Oxford RSE

The conda-pypi plugin available from conda 26.5

20 July 2026 by Timothy Poon7 minutes

Introduction

When I recently updated conda to 26.5, I was greeted by the following message:

  Did you know? You can install many PyPI packages with conda
  using the conda-pypi beta. Get started:
    https://docs.conda.io/projects/conda/en/stable/new-features.html

A plugin conda-pypi is now available in beta! The message suggests that, can conda install and pip install finally work together safely? This makes me excited as mixing conda install and pip install in a conda environment usually causes confusion when the dependencies are resolved. To find out more about it, I went to the documentation of conda-pypi to see what it is capable of and how it works.

How did I use pip within a conda environment

One of the most popular best practices around the usage of conda has been to avoid mixing it with pip. However, sometimes you have no choice as the package you want to install is not available via a conda channel or simply it is not up-to-date with the corresponding one in PyPI. I frequently helped scientists in setting standalone conda environments so they can launch a particular application in an isolated environment. Since a lot of applications were developed in-house, it is not surprising that they are not available in a conda channel so I will need to pip install inside a conda environment. This was what I did:

  1. Check what dependencies the package needs by performing a dry run: pip install --dry-run <PKG_TO_INSTALL>.
  2. Manually copy the dependencies (with their exact resolved versions) and conda install them.
  3. After installing all the dependencies via conda, run pip install --no-deps <PKG_TO_INSTALL> to install the actual package.

One of my colleagues rightly questioned my approach, saying why didn’t you just pip install <PKG_TO_INSTALL> after creating the conda environment if the environment will not be touched anymore. Perhaps I was being overly cautious and wanted to minimise the mixing of conda and pip for the sake of ‘purity’: it is satisfying to see all the installed packages are coming from the same channel after doing conda list. Other strategies also exist as documented in this conda-pypi page.

Why mixing pip install and conda install is usually bad

It is actually worth understanding why mixing conda install and pip install is generally a bad idea. Perhaps not surprising to know, conda and pip are two independent package managers and they record dependencies differently: conda has its own way of resolving dependencies and record those metadata by JSON files in the conda-meta directory; pip records the metadata of every installed package in separate directories with the format <NAME>-<VERSION>.dist-info (see here for the detailed specification) in the site-packages directory, e.g. numpy-2.5.0.dist-info.

A good thing about conda is that it also creates .dist-info directories in site-packages for each package installed via conda install, so pip can know what conda has installed because it understands .dist-info. However, pip install does not put anything in conda-meta so conda has no record about what pip has installed. This is the reason why touching a conda environment (e.g. conda install other packages or conda update) after pip install is generally a bad idea as conda only looks at what are inside conda-meta but not .dist-info, and this often results in a mismatch between the version reported in conda list and the actual imported one, or even you can import a ‘ghost’ package that is not shown in conda list!

How conda-pypi solves the issue

With the plugin conda-pypi enabled, conda will resolve dependencies in all your specified conda channels and the conda-pypi ‘channel’, which makes packages in PyPI available via conda install. Depending on the order of your channels and the availability, packages will be installed either from a conda channel or wheels directly from PyPI.

The important difference from a pure pip install inside a conda environment is that even if the wheels are coming straight from PyPI, it is conda which manages the installation and the recording of metadata, i.e. a JSON file will be created in conda-meta. conda operations afterwards such as conda update, conda install another package or conda remove will respect both metadata from packages installed from conda install and PyPI. No more manual dependencies installation or worry about conflicting installation within a conda environment if I use conda install and pip install!

A demo with NumPy

To understand how the conda-pypi plugin works, I experimented with installing numpy by pip install inside a conda environment to see if it delivers in practice. I would like to see how it works if I want to install numpy with some custom build options that can be specified during pip install and how I can make use of conda-pypi for a smoother experience of package management.

The default build options of numpy, not surprisingly, target a wide range of CPU architecture and it is usually quite optimised. However in some situations, you may want to turn off the optimisation for testing and debugging purposes (or just to appreciate the effort people put in for optimisation).

As a lot of scientific packages depend on numpy and if you have pip install-ed your own numpy in a conda environment, it will be a nightmare if you conda install anything after this. So I hope by using conda-pypi, I can now build my own numpy and then install any package afterwards from various conda channels.

Build and install NumPy by pip

First, set up your conda environment:

# cython etc. are build dependencies for NumPy
conda create -n npy-pip python=3.13 cython compilers openblas meson-python pkg-config

Then build and install numpy by pip:

python -m pip install --no-cache-dir numpy --no-build-isolation --no-binary numpy

Now if I inspect the output of numpy.show_config() on my machine (Apple M2):

  "SIMD Extensions": {
    "baseline": [
      "NEON",
      "NEON_FP16",
      "NEON_VFPV4",
      "ASIMD"
    ],
    "found": [
      "ASIMDHP",
      "ASIMDDP"
    ],
    "not found": [
      "ASIMDFHM"
    ]
  }

Build and install NumPy by pip with optimisation off

Create another conda environment with the same build dependencies as above, then build and install numpy by setting disable-optimization to true:

python -m pip install -Csetup-args=-Ddisable-optimization="true" --no-cache-dir numpy --no-build-isolation --no-binary numpy

In this installation, there are no SIMD extensions as shown in numpy.show_config().

Build and install NumPy with optimisation off by using conda-pypi

Both of the above approaches used pip install and if you look at conda-meta, there is no numpy-....json which records the installation and further conda operations that depend on the records in conda-meta would not know the existence of numpy with optimisation off, which are installed by pip.

How about using conda-pypi to build and install numpy with the optimisation off? After creating a fresh conda environment as above, you can first build the wheel without installing by pip:

python -m pip wheel -w wheels -Csetup-args=-Ddisable-optimization="true" --no-cache-dir numpy --no-build-isolation --no-binary numpy

The -w wheels puts the numpy wheel inside the wheels directory and you can use wherever you fancy. One of the features provided by conda-pypi allows you to convert wheels into .conda format which you can then use it for local installation.

# the name of the NumPy wheel will most likely be different in your case
conda pypi convert wheels/numpy-2.5.0-cp313-cp313-macosx_26_0_arm64.whl

# conda-pypi saves the converted .conda in ./conda-pypi-output by default
conda install conda-pypi-output/numpy-2.5.0-pypi_0.conda

As the installation is performed by conda, you will see something like numpy-....json inside conda-meta which means further conda operations will respect the existence of your numpy built with optimisation off! From now on, if you install packages that depend on numpy via conda install, such as scipy, it won’t install numpy from any conda channel as conda knows its existence.

Conclusion

We discussed the new conda-pypi plugin which is available from conda>=26.5 and its motivation, which is to provide a smoother experience when mixing pip install and conda install in a conda environment. We showed a demo that uses conda-pypi to install a custom built numpy into a conda environment and illustrated this approach would let conda respect the existence of it.

As of July 2026, conda-pypi is still in beta so I would not recommend using it in production. I think the best practice about not mixing conda and pip packages still holds, and whenever possible, one should install packages from a single conda channel. However, conda-pypi provides a much smoother workflow if you find yourself installing something via pip in a conda environment and it is definitely worth keeping an eye on its development.

Resources

conda pip package management