private_attribute_cpp

True C++-style private attributes for Python, enforced at the C level.

A C++ extension module (private_attribute) that lets you declare attributes that are only accessible from inside the class that owns them — genuinely hidden from the outside world, from subclasses, and from hasattr(), dir(), and obj.__dict__. Unlike Python's name-mangling (_Class__name) and the leading-underscore convention, these attributes cannot be reached at all from outside the declaring class's own code.

Note: CPython only. Requires Python ≥ 3.10.

Table of contents

• Why?
• Installation
• Quick start
• Core concepts
  • Declaring private attributes
  • Custom name generation (private_func)
  • Decorated methods (PrivateWrapProxy)
  • Inheritance and use_parents
  • Cross-class access (PrivateSuper and PrivateClassSuper)
• API reference
• Custom metaclasses
• Behavioral notes & limitations
• Building from source
• Running the tests
• License

Why?

Python gives you a convention (_private) and a name-mangling trick (__name → _ClassName__name), but neither is an actual access barrier. Any code that knows (or can guess) the mangled name can still read and mutate your "private" state.

private_attribute implements the real thing in C++:

• Private attributes are not stored in __dict__.
• They are unreachable from outside the class that declares them — including from subclasses and from getattr/hasattr/dir.
• Inside the declaring class, they look and behave like ordinary attributes.

It does this by installing C-struct descriptors for __getattribute__ / __getattr__ / __setattr__ / __delattr__ / __del__ (and __call__ on registered metaclasses) and storing private values under generated, collision-resistant names (a tuple of two SHA-256 hashes).

Installation

pip install private-attribute-cpp

Prebuilt wheels are published for CPython 3.10–3.15 (including free-threaded t builds) across Linux, macOS, and Windows.

Quick start

from private_attribute import PrivateAttrBase

class BankAccount(PrivateAttrBase):
    __private_attrs__ = ["_balance"]

    def __init__(self, balance):
        self._balance = balance

    def deposit(self, amount):
        self._balance += amount      # ✅ accessible inside the class

    @property
    def balance(self):
        return self._balance

acct = BankAccount(100)
print(acct.balance)     # 100
acct.deposit(50)
print(acct.balance)     # 150

print(hasattr(acct, "_balance"))   # False — truly hidden
print(acct._balance)               # AttributeError: private attribute

Core concepts

Declaring private attributes

List the private names in the __private_attrs__ class attribute:

from private_attribute import PrivateAttrBase

class MyClass(PrivateAttrBase):
    __private_attrs__ = ["a", "b", "c"]   # string, or a sequence of strings

    def __init__(self):
        self.a = 1
        self.b = 2
        self.c = 3

    def show(self):
        print(self.a, self.b, self.c)     # (1, 2, 3) — visible inside

obj = MyClass()
obj.show()                                # 1 2 3
print(hasattr(obj, "a"))                  # False
print(hasattr(obj, "b"))                  # False
print(hasattr(obj, "c"))                  # False

• Private attributes may be set on instances or on the class itself (class-level private state, e.g. for @classmethod).
• They also cover private methods: declare a method name in __private_attrs__ and it becomes callable only from the class's own code.
• They work alongside __slots__ (see notes below).

Custom name generation (private_func)

By default private values are stored under generated hash names. You can control how the storage name is derived with a private_func callable, which receives the object id and the attribute name:

from private_attribute import PrivateAttrBase

def my_generate_func(obj_id, attr_name):
    return f"_hidden_{obj_id}_{attr_name}"

class MyClass(PrivateAttrBase, private_func=my_generate_func):
    __private_attrs__ = ["secret"]

    def __init__(self):
        self.secret = 42

    def read(self):
        return self.secret

private_func may be passed either as a keyword in the class definition (shown above) or on PrivateAttrBase itself. It is optional; the default generates a tuple of two SHA-256 hashes.

Decorated methods (PrivateWrapProxy)

Decorators that return a plain wrapper object (anything other than property / classmethod / staticmethod) would normally break private attribute access, because the wrapper's bytecode is not recognized as belonging to the class. PrivateWrapProxy makes arbitrary decorators compatible:

from private_attribute import PrivateAttrBase, PrivateWrapProxy

def deco(func):
    def wrapper(*args, **kwargs):
        return func(*args, **kwargs)
    return wrapper

class MyClass(PrivateAttrBase):
    __private_attrs__ = ["a", "b"]

    @PrivateWrapProxy(deco)
    def method1(self):
        return self.a

    @PrivateWrapProxy(deco)
    @PrivateWrapProxy(deco)
    def method2(self):
        return self.b

PrivateWrapProxy(decorator) returns a _PrivateWrap object that wraps the decorator application. It exposes two public attributes:

• .result — the final decorated result.
• .funcs — a read-only list proxy (_listproxy) of the original (pre-decoration) functions.

