing() from Maybe
data Just(n) from Maybe

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

Python: Can’t be done without a series of method definitions for each data type. See the compiled code for the Python syntax.


call(func, /, *args, **kwargs)

Coconut’s call simply implements function application. Thus, call is effectively equivalent to

def call(f, /, *args, **kwargs) = f(*args, **kwargs)

call is primarily useful as an operator function for function application when writing in a point-free style.

Deprecated: of is available as a deprecated alias for call. Note that deprecated features are disabled in --strict mode.


safe_call(func, /, *args, **kwargs)

Coconut’s safe_call is a version of call that catches any Exceptions and returns an Expected containing either the result or the error.

safe_call is effectively equivalent to:

def safe_call(f, /, *args, **kwargs):
        return Expected(f(*args, **kwargs))
    except Exception as err:
        return Expected(error=err)

To define a function that always returns an Expected rather than raising any errors, simply decorate it with @safe_call$.



res, err = safe_call(=> 1 / 0) |> fmap$(.+1)

Python: Can’t be done without a complex Expected definition. See the compiled code for the Python syntax.


ident(x, *, side_effect=None)

Coconut’s ident is the identity function, generally equivalent to x => x.

ident also accepts one keyword-only argument, side_effect, which specifies a function to call on the argument before it is returned. Thus, ident is effectively equivalent to:

def ident(x, *, side_effect=None):
    if side_effect is not None:
    return x

ident is primarily useful when writing in a point-free style (e.g. in combination with lift) or for debugging pipes where ident$(side_effect=print) can let you see what is being piped.



Coconut’s const simply constructs a function that, whatever its arguments, just returns the given value. Thus, const is equivalent to a pickleable version of

def const(value) = (*args, **kwargs) => value

const is primarily useful when writing in a point-free style (e.g. in combination with lift).


flip(func, nargs=None)

Coconut’s flip(f, nargs=None) is a higher-order function that, given a function f, returns a new function with reversed argument order. If nargs is passed, only the first nargs arguments are reversed.

For the binary case, flip works as

flip(f, 2)(x, y) == f(y, x)

such that flip$(?, 2) implements the C combinator (flip in Haskell).

In the general case, flip is equivalent to a pickleable version of

def flip(f, nargs=None) =
    (*args, **kwargs) => (
        f(*args[::-1], **kwargs) if nargs is None
        else f(*(args[nargs-1::-1] + args[nargs:]), **kwargs)

lift and lift_apart#

lift(func, *func_args, **func_kwargs)#

Coconut’s lift built-in is a higher-order function that takes in a function and “lifts” it up so that all of its arguments are functions.

As a simple example, for a binary function f(x, y) and two unary functions g(z) and h(z), lift works as

lift(f)(g, h)(z) == f(g(z), h(z))

such that in this case lift implements the S' combinator (liftA2 or liftM2 in Haskell).

In the general case, lift is equivalent to a pickleable version of

def lift(f) = (
    (*func_args, **func_kwargs) =>
        (*args, **kwargs) =>
                *(g(*args, **kwargs) for g in func_args),
                **{k: h(*args, **kwargs) for k, h in func_kwargs.items()}

lift also supports a shortcut form such that lift(f, *func_args, **func_kwargs) is equivalent to lift(f)(*func_args, **func_kwargs).

lift_apart(func, *func_args, **func_kwargs)#

Coconut’s lift_apart built-in is very similar to lift, except instead of duplicating the final arguments to each function, it separates them out.

For a binary function f(x, y) and two unary functions g(z) and h(z), lift_apart works as

lift_apart(f)(g, h)(z, w) == f(g(z), h(w))

such that in this case lift_apart implements the D2 combinator.

