Introduction

If you want your Python package to install TensorFlow automatically, the direct mechanism is to declare TensorFlow as a dependency in your package metadata. In older packaging layouts that means install_requires in setup.py, but in modern packaging the better place is usually pyproject.toml or setup.cfg because setup.py is no longer the preferred home for simple dependency declarations.

The basic setup.py approach

A traditional setup.py file can declare TensorFlow in install_requires so pip install your-package also installs TensorFlow.

1from setuptools import setup, find_packages
2
3setup(
4    name="my_project",
5    version="0.1.0",
6    packages=find_packages(),
7    install_requires=[
8        "tensorflow>=2.16,<3",
9        "numpy>=1.26",
10    ],
11)

That is enough for pip to understand that TensorFlow is part of the dependency graph.

Use optional extras when TensorFlow is not required for every user

TensorFlow is a heavy dependency. It increases install time, download size, and platform complexity. If only part of your package needs it, use an optional extra instead of forcing every user to install it.

1from setuptools import setup, find_packages
2
3setup(
4    name="my_project",
5    version="0.1.0",
6    packages=find_packages(),
7    install_requires=["numpy>=1.26"],
8    extras_require={
9        "tensorflow": ["tensorflow>=2.16,<3"],
10    },
11)

Then users can choose:

pip install .[tensorflow]

This is often the better design for libraries because it keeps the base installation lighter.

Modern packaging usually means pyproject.toml

For new projects, dependency declarations are often cleaner in pyproject.toml.

1[project]
2name = "my-project"
3version = "0.1.0"
4dependencies = [
5  "numpy>=1.26",
6]
7
8[project.optional-dependencies]
9tensorflow = [
10  "tensorflow>=2.16,<3",
11]

That gives the same result for pip users while following current Python packaging conventions more closely.

Do not expect packaging metadata to solve platform prerequisites

Declaring TensorFlow in dependencies tells pip what package to install. It does not solve every environment issue automatically. TensorFlow can still fail to install or run correctly if the Python version, operating system, CPU architecture, or accelerator libraries are incompatible.

That means "install automatically" has limits. Packaging metadata can request the dependency, but it cannot guarantee that every target machine satisfies TensorFlow's platform requirements.

Avoid shelling out from setup.py to run pip manually

A common anti-pattern is trying to run pip install tensorflow inside setup.py with subprocess or os.system. That is fragile and should be avoided.

Bad pattern:

1import os
2from setuptools import setup
3
4os.system("pip install tensorflow")
5setup(name="bad_example")

This breaks build isolation assumptions, hides dependency management from packaging tools, and makes installs harder to reproduce. Declare dependencies; do not manually run pip from package setup code.

Pin versions thoughtfully

TensorFlow is a fast-moving dependency with platform-specific wheels and occasional ecosystem friction. If your package depends on a specific API range, express that range clearly.

Good dependency hygiene usually means:

• set a lower bound you have tested
• avoid overly broad upper ranges if future major versions could break you
• separate optional ML features from core package behavior when practical

This matters more for TensorFlow than for tiny utility libraries because version mismatches can be expensive to debug.

Common Pitfalls

• Hard-coding a pip install tensorflow command inside setup.py instead of declaring a dependency.
• Forcing TensorFlow on every user when it should be an optional extra.
• Assuming dependency metadata can fix unsupported Python, OS, or hardware combinations.
• Using only setup.py in a new project when pyproject.toml would be cleaner.
• Leaving the TensorFlow version range too loose for a package that depends on specific APIs.

Summary

• To install TensorFlow automatically, declare it as a dependency rather than running pip manually.
• In legacy packaging, use install_requires in setup.py.
• In modern packaging, prefer pyproject.toml or optional extras when appropriate.
• TensorFlow is often better as an extra dependency than as a mandatory base dependency.
• Dependency metadata requests the install, but platform compatibility still has to be satisfied by the target environment.