pyflow 0.1.1

A modern Python dependency manager
pyflow-0.1.1 is not a library. version Build Status


Simple is better than complex - The Zen of Python

This tool manages Python installations and dependencies. It implements PEP 582 -- Python local packages directory and Pep 518 -- Specifying Minimum Build System Requirements for Python Projects

It keeps the latter isolated in the project directory, and runs Python in an environment which uses this directory. Per PEP 582, dependencies are stored in the project directory → __pypackages__3.7(etc) → lib.

It also includes some convenience features, like running standalone files in their own environment with no config, and running project functions directly from the CLI.

You don't need Python installed to use it; it will install the specified version of Python if not already installed. This feature currently only works on Windows, Ubuntu≥18.04, and Debian. For now, if on another OS, assume you must have the version of Python you wish to use installed, and available on the PATH.

Only works with Python ≥ 3.4.

Goal: Make using and publishing Python projects as simple as possible. Understanding Python environments shouldn't be required to use dependencies safely. We're attempting to fix each stumbling block in the Python workflow, so that it's as elegant as the language itself.


  • Windows, Ubuntu, or Debian: Download and run this installer or this deb .

  • A different Linux distro: Download this standalone binary and place it somewhere accessible by the system PATH. For example, /usr/bin.

  • If you have Rust installed: Run cargo install pyflow.

  • Mac: Build from source using the instructions near the bottom of this page, or install via cargo. If able, please build and PR a Mac binary.


  • (Optional) Run pyflow init in an existing project folder, or pyflow new projname to create a new project folder. init imports data from requirements.txt or Pipfile; new creates a folder with the basics.
  • Run pyflow install to sync dependencies with pyproject.toml, or add dependencies to it. pyproject.toml will be created if it doesn't exist.
  • Run pyflow or pyflow to run Python.

Quick-and-dirty start for quick-and-dirty scripts

  • Add the line __requires__ = [numpy, requests] somewhere in the script, where numpy and requsts are dependencies. Run pyflow script, where is the name of your script. This will set up an isolated environment for this script, and install dependencies as required, without altering any other environment. This is a safe way to run one-off Python files that aren't attached to a project, but have dependencies.

Why add another Python manager?

Pipenv and Poetry both address part of this problem. Some reasons why this tool is different:

  • It automatically manages Python installations and environments. You specify a Python version in pyproject.toml (if ommitted, the pyflow asks), and pyflow ensures that version is used. If it's not installed, pyflow downloads a binary, and uses that. If multiple installations are found for that version, it asks which to use.

  • By not using Python to install or run, it remains intallation-agnostic and environment-agnostic. This is important for making setup and use as simple and decison-free as possible. It's especially important on Linux, where there may be several versions of Python installed, with different versions and access levels. This avoids complications, especially for new users. It's common for Python-based CLI tools to not run properly when installed from pip due to the PATH not being configured in the expected way.

  • Its dependency resolution and locking is faster due to using a cached database of dependencies, vice downloading and checking each package, or relying on the incomplete data available on the pypi warehouse.

  • It keeps dependencies in the project directory, in __pypackages__. This is subtle, but reinforces the idea that there's no hidden state to be concerned with.

  • It will always use the specified version of Python. This is a notable problem, for example, with Poetry; it may pick the wrong installation (eg Python2 vice Python3), with no obvious way to change it.

  • Multiple versions of a dependency can be installed, allowing resolution of conflicting sub-dependencies. (ie: Your package requires Dep A>=1.0 and Dep B. Dep B requires Dep A==0.9) There are many cases where Poetry and Pipenv will fail to resolve dependencies, but we're able to by doing this. Try it for yourself with a few random dependencies from pypi; there's a good change you'll hit this problem using Poetry or Pipenv. Limitations: This will not work for some compiled dependencies, and attempting to package something using this will trigger an error.

My OS comes with Python, and Virtual environments are easy. What's the point of this?

Hopefully we're not replacing one problem with another.

Some people like the virtual-environment workflow - it requires only tools included with Python, and uses few console commands to create, and activate and environments. However, it may be tedius depending on workflow: The commands may be long depending on the path of virtual envs and projects, and it requires modifying the state of the terminal for each project, each time you use it, which you may find inconvenient or inelegant.

If you're satisified with an existing flow, there may be no reason to change, but I think we can do better. This is especially relevant for new Python users who haven't groked venvs, or are unaware of the hazards of working with a system Python.