In the general case, lift_apart is equivalent to a pickleable version of

def lift_apart(f) = (
    (*func_args, **func_kwargs) =>
        (*args, **kwargs) =>
                *(f(x) for f, x in zip(func_args, args, strict=True)),
                **{k: func_kwargs[k](kwargs[k]) for k in func_kwargs.keys() | kwargs.keys()},

lift_apart supports the same shortcut form as lift.



xs_and_xsp1 = ident `lift(zip)` map$(=>_+1)
min_and_max = lift(,)(min, max)
plus_and_times = (+) `lift(,)` (*)


def xs_and_xsp1(xs):
    return zip(xs, map(lambda x: x + 1, xs))
def min_and_max(xs):
    return min(xs), max(xs)
def plus_and_times(x, y):
    return x + y, x * y


first_false_and_last_true = (
    lift(,)(ident, reversed)
    ..*> lift_apart(,)(dropwhile$(bool), dropwhile$(not))
    ..*> lift_apart(,)(.$[0], .$[0])


from itertools import dropwhile

def first_false_and_last_true(xs):
    rev_xs = reversed(xs)
    return (
        next(dropwhile(bool, xs)),
        next(dropwhile(lambda x: not x, rev_xs)),

and_then and and_then_await#

and_then(first_async_func, second_func)

and_then_await(first_async_func, second_async_func)

Coconut provides the and_then and and_then_await built-ins for composing async functions. Specifically:

  • To forwards compose an async function async_f with a normal function g (such that g is called on the result of awaiting async_f), write async_f `and_then` g.

  • To forwards compose an async function async_f with another async function async_g (such that async_g is called on the result of awaiting async_f, and then async_g is itself awaited), write async_f `and_then_await` async_g.

  • To forwards compose a normal function f with an async function async_g (such that async_g is called on the result of f), just write f ..> async_g.

Note that all of the above will always result in the resulting composition being an async function.

The built-ins are effectively equivalent to:

def and_then[**T, U, V](
    first_async_func: async (**T) -> U,
    second_func: U -> V,
) -> async (**T) -> V =
    async def (*args, **kwargs) => (
        first_async_func(*args, **kwargs)
        |> await
        |> second_func

def and_then_await[**T, U, V](
    first_async_func: async (**T) -> U,
    second_async_func: async U -> V,
) -> async (**T) -> V =
    async def (*args, **kwargs) => (
        first_async_func(*args, **kwargs)
        |> await
        |> second_async_func
        |> await

Like normal function composition, and_then and and_then_await will preserve all metadata attached to the first function in the composition.



load_and_send_data = (
    `and_then` proc_data
    `and_then_await` send_data


async def load_and_send_data():
    return await send_data(proc_data(await load_data_async()))

Built-Ins for Working with Iterators#

Enhanced Built-Ins#

Coconut’s map, zip, filter, reversed, and enumerate objects are enhanced versions of their Python equivalents that support:

  • The ability to be iterated over multiple times if the underlying iterators can be iterated over multiple times.

    • Note: This can lead to different behavior between Coconut built-ins and Python built-ins. Use py_ versions if the Python behavior is necessary.

  • reversed

  • repr

  • Optimized normal (and iterator) indexing/slicing (map, zip, reversed, and enumerate but not filter).

  • len (all but filter) (though bool will still always yield True).

  • PEP 618 zip(..., strict=True) support on all Python versions.

  • Added strict=True support to map as well (enforces that iterables are the same length in the multi-iterable case; uses zip under the hood such that errors will show up as zip(..., strict=True) errors).

  • 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

Indexing into other built-ins#

Though Coconut provides random access indexing/slicing to range, map, zip, reversed, and enumerate, Coconut cannot index into built-ins like filter, takewhile, or dropwhile directly, as there is no efficient way to do so.

range(10) |> filter$(i => i>3) |> .[0]  # doesn't work

In order to make this work, you can explicitly use iterator slicing, which is less efficient in the general case:

range(10) |> filter$(i => i>3) |> .$[0]  # works

For more information on Coconut’s iterator slicing, see here.



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.


range(0, 12, 2)[4]  # 8

map((i => i*2), range(10))[2]  # 4

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


reduce(function, iterable[, initial], /)

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

Python Docs#

reduce(function, iterable[, initial])

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 initial 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 initial 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)))



reiterable wraps the given iterable to ensure that every time the reiterable is iterated over, it produces the same results. Note that the result need not be a reiterable object if the given iterable is already reiterable. reiterable uses tee under the hood and tee can be used in its place, though reiterable is generally recommended over tee.



