package ocaml-in-python

  1. Overview
  2. Docs
Effortless Python bindings for OCaml modules


Dune Dependency





Effortless Python bindings for OCaml modules

This library exposes all OCaml modules as Python modules, generating bindings on the fly.


  • OCaml >= 4.13

  • Python >= 3.7 (and >= 3.10 for pattern-matching support)


The package can be installed via opam:

  • opam install ocaml-in-python installs the latest release,

  • opam pin add -k path . && opam install ocaml-in-python executed in a clone of this repository installs the latest development version.

Once installed via opam, the package should be registered in the Python environment. There are two options:

  • either you register the package with pip using the following command:

pip install --editable "`opam var ocaml-in-python:lib`"
  • or you add the following definition to your environment:

export PYTHONPATH="`opam var share`/python/:$PYTHONPATH"


Standard library

A very simple mean to test that the bindings are working properly is to invoke the OCaml standard library from Python.

import ocaml
print( x : x + 1), [1, 2, 3]))
# => output: [2;3;4]

In the following example, we invoke the ref function from the OCaml standard library to create a value of type int ref (a mutable reference to an integer), and the following commands show that the reference can be mutated from Python (a reference is a record with a mutable field contents) and from OCaml (here by invoking the OCaml function incr).

>>> x = ocaml.ref(1, type=int)
>>> x
>>> x.contents = 2
>>> x
>>> ocaml.incr(x)
>>> x

OCaml module compiled on the fly

In the following example, we compile a module on the fly from Python.

import ocaml

m = ocaml.compile(r'''
  let hello x = Printf.printf "Hello, %s!\n%!" x

  type 'a tree = Node of { label : 'a; children : 'a tree list }

  let rec height (Node { label = _; children }) =
    1 + List.fold_left (fun accu tree -> max accu (height tree)) 0 children

  let rec of_list nodes =
    match nodes with
    | [] -> invalid_arg "of_list"
    | [last] -> Node { label = last; children = [] }
    | hd :: tl -> Node { label = hd; children = [of_list tl] }

# => output: Hello, world!

  m.Node(label=1, children=[m.Node(label=2, children=[])])))
# => output: 2

print(m.of_list(["a", "b", "c"]))
# => output: Node {label="a";children=[Node {label="b";children=[Node {label="c";children=[]}]}]}

except ocaml.Invalid_argument as e:
    # => output: Stdlib.Invalid_argument("of_list")

It is worth noticing that there is no need for type annotations: bindings are generated with respect to the interface obtained by type inference.

Requiring a library with findlib

In the following example, we call the OCaml library parmap from Python.

import ocaml


from ocaml import Parmap

  (lambda x : x + 1), Parmap.A([1, 2, 3]), ncores=2))
# => output: [2, 3, 4]

The function ocaml.require uses ocamlfind to load parmap. Bindings are generated as soon as ocaml.Parmap is accessed (in the example, at line from ocaml import Parmap). Parmap.A is one of the two constructors of the type Parmap.sequence.

Conversion rules

The generation of bindings is driven by the types exposed by the compiled module interfaces (*.cmi files): relying on the *.cmi files allows the bindings to cover most of the OCaml definitions (there are some limitations though, see below) and to use the inferred types for modules whose interface is not explicitly specified by a .mli file.

Built-in types

The following conversions are defined for built-in types:

  • OCaml int, nativeint int32, int64 are mapped to Python int;

import ocaml

# => output: 42
print(ocaml.int_of_string("5") + 1)
# => output: 6
  • OCaml string is mapped to Python str

import ocaml

ocaml.print_endline("Hello, World!")
# => output: Hello, World!
print(ocaml.String.make(3, "a") + "b")
# => output: aaab
  • OCaml char is mapped to Python str with a single character

import ocaml

# => output: 97
# => output: A
  • OCaml bool is mapped to Python bool (beware of different case convention: OCaml values false and true are mapped to Python values False and True respectively)

import ocaml

# => output: False
# => output: true
  • OCaml float is mapped to Python float, and functions taking floats as arguments can take benefit from the Python automatic coercion from int to float

import ocaml

# => output: 1.0
# => output: 1.0
  • OCaml array is mapped to a dedicated class ocaml.array, which supports indexing, enumeration, pattern-matching (with Python >= 3.10) and in-place modification. When an OCaml array is converted to a Python object, the elements are converted on demand. There is an implicit coercion to array from all Python iterable types such as Python lists (but in-place modification is lost).

import ocaml

arr = ocaml.Array.make(3, 0)
arr[1] = 1
print(ocaml.Array.fold_left((lambda x,y : x + y), 0, arr))
# => output: 1
ocaml.Array.sort(, arr)
# => output: [0, 0, 1]
print( x: x + 1), range(0, 4)))
# => output: [|1;2;3;4|]

