Skip to main content
Version: 2.10 (deprecated)

Third-party dependencies

How to use third-party Python libraries in your project.

Pants handles dependencies with more precision than traditional Python workflows. Traditionally, you have a single heavyweight virtual environment that includes a large set of dependencies, whether or not you actually need them for your current task.

Instead, Pants understands exactly which dependencies every file in your project needs, and efficiently uses just that subset of dependencies needed for the task.

❯ ./pants dependencies src/py/

❯ ./pants dependencies --transitive src/py/

Among other benefits, this precise and automatic understanding of your dependencies gives you fine-grained caching. This means, for example, that if none of the dependencies for a particular test file have changed, the cached result can be safely used.

First, turn off old-style macros

If you're starting a new project, set the below. This will become the default in Pants 2.11.

use_deprecated_python_macros = false

If you already are using Pants, follow the instructions Pants prints when upgrading to Pants 2.10 to upgrade to the new mechanism when ready.

Teaching Pants your "universe"(s) of dependencies

For Pants to know which dependencies each file uses, it must first know which specific dependencies are in your "universe", i.e. all the third-party dependencies your project directly uses.

By default, Pants uses a single universe for your whole project, but it's possible to set up multiple. See the header "Multiple resolves" in the "Lockfiles" section.

Each third-party dependency you directly use is modeled by a python_requirement target:


You do not need a python_requirement target for transitive dependencies, i.e. requirements that you do not directly import.

To minimize boilerplate, Pants has target generators to generate python_requirement targets for you:

  • python_requirements for requirements.txt.
  • poetry_requirements for Poetry projects.


The python_requirements() target generator parses a requirements.txt-style file to produce a python_requirement target for each entry.

For example:

dataclasses ; python_version<'3.7'

If the file uses a different name than requirements.txt, set source like this:

Where should I put the requirements.txt?

You can name the file whatever you want, and put it wherever makes the most sense for your project.

In smaller repositories that only use Python, it's often convenient to put the file at the "build root" (top-level), as used on this page.

For larger repositories or multilingual repositories, it's often useful to have a 3rdparty or 3rdparty/python directory. Rather than the target's address being //:reqs#my_requirement, its address would be 3rdparty/python:reqs#my_requirement, for example; or 3rdparty/python#my_requirement if you leave off the name field for python_requirements. See Target Generation.


The poetry_requirements() target generator parses the Poetry section in pyproject.toml to produce a python_requirement target for each entry.

python = "^3.8"
requests = {extras = ["security"], version = "~1"}
flask = "~1.12"

isort = "~5.5"

See the section "Lockfiles" below for how you can also hook up poetry.lock to Pants.

How dependencies are chosen

Once Pants knows about your "universe"(s) of dependencies, it determines which subset should be used through dependency inference. Pants will read your import statements, like import django, and map it back to the relevant python_requirement target. Run ./pants dependencies path/to/ or ./pants dependencies path/to:target to confirm this works.

If dependency inference does not work—such as because it's a runtime dependency you do not import—you can explicitly add the python_requirement target to the dependencies field, like this:

# We don't have an import statement for this dep, so inference
# won't add it automatically. We add it explicitly instead.

Use modules and module_mapping when the module name is not standard

Some dependencies expose a module different than their project name, such as beautifulsoup4 exposing bs4. Pants assumes that a dependency's module is its normalized name—i.e. My-distribution exposes the module my_distribution. If that default does not apply to a dependency, it will not be inferred.

Pants already defines a default module mapping for some common Python requirements, but you may need to augment this by teaching Pants additional mappings:

# `modules` and `module_mapping` is only needed for requirements where
# the defaults do not work.


module_mapping={"my_distribution": ["custom_module"]},

module_mapping={"my_distribution": ["custom_module"]},

If the dependency is a type stub, and the default does not work, set type_stub_modules on the python_requirement target, and type_stubs_module_mapping on the python_requirements and poetry_requirements target generators. (The default for type stubs is to strip off types-, -types, -stubs, and stubs-. So, types-requests gives type stubs for the module requests.)

Warning: multiple versions of the same dependency

It's invalid in Python to have conflicting versions of the same requirement, e.g. Django==2 and Django==3. Instead, Pants supports "multiple resolves" (i.e. multiple lockfiles), as explained in the below section on lockfiles.

When you have multiple targets for the same dependency and they belong to the same resolve ("lockfile"), dependency inference will not work due to ambiguity. If you're using lockfiles—which we strongly recommend—the solution is to set the resolve field for problematic python_requirement targets so that each resolve has only one requirement and there is no ambiguity.

This ambiguity is often a problem when you have 2+ requirements.txt or pyproject.toml files in your project, such as project1/requirements.txt and project2/requirements.txt both specifying django. You may want to set up each poetry_requirements/python_requirements target generator to use a distinct resolve so that there is no overlap. Alternatively, if the versions are the same, you may want to consolidate the requirements into a common file.