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


starmap(function, iterable)

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


zip_longest(*iterables, fillvalue=None)

Coconut provides an enhanced version of itertools.zip_longest as a built-in under the name zip_longest. zip_longest supports all the same features as Coconut’s enhanced zip as well as the additional attribute fillvalue.

Python Docs#

zip_longest(*iterables, fillvalue=None)

Make an iterator that aggregates elements from each of the iterables. If the iterables are of uneven length, missing values are filled-in with fillvalue. Iteration continues until the longest iterable is exhausted. Roughly equivalent to:

def zip_longest(*args, fillvalue=None):
    # zip_longest('ABCD', 'xy', fillvalue='-') --> Ax By C- D-
    iterators = [iter(it) for it in args]
    num_active = len(iterators)
    if not num_active:
    while True:
        values = []
        for i, it in enumerate(iterators):
                value = next(it)
            except StopIteration:
                num_active -= 1
                if not num_active:
                iterators[i] = repeat(fillvalue)
                value = fillvalue
        yield tuple(values)

If one of the iterables is potentially infinite, then the zip_longest() function should be wrapped with something that limits the number of calls (for example iterator slicing or takewhile). If not specified, fillvalue defaults to None.



result = zip_longest(range(5), range(10))


import itertools
result = itertools.zip_longest(range(5), range(10))


takewhile(predicate, iterable, /)

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 = numiter |> takewhile$(x => x < 0)


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


dropwhile(predicate, iterable, /)

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 = numiter |> dropwhile$(x => x < 0)


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


flatten(iterable, levels=1)

Coconut provides an enhanced version of itertools.chain.from_iterable as a built-in under the name flatten with added support for reversed, repr, in, .count(), .index(), and fmap.

By default, flatten only flattens the top level of the given iterable/array. If levels is passed, however, it can be used to control the number of levels flattened, with 0 meaning no flattening and None flattening as many iterables as are found. Note that if levels is set to any non-None value, the first levels levels must be iterables, or else an error will be raised.

Python Docs#


Alternate constructor for chain(). Gets chained inputs from a single iterable argument that is evaluated lazily. Roughly equivalent to:

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


iter_of_iters = [[1, 2], [3, 4]]
flat_it = iter_of_iters |> flatten |> list


from itertools import chain
iter_of_iters = [[1, 2], [3, 4]]
flat_it = list(chain.from_iterable(iter_of_iters))


scan(function, iterable[, initial])

Coconut provides a modified version of itertools.accumulate with opposite argument order as scan that also supports repr, len, and func/iter/initial 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(function, iterable[, initial])

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 function. (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.

If no initial is given, roughly equivalent to:

def scan(function, 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 = function(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


count(start=0, step=1)

Coconut provides a modified version of itertools.count that supports in, normal slicing, optimized iterator slicing, the standard count and index sequence methods, repr, and start/step attributes as a built-in under the name count. If the step parameter is set to None, count will behave like itertools.repeat instead.

Since count supports slicing, count()[...] can be used as a version of range that can in some cases be more readable. In particular, it is easy to accidentally write range(10, 2) when you meant range(0, 10, 2), but it is hard to accidentally write count()[10:2] when you mean count()[:10:2].

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
        if step:
          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.


cycle(iterable, times=None)

Coconut’s cycle is a modified version of itertools.cycle with a times parameter that controls the number of times to cycle through iterable before stopping. cycle also supports in, slicing, len, reversed, .count(), .index(), and repr.

Python Docs#


Make an iterator returning elements from the iterable and saving a copy of each. When the iterable is exhausted, return elements from the saved copy. Repeats indefinitely. Roughly equivalent to:

def cycle(iterable):
    # cycle('ABCD') --> A B C D A B C D A B C D ...
    saved = []
    for element in iterable:
        yield element
    while saved:
        for element in saved:
              yield element

Note, this member of the toolkit may require significant auxiliary storage (depending on the length of the iterable).



cycle(range(2), 2) |> list |> print


from itertools import cycle, islice
print(list(islice(cycle(range(2)), 4)))


cartesian_product(*iterables, repeat=1)

Coconut provides an enhanced version of itertools.product as a built-in under the name cartesian_product with added support for len, repr, in, .count(), and fmap.