Pipenv improves the workflow by automating environment use, and allowing reproducable dependency resolution. Poetry improves upon Pipenv's API, speed, and dependency resolution, as well as improving the packaging and distributing process by using a consolidating project config. Both are sensitive to the Python environment used to run them. This tool attempts to improve upon both in the areas listed in the section above. Its goal is to be as intuitive as possible.

Conda addresses these problems elegantly, but maintains a separate repository of binaries from PyPi. If all packages you need are available on Conda, it may be the best solution. If not, it requires falling back to Pip, which means using two separate package managers.

When building and deploying packages, a set of degenerate files are traditionally used:, setup.cfg, requirements.txt and We use pyproject.toml as the single-source of project info required to build and publish.

A thoroughly biased feature table

These tools have different scopes and purposes:

Name Pip + venv Pipenv Poetry pyenv pythonloc Conda this
Manages dependencies
Manages Python installations
Included with Python
Stores packages with project
Locks dependencies
Requires changing session state
Easy script access
Clean build/publish flow
Supports old Python versions with virtualenv


  • Optionally, create a pyproject.toml file in your project directory. Otherwise, this file will be created automatically. You may wish to use pyproject new to create a basic project folder (With a .gitignore, source directory etc), or pyproject init to populate info from requirements.txt or Pipfile. See PEP 518 for details.

Example contents:

py_version = "3.7"
name = "runcible"
version = "0.1.0"
author = "John Hackworth"

numpy = "^1.16.4"
diffeqpy = "1.1.0"

The [tool.pyflow] section is used for metadata. The only required item in it is py_version, unless building and distributing a package. The [tool.pypyackage.dependencies] section contains all dependencies, and is an analog to requirements.txt.

You can specify extra dependencies, which will only be installed when passing explicit flags to pyflow install, or when included in another project with the appropriate flag enabled. Ie packages requiring this one can enable with pip install -e etc.

test = ["pytest", "nose"]
secure = ["crypto"]

If you'd like to an install a dependency with extras, use syntax like this:

ipython = { version = "^7.7.0", extras = ["qtconsole"] }

For details on how to specify dependencies in this Cargo.toml-inspired semvar format, reference this guide.

We also attempt to parse metadata and dependencies from tool.poetry sections of pyproject.toml, so there's no need to modify the format if you're using that.

You can specify direct entry points to parts of your program using something like this in pyproject.toml:

# ...
scripts = { name = "module:function" }

Where you replace name, function, and module with the name to call your script with, the function you wish to run, and the module it's in respectively. This is similar to specifying scripts in for built packages. The key difference is that functions specified here can be run at any time, without having to build the package. Run with pyflow scriptname to do this.

If you run pyflow package on on a package using this, the result will work like normal script entry points for somone using the package, regardless of if they're using this tool.

What you can do

Managing dependencies:

  • pyflow install - Install all packages in pyproject.toml, and remove ones not (recursively) specified. If an environment isn't already set up for the version specified in pyproject.toml, sets one up. If no version is specified, it asks you.
  • pyflow install requests - If you specify one or more packages after install, those packages will be added to pyproject.toml and installed.
  • pyflow install numpy==1.16.4 matplotlib>=3.1 - Example with multiple dependencies, and specified versions
  • pyflow uninstall requests - Remove one or more dependencies

Running REPL and Python files in the environment:

  • pyflow - Run a Python REPL
  • pyflow - Run a python file
  • pyflow python or pyflow python - alternate syntax for the above two
  • pyflow ipython, pyflow black etc - Run a CLI tool like ipython, or a project function. For the former, this must have been installed by a dependency; for the latter, it's specfied under [tool.pyflow], scripts
  • pyflow run ipython - alternate syntax for the above
  • pyflow script - Run a one-off script, outside a project directory, with per-file package management

Building and publishing:

  • pyflow package - Package for distribution (uses setuptools internally, and builds both source and wheel.)
  • pyflow package --features "test all" - Package for distribution with features enabled, as defined in pyproject.toml
  • pyflow publish - Upload to PyPi (Repo specified in pyproject.toml. Uses Twine internally.)


  • pyflow list - Display all installed packages and console scripts
  • pyflow new projname - Create a directory containing the basics for a project: a readme, pyproject.toml, .gitignore, and directory for code
  • pyflow init - Create a pyproject.toml file in an existing project directory. Pull info from requirements.text and Pipfile as required.
  • pyflow reset - Remove the environment, and uninstall all packages
  • pyflow clear - Clear the global cache of downloaded packages, eg in ~/.local/share/pyflow (Linux) or ~\AppData\Roaming\pyflow (Windows) and the global cache of one-off script environments, in ~/.local/share/pyflow/script-envs.
  • pyflow -V - Get the current version of this tool
  • pyflow help Get help, including a list of available commands

How installation and locking work

Running pyflow install syncs the project's installed dependencies with those specified in pyproject.toml. It generates pyflow.lock, which on subsequent runs, keeps dependencies each package a fixed version, as long as it continues to meet the constraints specified in pyproject.toml. Adding a package name via the CLI, eg pyflow install matplotlib simply adds that requirement before proceeding. pyflow.lock isn't meant to be edited directly.