To stack decorators, resolve name conflicts, or fix decorator ordering, chain .result.<attr> onto the wrapped method and pass the method itself as the second argument (which saves the previous functions):

class MyClass(PrivateAttrBase):
    __private_attrs__ = ["a", "b", "c"]

    @PrivateWrapProxy(deco1)
    def method1(self):
        return self.a

    @PrivateWrapProxy(method1.result.name1, method1)
    @PrivateWrapProxy(method1.result.name2, method1)
    def method1(self):        # final implementation; earlier ones are kept
        return self.b

    @PrivateWrapProxy(deco2)
    def method2(self):
        return self.c

The _PrivateWrap object is transparent in normal use: it forwards __getattr__, __call__, and descriptor __get__ to its .result, so callers interact with the wrapped method as usual.

Inheritance and use_parents

Since v2.1.0, a subclass's code cannot access its parent's private attributes — a parent's private attribute is reachable only from the parent's own code, or from a class that re-declares the same name in its own __private_attrs__. When a subclass re-declares the same name, the two are stored separately (instance attributes per declaring class, class-level attributes per class).

Since v2.2.0, you can opt in to inheriting all ancestors' private names with use_parents=True:

class Parent(PrivateAttrBase):
    __private_attrs__ = ["_secret"]
    def __init__(self):
        self._secret = 1
    def get_secret(self):
        return self._secret

class Child(Parent, use_parents=True):
    __private_attrs__ = []
    def __init__(self):
        super().__init__()
        self._secret = 99        # stored in Child's own storage
    def read_own(self):
        return self._secret

obj = Child()
obj.get_secret()   # 1  — parent code sees the parent's value
obj.read_own()     # 99 — child code sees its own value

The merged names live in the subclass's own storage — they never overwrite the parent's values. The Python-level __private_attrs__ tuple is left unchanged; only the internal name set is merged. Default is False.

Cross-class access (PrivateSuper and PrivateClassSuper)

Since v3.0.0 you can explicitly reach a parent class's class-level private attribute — including the value it stored "for" an intermediate class — with PrivateSuper (instance flavour) or PrivateClassSuper (@classmethod flavour):

import private_attribute

class Grand(private_attribute.PrivateAttrBase):
    __private_attrs__ = ["_g"]
    _g = "grand"

    @classmethod
    def set_g(cls, value):
        cls._g = value

class Mid(Grand):
    __private_attrs__ = ["_m"]
    _m = "mid"

class Leaf(Mid, use_parents=True):
    __private_attrs__ = []

    def read_grand_shadow(self):
        # A=Leaf (running code), B=Grand (owning class), C=Mid (defined-for class)
        return private_attribute.PrivateSuper(Leaf, Grand, Mid, self)._g

    @classmethod
    def read_grand_shadow_cls(cls):
        return private_attribute.PrivateClassSuper(Leaf, Grand, Mid, cls)._g

Mid.set_g("written-on-mid")
Leaf().read_grand_shadow()      # "written-on-mid"
Leaf.read_grand_shadow_cls()    # "written-on-mid"

• A — the type whose code is currently running.
• B — a parent type of A whose class-level private attribute is accessed.
• C — the class the private value was defined for; it must be B, A, or a class that is both a subclass of B and a parent of A.
• the final argument is an instance of A (for PrivateSuper) or a subclass of A (for PrivateClassSuper).

The attribute is resolved from B's class-level storage for C and its __get__(obj, A) (or __get__(None, cls)) is invoked. This is for class-level private attributes only — it is not for instance-attribute cross-type access.