Additionally, cartesian_product includes special support for numpy objects, in which case a multidimensional array is returned instead of an iterator.

Python Docs#

cartesian_product(*iterables, repeat=1)

Cartesian product of input iterables.

Roughly equivalent to nested for-loops in a generator expression. For example, cartesian_product(A, B) returns the same as ((x,y) for x in A for y in B).

The nested loops cycle like an odometer with the rightmost element advancing on every iteration. This pattern creates a lexicographic ordering so that if the input’s iterables are sorted, the product tuples are emitted in sorted order.

To compute the product of an iterable with itself, specify the number of repetitions with the optional repeat keyword argument. For example, product(A, repeat=4) means the same as cartesian_product(A, A, A, A).

This function is roughly equivalent to the following code, except that the actual implementation does not build up intermediate results in memory:

def cartesian_product(*args, repeat=1):
    # product('ABCD', 'xy') --> Ax Ay Bx By Cx Cy Dx Dy
    # product(range(2), repeat=3) --> 000 001 010 011 100 101 110 111
    pools = [tuple(pool) for pool in args] * repeat
    result = [[]]
    for pool in pools:
        result = [x+[y] for x in result for y in pool]
    for prod in result:
        yield tuple(prod)

Before cartesian_product() runs, it completely consumes the input iterables, keeping pools of values in memory to generate the products. Accordingly, it is only useful with finite inputs.



v = [1, 2]
assert cartesian_product(v, v) |> list == [(1, 1), (1, 2), (2, 1), (2, 2)]


from itertools import product
v = [1, 2]
assert list(product(v, v)) == [(1, 1), (1, 2), (2, 1), (2, 2)]



Coconut’s multi_enumerate enumerates through an iterable of iterables. multi_enumerate works like enumerate, but indexes through inner iterables and produces a tuple index representing the index in each inner iterable. Supports indexing.

For numpy objects, uses np.nditer under the hood. Also supports len for numpy arrays.



>>> [1, 2;; 3, 4] |> multi_enumerate |> list
[((0, 0), 1), ((0, 1), 2), ((1, 0), 3), ((1, 1), 4)]


array = [[1, 2], [3, 4]]
enumerated_array = []
for i in range(len(array)):
    for j in range(len(array[i])):
        enumerated_array.append(((i, j), array[i][j]))


groupsof(n, iterable, fillvalue=...)

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. If that is not the desired behavior, fillvalue can be passed and will be used to pad the end of the last tuple to length n.

Additionally, groupsof supports len when iterable supports len.



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


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


windowsof(size, iterable, fillvalue=..., step=1)

windowsof produces an iterable that effectively mimics a sliding window over iterable of size size. step determines the spacing between windows.

If size is larger than iterable, windowsof will produce an empty iterable. If that is not the desired behavior, fillvalue can be passed and will be used in place of missing values. Also, if fillvalue is passed and the length of the iterable is not divisible by step, fillvalue will be used in that case to pad the last window as well. Note that fillvalue will only ever appear in the last window.

Additionally, windowsof supports len when iterable supports len.



assert "12345" |> windowsof$(3) |> list == [("1", "2", "3"), ("2", "3", "4"), ("3", "4", "5")]

Python: Can’t be done without the definition of windowsof; see the compiled header for the full definition.


all_equal(iterable, to=...)

Coconut’s all_equal built-in takes in an iterable and determines whether all of its elements are equal to each other.

If to is passed, all_equal will check that all the elements are specifically equal to that value, rather than just equal to each other.

Note that all_equal assumes transitivity of equality, that != is the negation of ==, and that empty arrays always have all their elements equal.

Special support is provided for numpy objects.



all_equal([1, 1, 1])
all_equal([1, 1, 2])


sentinel = object()
def all_equal(iterable):
    first_item = sentinel
    for item in iterable:
        if first_item is sentinel:
            first_item = item
        elif first_item != item:
            return False
    return True
all_equal([1, 1, 1])
all_equal([1, 1, 2])


tee(iterable, n=2)

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

Though tee is not deprecated, reiterable is generally recommended over tee.

