pkg-deploy
Modern Python Package Deployment Tool
What is it
pkg-deploy is a comprehensive Python package deployment tool that streamlines the process of building, versioning, and publishing Python packages to PyPI and private repositories. It supports both standard and Cython builds, automatic version management, and seamless Git integration.
Features
- Automatic Version Management: Support for semantic versioning with patch, minor, major, alpha, beta, and release candidate bumps
- Flexible Build System: Standard Python builds and optimized Cython compilation
- Code Minification: Optionally shrink the code before Cython compilation
- Source Leak Check: Cython wheels are rejected before upload if any
.pyor intermediate C source slipped in - Multiple Repository Support: Deploy to PyPI, private Nexus repositories, and custom package indexes
- Git Integration: Automatic tagging and commit management
- Environment Detection: Native support for both pip and uv virtual environments
- Dry Run Mode: Build the wheel without publishing it, bumping the version, or touching Git
- Interactive Authentication: Secure credential input with API token support
- Automatic Cleanup: Clean removal of build artifacts after deployment
Installation
From PyPI
pip install pkg-deploy
Quick Start
Basic Usage
After installing pkg-deploy , navigate to your project directory (the folder containing your pyproject.toml ) and use the pkg-deploy command directly. Example project structure:
my-package/
├── src/
│ └── my_package/
│ ├── __init__.py
│ └── main.py
├── pyproject.toml
└── README.md
Running the Deployment
Use the pkg-deploy command with your desired arguments:
# Deploy with patch version bump to PyPI
pkg-deploy --repository-name pypi --version-type patch
# Deploy to private repository with minor version bump
pkg-deploy --repository-url https://nexus.example.com/repository/pypi-internal/ \
--username admin --password secret --version-type minor
# Dry run to test configuration
pkg-deploy --repository-name pypi --version-type patch --dry-run
Configuration
pyproject.toml
pyproject.toml is required. pkg-deploy refuses to start without one, and it must declare
[project].name and [project].version — the version is read from and written back to this file
only, so a setup.py cannot stand in for it.
Ensure your pyproject.toml includes the required project metadata:
[build-system]
requires = ["setuptools>=70"]
build-backend = "setuptools.build_meta"
[project]
name = "your-package-name"
version = "1.0.0"
description = "Your package description"
requires-python = ">=3.10"
authors = [
{ name = "Your Name", email = "your.email@example.com" },
]
readme = "README.md"
dependencies = [
"dependency1",
"dependency2",
]
[tool.setuptools]
package-dir = {"" = "src"}
zip-safe = false
[tool.setuptools.packages.find]
where = ["src"]
# Optional: extra files to bump alongside pyproject.toml (otherwise only pyproject.toml changes)
[[tool.bumpversion.file]]
filename = "src/your_package/__init__.py"
search = '__version__ = "{current_version}"'
replace = '__version__ = "{new_version}"'
Package Directory Resolution
pkg-deploy automatically resolves the package directory using the following priority:
- Explicit --package-dir: If specified via command line, this path is used directly
- pyproject.toml Configuration: Automatically parsed from setuptools configuration:
tool.setuptools.packages.find.where- source directory locationtool.setuptools.package-dir- package directory mapping
- Default Fallback: Uses
project_dir/package_namebased on the project name
Example configurations in pyproject.toml:
# Standard src layout
[tool.setuptools]
package-dir = {"" = "src"}
[tool.setuptools.packages.find]
where = ["src"]
# Custom package location
[tool.setuptools]
package-dir = {"my_package" = "custom/path"}
Both the src layout and the flat layout (package directly under the project root, no
src/) are supported, including for Cython builds. pkg-deploy tells them apart by whether the
resolved directory contains an __init__.py: if it does, it is the package itself and its parent
is treated as the source root; otherwise it is a container such as src/. The resolved directory
must be inside the project directory.