API reference

Class-definition options:

Custom metaclasses

To combine private attributes with another metaclass (e.g. ABCMeta), put a create_new slot directly in the metaclass. create_new takes care of registration and control automatically:

from abc import ABCMeta, abstractmethod
import private_attribute

class PrivateAbcMeta(ABCMeta):
    __new__ = private_attribute.create_new()

class MyClass(metaclass=PrivateAbcMeta):
    __private_attrs__ = ("_a",)
    __slots__ = ()

    @abstractmethod
    def my_function(self): ...

class MyImplement(MyClass):
    __private_attrs__ = ("_a",)
    def __init__(self, value=1):
        self._a = value
    def my_function(self):
        return self._a

a = MyImplement(1)
a.my_function()          # 1
a._a                     # AttributeError: private attribute
MyClass()                # TypeError: Can't instantiate abstract class ...

create_new also works as a decorator:

class PrivateAbcMeta(ABCMeta):
    @private_attribute.create_new
    def __new__(cls, name, bases, attrs, **kwargs):
        return super().__new__(cls, name, bases, attrs, **kwargs)

Custom special-method slots on the metaclass use create_meta_slot, and on a private-attribute class use create_slot. Both take the slot name first and work either as an empty slot or as a decorator:

class PrivateAbcMeta(ABCMeta):
    __new__ = private_attribute.create_new()
    __setattr__ = private_attribute.create_meta_slot("__setattr__")()

class MyClass(private_attribute.PrivateAttrBase):
    __setattr__ = private_attribute.create_slot("__setattr__")()

The old prepare / postprocess pair and register_metaclass are deprecated (they emit a warning) and kept only for backward compatibility; new code should use create_new.

Behavioral notes & limitations

• __private_attrs__ must be a string or a sequence of strings.
• Names declared in __slots__ cannot also be in __private_attrs__; however, private attributes may still be set in methods even when __slots__ is defined (they live in the private-attribute storage, not __dict__).
• Instances of PrivateAttrBase (and subclasses) are not picklable by default.
• Private names are ultimately rewritten to a tuple of two hashes (unless a custom private_func is supplied).
• _PrivateWrap objects are transparently unwrapped back to the original object during class creation.
• Do not pass PrivateWrapProxy a decorator that itself returns a _PrivateWrap — this raises TypeError.
• A class nested inside another class cannot use the outer class's private attributes.
• If a parent class defines an attribute that is not in its __private_attrs__ (nor a PrivateAttrType instance), a child class must not list that name in its own __private_attrs__.
• v2.1.0 — subclass code can no longer read/write a parent's private attributes; re-declaring the name stores it separately.
• v2.2.0 — a subclass must not define, in its own class body, a name that is private to a parent, unless it declares the name in its own __private_attrs__ or opts in with use_parents=True.
• v2.2.1 — prepare locks private-attribute-type creation so only one type is being built (type.__new__) at a time; postprocess unlocks it. The lock is released automatically on failure. (Deprecated; see create_new.)
• v2.2.3 — prefer the with form of prepare/postprocess to guarantee the lock is released. (Deprecated; see create_new.)
• CPython may reset the special-method slot descriptors (e.g. when you assign __getattribute__ / __setattr__). Use ensure_type (or ensure_metaclass) to re-inject them, and avoid setting those methods on private-attribute classes directly.
• v3 — private attribute interception is implemented through C-struct descriptors installed as __getattribute__ / __getattr__ / __setattr__ / __delattr__ / __del__ (and __call__ on registered metaclasses), rather than by directly setting tp_getattro / tp_setattro / tp_finalize. Only tp_traverse / tp_clear are set directly. A public cache keyed by object id records attribute names already confirmed non-private, so they go straight to the user callback without re-classification. Registered types and created types must be heap types (static types are rejected).
• register_metaclass requires a metaclass that supports weak references. (Deprecated; use create_new, which registers the metaclass automatically.)
• Do not set __static_attributes__ on a private-attribute class; it is removed automatically.

Building from source

The module is built with scikit-build-core and CMake (C++17). It vendors the header-only PicoSHA2 library for SHA-256 hashing.

git clone https://github.com/Locked-chess-official/private_attribute_cpp.git
cd private_attribute_cpp
pip install .

To build a wheel instead:

pip install build
python -m build

Set PRIVATE_ATTR_KEEP_DEBUG=1 to retain debug symbols in a Linux build (otherwise -g0 is applied).

Running the tests

pip install pytest
python -m pytest tests/ -v

The test suite covers attribute hiding, private methods, class-level private state, decorator wrapping, inheritance semantics (use_parents), ABC/metaclass integration, __slots__, and GC/cycle-collection regressions.

License

MIT © 2026 Locked-chess-official.

PicoSHA2 is distributed under the MIT License (© 2017 okdshin).