Custom tee/reiterable implementations for custom Containers/Collections should be put in the __copy__ method. Note that all Sequences/Mappings/Sets are always assumed to be reiterable even without calling __copy__.

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)


consume(iterable, keep_last=0)

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 a sequence (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)

Built-Ins for Parallelization#

process_map and thread_map#

process_map(function, *iterables, *, chunksize=1, strict=False, stream=False, ordered=True)#

Coconut provides a multiprocessing-based version of map under the name process_map. process_map makes use of multiple processes, and is therefore much faster than map for CPU-bound tasks. If any exceptions are raised inside of process_map, a traceback will be printed as soon as they are encountered. Results will be in the same order as the input unless ordered=False.

process_map never loads the entire input iterator into memory, though by default it does consume the entire input iterator as soon as a single output is requested. Results can be streamed one at a time when iterating by passing stream=True, however note that stream=True requires that the resulting iterator only be iterated over inside of a process_map.multiple_sequential_calls block (see below).

Because process_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 process_map occur inside of an if __name__ == "__main__" guard.

process_map supports a chunksize argument, which determines how many items are passed to each process at a time. Larger values of chunksize are recommended when dealing with very long iterables. Additionally, in the multi-iterable case, strict can be set to True to ensure that all iterables are the same length.

Deprecated: parallel_map is available as a deprecated alias for process_map. Note that deprecated features are disabled in --strict mode.


If multiple sequential calls to process_map need to be made, it is highly recommended that they be done inside of a with process_map.multiple_sequential_calls(): block, which will cause the different calls to use the same process pool and result in process_map immediately returning a list rather than a process_map object. If multiple sequential calls are necessary and the laziness of process_map is required, then the process_map objects should be constructed before the multiple_sequential_calls block and then only iterated over once inside the block.

process_map.multiple_sequential_calls also supports a max_workers argument to set the number of processes. If max_workers=None, Coconut will pick a suitable max_workers, including reusing worker pools from higher up in the call stack.

thread_map(function, *iterables, *, chunksize=1, strict=False, stream=False, ordered=True)#

Coconut provides a multithreading-based version of process_map under the name thread_map. thread_map and thread_map.multiple_sequential_calls behave identically to process_map except that they use multithreading instead of multiprocessing, and are therefore primarily useful only for IO-bound tasks due to CPython’s Global Interpreter Lock.

Deprecated: concurrent_map is available as a deprecated alias for thread_map. Note that deprecated features are disabled in --strict mode.

Python Docs#

process_map(func, *iterables, chunksize=1)

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.

process_map chops the iterable into a number of chunks which it submits to the process pool as separate tasks. The (approximate) size of these chunks can be specified by setting chunksize to a positive integer. For very long iterables using a large value for chunksize can make the job complete much faster than using the default value of 1.

thread_map(func, *iterables, chunksize=1)

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.

thread_map chops the iterable into a number of chunks which it submits to the thread pool as separate tasks. The (approximate) size of these chunks can be specified by setting chunksize to a positive integer. For very long iterables using a large value for chunksize can make the job complete much faster than using the default value of 1.



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


import functools
from multiprocessing import Pool
with Pool() as pool:
    print(list(pool.imap(functools.partial(pow, 2), range(100))))


thread_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())))

collectby and mapreduce#

collectby(key_func, iterable, value_func=None, *, reduce_func=None, collect_in=None, reduce_func_init=..., map_using=None)#

collectby(key_func, iterable) collects the items in iterable into a dictionary of lists keyed by key_func(item).

If value_func is passed, instead collects value_func(item) into each list instead of item.

If reduce_func is passed, instead of collecting the items into lists, reduce over the items of each key with reduce_func, effectively implementing a MapReduce operation. If keys are intended to be unique, set reduce_func=False (reduce_func=False is also the default if collect_in is passed). If reduce_func is passed, then reduce_func_init may also be passed, and will determine the initial value when reducing with reduce_func.