⚠️ Important: When package directory is configured in multiple places within
pyproject.toml(such as bothtool.setuptools.packages.find.whereandtool.setuptools.package-dir), all configurations must point to the same directory. If they differ,pkg-deploywill raise aValueErrorwith the message "Package directory from toml are not the same".
.pypirc Configuration
For repository authentication, create a .pypirc file in your user home directory:
- Unix/macOS:
~/.pypirc - Windows:
C:\Users\username\.pypirc
[distutils]
index-servers =
pypi
private-repo
[pypi]
repository = https://upload.pypi.org/legacy/
username = __token__
password = pypi-your-api-token-here
[private-repo]
repository = https://nexus.example.com/repository/pypi-internal/
username = your-username
password = your-password
Usage Examples
Version Management
# Patch version bump (1.0.0 -> 1.0.1)
pkg-deploy --repository-name pypi --version-type patch
# Minor version bump (1.0.1 -> 1.1.0)
pkg-deploy --repository-name pypi --version-type minor
# Major version bump (1.1.0 -> 2.0.0)
pkg-deploy --repository-name pypi --version-type major
# Use specific version
pkg-deploy --repository-name pypi --new-version 2.1.0
# Pre-release cycle. From a final release, alpha/beta/rc start a cycle on the
# NEXT patch version, because PEP 440 sorts 1.2.4a1 before 1.2.4 and a pre-release
# published after its final would be invisible to pip.
pkg-deploy --repository-name pypi --version-type alpha # 1.2.4 -> 1.2.5a1
pkg-deploy --repository-name pypi --version-type alpha # 1.2.5a1 -> 1.2.5a2
pkg-deploy --repository-name pypi --version-type beta # 1.2.5a2 -> 1.2.5b1
pkg-deploy --repository-name pypi --version-type rc # 1.2.5b1 -> 1.2.5rc1
pkg-deploy --repository-name pypi # 1.2.5rc1 -> 1.2.5 (patch finalises)
Rules for pre-releases:
- Stages only move forward (
alpha→beta→rc); any can be skipped, none can be revisited.1.2.5rc1 --version-type alphais refused — use--new-versionto force a value. patch(the default) on a pre-release drops the suffix:1.2.5a1,1.2.5b1and1.2.5rc1all finalise to1.2.5.minor/majoron a pre-release abandon the cycle:1.2.5rc1→1.3.0/2.0.0.- A pre-release of the next minor (
1.3.0a1) is not inferred — pass it explicitly with--new-version 1.3.0a1.
Cython Builds
# Build with Cython optimization
pkg-deploy --repository-name pypi --version-type patch --cython
# Cython build for private repository
pkg-deploy --repository-url https://nexus.example.com/repository/pypi-internal/ \
--username user_name --password secret --cython
# Cython build with code minification
pkg-deploy --repository-name pypi --version-type patch --cython --minify
How setup.py is handled
A Cython build needs a setup.py. pkg-deploy decides what to do based on what it finds in
the project directory:
| Situation | Behaviour |
|---|---|
No setup.py |
One is generated for the build and deleted again during cleanup |
setup.py exists and contains cythonize |
Yours is used as-is and never overwritten |
setup.py exists without cythonize |
The build stops with FileExistsError — back it up or migrate it to pyproject.toml |
pkg-deploy also rewrites [build-system] in pyproject.toml for the duration of the build
(adding setuptools/Cython to requires and forcing build-backend to
setuptools.build_meta), then restores it afterwards. Any version bump you asked for survives
the restore; only the injected build-system entries are removed.
The Cython version actually used comes from your [build-system].requires. If you want a
specific one, pin it there — the setup_requires line in the generated setup.py is a legacy
field and does not control it.
Multiple Python Version Builds with cibuildwheel
pkg-deploy can hand the build to cibuildwheel, which produces one wheel per Python version
listed in your [tool.cibuildwheel] configuration in a single run.
cibuildwheel builds for the operating system it is running on: on Linux it produces manylinux
wheels (inside Docker), on macOS macOS wheels, on Windows Windows wheels. To cover several
platforms, run pkg-deploy --cibuildwheel once on each — it does not cross-compile.