Each dependency listed in pyproject.toml is checked for a compatible match in pyflow.lock If a constraint is met by something in the lock file, the version we'll sync will match that listed in the lock file. If not met, a new entry is added to the lock file, containing the highest version allowed by pyproject.toml. Once complete, packages are installed and removed in order to exactly meet those listed in the updated lock file.

This tool downloads and unpacks wheels from pypi, or builds wheels from source if none are availabile. It verifies the integrity of the downloaded file against that listed on pypi using SHA256, and the exact versions used are stored in a lock file.

When a dependency is removed from pyproject.toml, it, and its subdependencies not also required by other packages are removed from the __pypackages__ folder.

How dependencies are resolved

Compatible versions of dependencies are determined using info from the PyPi Warehouse (available versions, and hash info), and the pydeps database. We use pydeps, which is built specifically for this project, due to inconsistent dependency information stored on pypi. A dependency graph is built using this cached database. We attempt to use the newest compatible version of each package.

If all packages are either only specified once, or specified multiple times with the same newest-compatible version, we're done resolving, and ready to install and sync.

If a package is included more than once with different newest-compatible versions, but one of those newest-compatible is compatible with all requirements, we install that one. If not, we search all versions to find one that's compatible.

If still unable to find a version of a package that satisfies all requirements, we install multiple versions of it as-required, store them in separate directories, and modify their parents' imports as required.

Note that it may be possible to resolve dependencies in cases not listed above, instead of installing multiple versions. Ie we could try different combinations of top-level packages, check for resolutions, then vary children as-required down the hierarchy. We don't do this because it's slow, has no guarantee of success, and involves installing older versions of packages.


  • Installing from sources other than pypi (eg repos)
  • Installing multiple versions of a dependency may not work if it uses compiles code
  • The lock file is missing some info like hashes
  • Adding a dependency via the CLI with a specific version constraint, or extras.
  • Developer requirements
  • Packaging and publishing projects that use compiled extensions
  • Dealing with multiple-installed-versions of a dependency that uses importlib or dynamic imports

Building and uploading your project to PyPi

In order to build and publish your project, additional info is needed in pyproject.toml, that mimics what would be in Example:

name = "everythingkiller"
py_version = "3.6"
version = "0.1.0"
author = "Fraa Erasmas"
author_email = "raz@edhar.math"
description = "Small, but packs a punch!"
homepage = "https://everything.math"
repository = ""
license = "MIT"
classifiers = [
    "Topic :: System :: Hardware",
    "Topic :: Scientific/Engineering :: Human Machine Interfaces",
scripts = { activate = "jeejah:activate" }

numpy = "^1.16.4"
manim = "0.1.8"
ipython = {version = "^7.7.0", extras=["qtconsole"]}

Building this from source

If you’d like to build from source, download and install Rust, clone the repo, and in the repo directory, run cargo build --release.

Ie on Linux:

curl -sSf | sh
git clone
cd pyflow
cargo build --release


If installed via Cargo, run cargo install pyflow --force.


If you notice unexpected behavior or missing features, please post an issue, or submit a PR. If you see unexpected behavior, it's probably a bug! Post an issue listing the dependencies that did not install correctly.

Why not to use this

  • It's adding another tool to an already complex field.
  • Most of the features here are already provided by a range of existing packages, like the ones in the table above.
  • The field of contributers is expected to be small, since it's written in a different language.
  • Dependency managers like Pipenv and Poetry work well enough for many cases, and have dedicated dev teams, and large userbases.
  • Conda in particular handles many things this does quite well.

Dependency cache repo:

  • Github Example API calls:, This pulls all top-level dependencies for the requests package, and the dependencies for version 2.21.0 respectively. There is also a POST API for pulling info on specified versions. The first time this command is run for a package/version combo, it may be slow. Subsequent calls, by anyone, should be fast. This is due to having to download and install each package on the server to properly determine dependencies, due to unreliable information on the pypi warehouse.

Python binary sources:


  • Make sure the pyflow binary is accessible in your path. If installing via a deb, msi, or Cargo, this should be set up automatically.
  • Make sure __pypackages__ is in your .gitignore file.