If collect_in is passed, initializes the collection from collect_in rather than as a collections.defaultdict (if reduce_func=None) or an empty dict (otherwise). Additionally, reduce_func defaults to False rather than None when collect_in is passed. Useful when you want to collect the results into a pandas.DataFrame.

If map_using is passed, calculates key_func and value_func by mapping them over the iterable using map_using as map. Useful with process_map/thread_map. See .using_threads and .using_processes methods below for simple shortcut methods that make use of map_using internally.

collectby is similar to itertools.groupby except that collectby aggregates common elements regardless of their order in the input iterable, whereas groupby only aggregates common elements that are adjacent in the input iterable.

mapreduce(key_value_func, iterable, *, reduce_func=None, collect_in=None, reduce_func_init=..., map_using=None)#

mapreduce(key_value_func, iterable) functions the same as collectby, but allows calculating the keys and values together in one function. key_value_func must return a 2-tuple of (key, value).

collectby.using_threads(key_func, iterable, value_func=None, *, reduce_func=None, collect_in=None, reduce_func_init=..., ordered=False, chunksize=1, max_workers=None)#
collectby.using_processes(key_func, iterable, value_func=None, *, reduce_func=None, collect_in=None, reduce_func_init=..., ordered=False, chunksize=1, max_workers=None)#
mapreduce.using_threads(key_value_func, iterable, *, reduce_func=None, collect_in=None, reduce_func_init=..., ordered=False, chunksize=1, max_workers=None)#
mapreduce.using_processes(key_value_func, iterable, *, reduce_func=None, collect_in=None, reduce_func_init=..., ordered=False, chunksize=1, max_workers=None)#

These shortcut methods call collectby/mapreduce with map_using set to process_map/thread_map, properly managed using the .multiple_sequential_calls method and the stream=True argument of process_map/thread_map. reduce_func will be called as soon as results arrive, and by default in whatever order they arrive in (to enforce the original order, pass ordered=True).

To make multiple sequential calls to collectby.using_threads()/mapreduce.using_threads(), manage them using thread_map.multiple_sequential_calls(). Similarly, use process_map.multiple_sequential_calls() to manage .using_processes().

Note that, for very long iterables, it is highly recommended to pass a value other than the default 1 for chunksize.

As an example, mapreduce.using_processes is effectively equivalent to:

def mapreduce.using_processes(key_value_func, iterable, *, reduce_func=None, ordered=False, chunksize=1, max_workers=None):
    with process_map.multiple_sequential_calls(max_workers=max_workers):
        return mapreduce(


user_balances = (
    |> collectby$(.user, value_func=.balance, reduce_func=(+))


from collections import defaultdict

user_balances = defaultdict(int)
for item in balance_data:
    user_balances[item.user] += item.balance


async_map(async_func, *iters, strict=False)

async_map maps async_func over iters asynchronously using anyio, which must be installed for async_func to work. strict functions as in map/zip, enforcing that all the iters must have the same length.

Equivalent to:

async def async_map[T, U](
    async_func: async T -> U,
    *iters: T$[],
    strict: bool = False
) -> U[]:
    """Map async_func over iters asynchronously using anyio."""
    import anyio
    results = []
    async def store_func_in_of(i, args):
        got = await async_func(*args)
        results.extend([None] * (1 + i - len(results)))
        results[i] = got
    async with anyio.create_task_group() as nursery:
        for i, args in enumerate(zip(*iters, strict=strict)):
            nursery.start_soon(store_func_in_of, i, args)
    return results


async def load_pages(urls) = (
    |> async_map$(load_page)
    |> await


import anyio

async def load_pages(urls):
    results = [None] * len(urls)
    async def proc_url(i, url):
        results[i] = await load_page(url)
    async with anyio.create_task_group() as nursery:
        for i, url in enumerate(urls)
          nursery.start_soon(proc_url, i, url)
    return results

Typing-Specific Built-Ins#


The TYPE_CHECKING variable is set to False at runtime and True during type_checking, allowing you to prevent your typing imports and TypeVar definitions from being executed at runtime. By wrapping your typing imports in an if TYPE_CHECKING: block, you can even use the typing module on Python versions that don’t natively support it. Furthermore, TYPE_CHECKING can also be used to hide code that is mistyped by default.