We strongly recommend using lockfiles because they make your builds more stable so that new releases of dependencies will not break your project. They also reduce the risk of supply chain attacks.

Pants has two types of lockfiles:

  • User lockfiles, for your own code such as packaging binaries and running tests.
  • Tool lockfiles, to install tools that Pants runs like Pytest and Flake8.

With both types of lockfiles, Pants can generate the lockfile for you with the generate-lockfiles goal. However, there are several situations where this does not work properly, and you may need to generate the lockfile manually. This will be improved in future Pants versions. See the below section for more information.

User lockfiles

First, set [python].enable_resolves in pants.toml:

enable_resolves = true

By default, Pants will write the lockfile to 3rdparty/python/default.lock. If you want a different location, change [python].resolves like this:

enable_resolves = true
resolves = { python-default = "lockfile_path.txt" }

Then, use ./pants generate-lockfiles to generate the lockfile.

❯ ./pants generate-lockfiles
19:00:39.26 [INFO] Completed: Generate lockfile for python-default
19:00:39.29 [INFO] Wrote lockfile for the resolve `python-default` to 3rdparty/python/default.lock

Alternatively, if you are manually generating the lockfile, set [python].resolves_generate_lockfiles, and point [python].resolves to the path of your lockfile. Pants will still consume it like normal, only it will not manage it for you such as checking when it needs to be regenerated.

enable_resolves = true
resolves_generate_lockfiles = false
resolves = { python-default = "lockfile_path.txt" }
FYI: user lockfiles improve performance

As explained at the top of these docs, Pants only uses the subset of the "universe" of your dependencies that is actually needed for a build, such as running tests and packaging a wheel file. This gives fine-grained caching and has other benefits like built packages (e.g. PEX binaries) only including their true dependencies. However, this also means that you may need to resolve dependencies multiple times, which can be slow.

If you use lockfiles, Pants will optimize to only resolve your requirements one time for your project. Then, for each build, Pants will extract from that resolve the exact subset needed.

This greatly speeds up performance and improves caching for goals like test, run, package, and repl.

Multiple lockfiles

While it's often desirable to have a single lockfile for the whole project for simplicity and consistency, sometimes you may need multiple. This is necessary, for example, when you have conflicting versions of requirements, such as part of your code using Django 2 and other parts using Django 3.

Start by defining multiple "resolves", which are logical names for lockfile paths. For example:

enable_resolves = true
default_resolve = "web-app"

data-science = "3rdparty/python/data_science_lock.txt"
web-app = "3rdparty/python/web_app_lock.txt"

Then, teach Pants which resolves every python_requirement target belongs to through the resolve field. It will default to [python].default_resolve.


# Often, you will want to set `resolve` on the
# `poetry_requirements` and `python_requirements`
# target generators.
# You can use `overrides` if you only want to change
# some targets.
overrides={"requests": {"resolve": "web-app"}},

If you want the same requirement to show up in multiple resolves, you currently need to create a distinct target per resolve. This will be improved in Pants 2.11 through a new parametrize() mechanism.

# The same requirement in multiple resolves:


# Note that because BUILD files are Python, you could de-duplicate
# this by defining variables. You can also add a
# macro:

Then, run ./pants generate-lockfiles to generate the lockfiles. If the results aren't what you'd expect, adjust the prior step.

Finally, update your first-party targets like python_source / python_sources, python_test / python_tests, and pex_binary to set their resolve field. As before, the resolve field defaults to [python].default_resolve.


# You can use `overrides` to change certain generated targets
overrides={"": {"resolve": "data-science"}},


If a first-party target is compatible with multiple resolves—such as some utility code—you must for now create one target per resolve. This will be improved with Pants 2.11's parametrize feature.

All transitive dependencies of a target must use the same resolve. Pants's dependency inference already handles this for you by only inferring dependencies on targets that share the same resolve. If you incorrectly add a target from a different resolve to the dependencies field, Pants will error with a helpful message when building your code with goals like test, package, and run.

Tool lockfiles

Pants distributes a lockfile with each tool by default. However, if you change the tool's version and extra_requirements—or you change its interpreter constraints to not be compatible with our default lockfile—you will need to use a custom lockfile. Set the lockfile option in pants.toml for that tool, and then run ./pants generate-lockfiles.

version = "flake8==3.8.0"
lockfile = "3rdparty/flake8_lockfile.txt" # This can be any path you'd like.

extra_requirements.add = ["pytest-icdiff"]
lockfile = "3rdparty/pytest_lockfile.txt"
❯  ./pants generate-lockfiles
19:00:39.26 [INFO] Completed: Generate lockfile for flake8
19:00:39.27 [INFO] Completed: Generate lockfile for pytest
19:00:39.29 [INFO] Wrote lockfile for the resolve `flake8` to 3rdparty/flake8_lockfile.txt
19:00:39.30 [INFO] Wrote lockfile for the resolve `pytest` to 3rdparty/pytest_lockfile.txt

You can also run ./pants generate-lockfiles --resolve=tool, e.g. --resolve=flake8, to only generate that tool's lockfile rather than generating all lockfiles.