Benefits
- Multiple Python Versions: One run yields a wheel for every version in
build = "cp38-* cp39-* ..." - Proper Platform Tags: Wheels are tagged and (on Linux) audited for manylinux compatibility
Requirements
- Docker (Linux only): Docker must be installed and running when building on Linux
- cibuildwheel: Installed together with
pkg-deploy - Sufficient disk space: One build environment per Python version
Configuration
Add cibuildwheel configuration to your pyproject.toml (Optional):
[tool.cibuildwheel]
# Specify which Python versions to build for
build = "cp38-* cp39-* cp310-* cp311-* cp312-* cp313-*"
before-build = "pip install Cython"
[tool.cibuildwheel.linux]
# Use manylinux images for better compatibility
before-all = "yum install -y gcc || apt-get update && apt-get install -y gcc"
[tool.cibuildwheel.macos]
# Ensure Cython is available for macOS builds
before-build = "pip install Cython"
[tool.cibuildwheel.windows]
# Ensure Cython is available for Windows builds
before-build = "pip install Cython"
Usage Examples
# Cython build for every configured Python version
pkg-deploy --repository-name pypi --version-type patch --cython --cibuildwheel
# Same, to a private repository
pkg-deploy --repository-url https://nexus.example.com/repository/pypi-internal/ \
--username admin --password secret --cython --cibuildwheel
# Dry run to check the cibuildwheel configuration
pkg-deploy --repository-name pypi --cython --cibuildwheel --dry-run
Limitations
- Build Time: Each Python version is compiled separately, so runs take proportionally longer
- Docker Dependency: Linux builds require Docker to be installed and running
- One platform per run: Wheels for other operating systems need a run on that system
Advanced Options
# Custom project directory
pkg-deploy --project-dir /path/to/project --repository-name pypi
# Custom package directory
pkg-deploy --package-dir /path/to/package --repository-name pypi
# Publish without recording the release in Git (version bump is reverted)
pkg-deploy --repository-name pypi --new-version 1.2.3 --discard-version-bump
# Verbose logging
pkg-deploy --repository-name pypi --verbose
# Dry run with verbose output
pkg-deploy --repository-name pypi --dry-run --verbose
Command Line Interface
Arguments
--project-dir: Project directory (default: current directory)--package-dir: Package directory path (default: auto-resolved from pyproject.toml or project name)--version-type, -vt: Version bump type: patch, minor, major, alpha, beta, or rc--new-version, -v: Specify exact version number (overrides version-type). Must matchMAJOR.MINOR.PATCHwith an optionalaN/bN/rcNsuffix —1.2,1.2.3.post1and1.2.3dev1are rejected--cython, -c: Enable Cython compilation for performance--minify, -m: Minify the code before compilation to reduce its size. Must be used together with--cython--cibuildwheel: Build with cibuildwheel — one wheel per configured Python version, for the platform you run it on (requires Docker on Linux)--repository-name, -rn: Repository name from .pypirc configuration (e.g., 'pypi', 'testpypi')--repository-url, -ru: Repository upload URL (prompts for username/password if not in .pypirc)--username, -u: Authentication username (optional if configured in .pypirc)--password, -p: Authentication password/token (optional if configured in .pypirc)--discard-version-bump: Leave the repository untouched — no bump commit, no tag, no push, and the version bump written topyproject.tomlis reverted after a successful upload. The published version is then recorded nowhere in the repo, so the next deploy resolves to that same version again and the upload clashes; pair it with--new-version--skip-git-status-check: Skip Git status validation before deployment--dry-run: Build the wheel and log what would be uploaded, without publishing it, bumping the version, or touching Git. The build itself really runs, sodist/andbuild/are created and then cleaned up--verbose, -V: Debug logging, including the full output of the build tool (which packages the isolated build environment installed, which modules were cythonized, compiler warnings)
Environment Support
UV Virtual Environments
pkg-deploy detects a uv-managed environment by looking for a uv = ... marker in the
pyvenv.cfg of the active interpreter. When it finds one it builds with uv build --wheel
(installing uv first if it is missing); otherwise it falls back to python -m build --wheel.