Python Docs#

A special constant that is assumed to be True by 3rd party static type checkers. It is False at runtime. Usage:

    import expensive_mod

def fun(arg: expensive_mod.SomeType) -> None:
    local_var: expensive_mod.AnotherType = other_fun()


    from typing import List
x: List[str] = ["a", "b"]
    def factorial(n: int) -> int: ...
    def factorial(0) = 1
    addpattern def factorial(n) = n * factorial(n-1)


    from typing import TYPE_CHECKING
except ImportError:

    from typing import List
x: List[str] = ["a", "b"]
    from typing import TYPE_CHECKING
except ImportError:

    def factorial(n: int) -> int: ...
    def factorial(n):
        if n == 0:
            return 1
            return n * factorial(n-1)

reveal_type and reveal_locals#

When using MyPy, reveal_type(<expr>) will cause MyPy to print the type of <expr> and reveal_locals() will cause MyPy to print the types of the current locals(). At runtime, reveal_type(x) is always the identity function and reveal_locals() always returns None. See the MyPy documentation for more information.



> coconut --mypy
Coconut Interpreter vX.X.X:
(enter 'exit()' or press Ctrl-D to end)
>>> reveal_type(fmap)
<function fmap at 0x00000239B06E2040>
<string>:17: note: Revealed type is 'def [_T, _U] (func: def (_T`-1) -> _U`-2, obj: typing.Iterable[_T`-1]) -> typing.Iterable[_U`-2]'


    from typing import TYPE_CHECKING
except ImportError:

    def reveal_type(x):
        return x

from coconut.__coconut__ import fmap

Coconut API#


coconut.embed(kernel=None, depth=0, **kwargs)

If kernel=False (default), embeds a Coconut Jupyter console initialized from the current local namespace. If kernel=True, launches a Coconut Jupyter kernel initialized from the local namespace that can then be attached to. The depth indicates how many additional call frames to ignore. kwargs are as in IPython.embed or IPython.embed_kernel based on kernel.

Recommended usage is as a debugging tool, where the code from coconut import embed; embed() can be inserted to launch an interactive Coconut shell initialized from that point.

Automatic Compilation#

Automatic compilation lets you simply import Coconut files directly without having to go through a compilation step first. Automatic compilation can be enabled either by importing coconut.api before you import anything else, or by running coconut --site-install.

Once automatic compilation is enabled, Coconut will check each of your imports to see if you are attempting to import a .coco file and, if so, automatically compile it for you. Note that, for Coconut to know what file you are trying to import, it will need to be accessible via sys.path, just like a normal import.

Automatic compilation always compiles with --target sys --line-numbers --keep-lines by default. On Python 3.4+, automatic compilation will use a __coconut_cache__ directory to cache the compiled Python. Note that __coconut_cache__ will always be removed from __file__.

Automatic compilation is always available in the Coconut interpreter or when using coconut-run. When using auto compilation through the Coconut interpreter, any compilation options passed in will also be used for auto compilation. Additionally, the interpreter always allows importing from the current working directory, letting you easily compile and play around with a .coco file simply by running the Coconut interpreter and importing it.

If using the Coconut interpreter, a reload built-in is always provided to easily reload (and thus recompile) imported modules.

Coconut Encoding#

While automatic compilation is the preferred method for dynamically compiling Coconut files, as it caches the compiled code as a .py file to prevent recompilation, Coconut also supports a special

# coding: coconut

declaration which can be added to .py files to have them treated as Coconut files instead. To use such a coding declaration, you’ll need to either run coconut --site-install or import coconut.api at some point before you first attempt to import a file with a # coding: coconut declaration. Like automatic compilation, the Coconut encoding is always available from the Coconut interpreter. Compilation always uses the same parameters as in the Coconut Jupyter kernel.


In addition to enabling automatic compilation, coconut.api can also be used to call the Coconut compiler from code instead of from the command line. See below for specifications of the different api functions.

Deprecated: coconut.convenience is a deprecated alias for coconut.api.



Gets a state object which stores the current compilation parameters. State objects can be configured with setup or cmd and then used in parse or other endpoints.

