Support installing with pip

[peastman]

The only packages provided for OpenFF-Toolkit are for conda. It would be really helpful to also provide pip packages. The ML ecosystem has standardized on pip rather than conda, and most MLIP packages are only available for pip (MACE, NequIP, AIMNet2, OrbMol, etc.). Mixing pip and conda packages in an environment often doesn't work and leads to a broken environment. This makes it difficult to use OpenFF-Toolkit with any MLIPs.

With a bit of work it's already possible to install OpenFF-Toolkit (or at least the parts of it that I need) with pip:

pip install git+https://github.com/openforcefield/openff-utilities.git
pip install git+https://github.com/openforcefield/openff-units.git
pip install git+https://github.com/openforcefield/openff-toolkit.git
pip install cachetools rdkit

Making this easier and supported would help a lot of people.

[j-wags]

We'd discussed this with our advisory board in late 2023 and decided not to move the Toolkit onto pypi. However some of the key blockers have now been resolved. Those were:

• Our only open-source AM1BCC charge assignment method (AmberTools) is not on PyPI (still true, but no longer a critical blocker since our newest FF uses NAGL charges. But people using anything except openff-2.3.0 will get an error that they need to install a non-pypi dep)
• OpenMM not completely supported on PyPI (now it is)
• Bespokefit deps not on PyPI (we're no longer updating bespokefit)

Remaining issues:

• It's a one-way choice: If we move to PyPI, it should be something we commit to in the long term, otherwise we risk having out of date zombie packages that confuse users
• If we do this, we should move all our major user-facing packages to pypi (not just toolkit), to avoid confusion if someone tries to eg. fetch a QC record using QCSubmit but that package is only available on conda.
• Some additional work supporting two distribution channels (mostly small, but headaches would be proportional to package divergence on conda vs pypi)

All this is to say: No decision on this yet, but in our current 2026-2027 roadmap planning multiple staff members proposed offering packages on pypi. I'll update here as those discussions progress.