# With Python 3.10:
match arr:
  case [0, 0, 1]:
# => output: Here

It is worth noticing that Array.make is a polymorphic function parameterized in the type of the elements of the constructed array, and by default the type parameter for polymorphic function with ocaml-in-python is Py.Object.t, the type of all Python objects. As such, the cells of the array arr defined above can contain any Python objects, not only integers.

arr[0] = "Test"
# => output: [|"Test";0;1|]

We can create an array with a specific types for cells by expliciting the type parameter of Array.make, by using the keyword parameter type.

arr = ocaml.Array.make(3, 0, type=int)
arr[0] = "Test"
# TypeError: 'str' object cannot be interpreted as an integer
  • OCaml list is mapped to a dedicated class ocaml.list, which supports indexing, enumeration and pattern-matching (with Python >= 3.10). When an OCaml list is converted to a Python object, the elements are converted on demand. There is an implicit coercion to list from all Python iterable types such as Python lists.

  • OCaml bytes is mapped to a dedicated class ocaml.bytes, which behaves as a mutable collection of characters.

  • OCaml option is mapped to a dedicated class ocaml.option, only for values of the form Some x where the type of x allows the value None. If the type of x does not contain a value None, the OCaml value Some x is mapped directly to the conversion of x. Conversely, the value Some x can be constructed with ocaml.Some(x). The OCaml value None is mapped to the Python value None.

print(ocaml.List.find_opt((lambda x : x > 1), [0,1], type=int))
# => output: None
print(ocaml.List.find_opt((lambda x : x > 1), [0,1,2], type=int))
# => output: 2
print(ocaml.List.find_opt((lambda x : x > 1), [0,1,2]))
# => output: Some(2)

In the last call to find_opt, the default type parameter is Py.Object.t which contains the value None.

  • OCaml exn is mapped to a dedicated class ocaml.exn, which is a sub-class of Python Error class, and exceptions are converted as other extension constructors: each exception is a sub-class of ocaml.exn, and values can be indexed (if the exception constructor takes parameters), accessed by field name (for inline records) and supports pattern-matching (with Python >= 3.10). There is an implicit coercion from other sub-classes of Python Error class to Py.E, the OCaml exception defined in pyml for Python exceptions. If an exception is raised between OCaml and Python code, the exception is converted and raised from one side to the other.

except ocaml.Failure as e:
# => output: Test
  • OCaml in_channel and out_channel are mapped to FileIO objects. In the following example, the OCaml function open_out is used to create a new file test, and the string Hello is written in this file through the Python method write. Then, the file test is opened for reading with the Python built-in open, and the channel is read with the OCaml function really_input_string.

with ocaml.open_out("test") as f:
with open("test", "r") as f:
   print(ocaml.really_input_string(f, 5))
# => ouput: Hello
  • OCaml floatarray is mapped to a dedicated class ocamlarray, which derives from numpy array (numpy is required for the support of floatarray).

Type constructors

  • OCaml functions of type 't_1 -> ... -> 't_n -> 'r are mapped to Python callable objects with n arguments. Labelled arguments are mapped to mandatory keyword arguments, and optional arguments are mapped to optional keyword arguments. For polymorphic functions, the type parameters are assumed to be Py.Object.t, except if there is a keyword argument type: the associated value can either be a single type if there is a single type parameter, or a tuple of types giving the type parameters in the order of their apparition in the function signature, or a dictionary whose keys are the names of the type parameters (e.g., "a" for 'a).

  • OCaml tuples of type 't_1 * ... * 't_n are mapped to OCaml tuples with n components.

Type definitions

Each OCaml type definition introduces a new Python class, except for type aliases, that are exposed as other names for the same class.

Records are accessible by field name or index (in the order of the field declarations), and the values of the fields are converted on demand. Mutable fields can be set in Python. In particular, the ref type defined in the OCaml standard library is mapped to the Python class ocaml.ref with a mutable field content. Records support pattern-matching (with Python >= 3.10). There is an implicit coercion from Python dictionaries with matching field names.

For variants, there is a sub-class by constructor, which behaves either as a tuple or as a record. The values of the arguments are converted on demand. Variants support pattern-matching (with Python >= 3.10).

Sub-module definitions

Sub-modules are mapped to classes, which are constructed on demand. For instance, the module Array.Floatarray is exposed as ocaml.Array.Floatarray, and, in particular, the function Array.Floatarray.create is available as ocaml.Array.Floatarray.create.


The following traits of the OCaml type system are not supported (yet):

  • records with polymorphic fields,

  • polymorphic variants,

  • objects,

  • functors,

  • first class modules.


Innovation. Community. Security.