If state is None, gets a new state object, whereas if state is False, the global state object is returned.


coconut.api.parse(code="", mode="sys", state=False, keep_internal_state=None)

Likely the most useful of the api functions, parse takes Coconut code as input and outputs the equivalent compiled Python code. mode is used to indicate the context for the parsing and state is the state object storing the compilation parameters to use as obtained from get_state (if False, uses the global state object). keep_internal_state determines whether the state object will keep internal state (such as what custom operators have been declared)—if None, internal state will be kept iff you are not using the global state.

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

  • "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

  • "lenient":

    • parser: lenient

      • Can parse any Coconut code, allows leading whitespace, and has no trailing newline.

    • header: none

  • "xonsh":

from coconut.api import parse
while True:
    exec(parse(input(), mode="block"))


coconut.api.setup(target=None, strict=False, minify=False, line_numbers=True, keep_lines=False, no_tco=False, no_wrap=False, *, state=False)

setup can be used to set up the given state object with the given compilation parameters, each corresponding to the command-line flag of the same name. target should be either None for the default target or a string of any allowable target.

If state is False, the global state object is used.


coconut.api.warm_up(streamline=True, enable_incremental_mode=False, *, state=False)

Can optionally be called to warm up the compiler and get it ready for parsing. Passing streamline will cause the warm up to take longer but will substantially reduce parsing times (by default, this level of warm up is only done when the compiler encounters a large file). Passing enable_incremental_mode will enable the compiler’s incremental mdoe, where parsing some string, then later parsing a continuation of that string, will yield substantial performance improvements.


coconut.api.cmd(args=None, *, argv=None, interact=False, default_target=None, default_jobs=None, state=False)

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, argv can be used to pass in arguments as in --argv and default_target can be used to set the default --target.

Has the same effect of setting the command-line flags on the given state object as setup (with the global state object used when state is False).


coconut.api.cmd_sys(args=None, *, argv=None, interact=False, default_target="sys", default_jobs="0", state=False)

Same as coconut.api.cmd but default_target is "sys" rather than None (universal) and default_jobs="0" rather than None ("sys"). Since cmd_sys defaults to not using multiprocessing, it is preferred whenever that might be a problem, e.g. if you’re not inside an if __name__ == "__main__" block on Windows.


coconut.api.coconut_exec(expression, globals=None, locals=None, state=False, keep_internal_state=None)

Version of exec which can execute Coconut code.


coconut.api.coconut_eval(expression, globals=None, locals=None, state=False, keep_internal_state=None)

Version of eval which can evaluate Coconut code.


coconut.api.auto_compilation(on=True, args=None, use_cache_dir=None)

Turns automatic compilation on or off. This function is called automatically when coconut.api is imported.

If args is passed, it will set the Coconut command-line arguments to use for automatic compilation. Arguments will be processed the same way as with coconut-run such that --quiet --target sys --keep-lines will all be set by default.

If use_cache_dir is passed, it will turn on or off the usage of a __coconut_cache__ directory to put compile files in rather than compiling them in-place. Note that __coconut_cache__ will always be removed from __file__.



Switches the breakpoint built-in which Coconut makes universally available to use coconut.embed instead of pdb.set_trace (or undoes that switch if on=False). This function is called automatically when coconut.api is imported.

find_packages and find_and_compile_packages#

coconut.api.find_packages(where=".", exclude=(), include=("*",))

coconut.api.find_and_compile_packages(where=".", exclude=(), include=("*",))

Both functions behave identically to setuptools.find_packages, except that they find Coconut packages rather than Python packages. find_and_compile_packages additionally compiles any Coconut packages that it finds in-place.

Note that if you want to use either of these functions in your, you’ll need to include coconut as a build-time dependency in your pyproject.toml. If you want setuptools to package your Coconut files, you’ll also need to add global-include *.coco to your and pass include_package_data=True to setuptools.setup.

# if you put this in your, your Coconut package will be compiled in-place whenever it is installed

from setuptools import setup
from coconut.api import find_and_compile_packages




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 api function, a CoconutException instance may be raised. coconut.api.CoconutException is provided to allow catching such errors.


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 process_map