Passing --cibuildwheel overrides both and runs cibuildwheel instead.
Either way the build runs in an isolated environment that installs your
[build-system].requires fresh, so that is where build-time pins such as Cython>=3.2 take
effect. Note that uv caches resolutions — run uv cache clean if you need to be certain you
picked up the newest matching release.
Security Considerations
API Tokens (Recommended)
For PyPI, use API tokens instead of passwords:
- Generate token at https://pypi.org/manage/account/token/
- Use
__token__as username - Use the generated token as password
Credential Storage
- Store credentials in
.pypircfor reusability - Credentials are handed to twine through the
TWINE_USERNAME/TWINE_PASSWORDenvironment variables of the child process, never on its command line, so they do not show up inpsor in any logged command - Leave
--username/--passwordoff the command line and letpkg-deployprompt for them instead — the prompt usesgetpass, so nothing is echoed or kept in your shell history
Troubleshooting
Common Issues
Git Repository Not Clean
Error: Git repo is NOT clean
Solution: Commit or stash your changes before deployment, or use --skip-git-status-check.
Missing Dependencies
Error: Missing required packages: build, twine
Solution: Install the missing packages. build, twine and tomlkit are always
required; --cython additionally requires Cython.
Cython Build Conflicts
Error: Cannot build Cython code: setup.py already exists
Solution: Back up the existing setup.py, add a cythonize call to it so pkg-deploy uses it
as-is, or migrate its configuration to pyproject.toml. See
How setup.py is handled.
Source Leak Detected
ValueError: Source leak detected in built wheel(s) - refusing to upload.
Solution: The listed files should have been compiled but were not. Check that every directory
holding those modules has an __init__.py.
Contradictory Package Directory
ValueError: Package directory from toml are not the same: ['src', 'lib']
Solution: Make tool.setuptools.packages.find.where and tool.setuptools.package-dir point at
the same directory, or pass --package-dir to override both.
Build Failures
ValueError: Cython build failed,
stdout: ...
stderr: ...
Solution: pkg-deploy reports the underlying build tool's output verbatim — read the stderr
section first. For --cibuildwheel on Linux, check that Docker is installed and running,
verify your [tool.cibuildwheel] configuration, and make sure there is enough disk space.
pkg-deploy itself does not check for Docker, so a missing daemon surfaces as a cibuildwheel
error inside this output.
Authentication Failures
Error: 403 Forbidden
Solution: Verify credentials, or use API tokens for PyPI. Omitting --username / --password
makes pkg-deploy prompt for them interactively.
Debug Mode
Enable verbose logging for detailed troubleshooting:
pkg-deploy --repository-name pypi --verbose --dry-run
pkg-deploy - Streamlining Python package deployment since 2025.
Metadata
Release files for pkg-deploy 1.1.2
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Reset |
|---|---|---|---|---|
| pkg_deploy-1.1.2-py3-none-any.whl | Python 3 | none | any | Details |
Release files / pkg_deploy-1.1.2-py3-none-any.whl
| Download URL | pkg_deploy-1.1.2-py3-none-any.whl |
|---|---|
| Size | 25.7 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
53b7d935b5619969d32633ff26eb6e75c09bf2bebde6eaede88455b16e8cf8fb
|
|
BLAKE2b-256 checksum How to use checksums |
a7f19f280014e9dce0f9ba0081e68431a2594245f3396e2700bdbdaddd2e1a0a
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/6.1.0 CPython/3.13.5
|