To manually manage lockfiles, set the option [tool].lockfile to your lockfile path like normal. Do not run the generate-lockfiles goal, and also set [python].invalid_lockfile_behavior so that Pants does not look for its metadata header:

invalid_lockfile_behavior = "ignore"

To disable lockfiles entirely for a tool, set [tool].lockfile = "<none>" for that tool. Although we do not recommend this!

generate-lockfiles goal vs manual lockfile generation

generate-lockfiles limitations

Categorically, the generate-lockfiles goal cannot yet handle three use cases:

  • Does not support [python-repos] if you have a custom index or repository other than PyPI.
  • Does not support [GLOBAL].ca_certs_path.
  • Does not support VCS (Git) requirements and local file requirements.

If you use any of these three features for a certain lockfile, unfortunately, you must manually generate that lockfile. Support for these use cases is coming in future Pants releases by teaching Pex to generate lockfiles via pip.

Several users have also had issues with generate-lockfiles returning a lockfile that gets generated successfully, but then errors due to missing transitive dependencies when Pants tries to install it. This is especially common with user lockfiles. For example:

Failed to resolve requirements from PEX environment @ /home/pantsbuild/.cache/pants/named_caches/pex_root/unzipped_pexes/42735ba5593c0be585614e50072f765c6a45be15.
Needed manylinux_2_28_x86_64-cp-37-cp37m compatible dependencies for:
1: colorama<0.5.0,>=0.4.0
Required by:
FingerprintedDistribution(distribution=rich 11.0.0 (/home/pantsbuild/.cache/pants/named_caches/pex_root/installed_wheels/4ce6259e437af26bac891ed2867340d4163662b9/rich-11.0.0-py3-none-any.whl), fingerprint='ff22612617b194af3cd95380174413855aad7240')
But this pex had no 'colorama' distributions.

Usually, the transitive dependency is in the lockfile, but it doesn't get installed because it has nonsensical environment markers, like this:

colorama==0.4.4; sys_platform == "win32" and python_version >= "3.6" and python_full_version >= "3.6.2" and python_full_version < "4.0.0" and (python_version >= "3.6" and python_full_version < "3.0.0" or python_full_version >= "3.5.0" and python_version >= "3.6") and (python_version >= "3.6" and python_full_version < "3.0.0" and sys_platform == "win32" or sys_platform == "win32" and python_version >= "3.6" and python_full_version >= "3.5.0") and (python_version >= "3.6" and python_full_version < "3.0.0" and platform_system == "Windows" or python_full_version >= "3.5.0" and python_version >= "3.6" and platform_system == "Windows")

For user lockfiles, the workaround is to treat the problematic transitive dependencies as direct inputs to the resolve by creating a python_requirement target, which usually causes the lockfile generator to handle things correctly. For example:

# This turns off dependency inference for these
# requirements, which you may want to do as they
# are transitive dependencies that should not be directly imported.
# If you are using multiple resolves, you may need to set the
# `resolve` field.

For tool lockfiles, add the problematic transitive dependencies to [tool].extra_requirements. For example:

version = "pylint>=2.11.0,<2.12"
extra_requirements.add = ["colorama"]

Then, regenerate the lock with generate-lockfiles.

You can also try manually removing the problematic environment markers, although you will need to remember to do this again whenever re-running generate-lockfiles.

Manual lockfile generation techniques

Pants is agnostic to how your lockfile is generated, as long as it's a valid requirements.txt-style file.

Users have had success with these three techniques to generate their user lockfiles:

venv + pip freezeCreate a script like the one below. If you have multiple resolves, run this script once per resolve.The lockfile will not have --hash, which is less secure for supply chain attacks. This does allow you to use VCS (Git) requirements, however.

The lockfile may not work on platforms and Python versions other than what was used to create the virtual env.
pip-compilepip-compile --generate-hashes --allow-unsafe -o lock.txt requirements.txtThe lockfile may not work on platforms and Python versions other than what was used to run pip-compile.

Will not capture any python_requirement targets declared explicitly in BUILD files or in pyproject.toml.

Does not account for multiple resolves.
Poetrypoetry export --dev -o lock.txtRequires that you are using Poetry for dependency management.

Will not capture any python_requirement targets declared explicitly in BUILD files or in requirements.txt.

Does not account for multiple resolves.

Script to manually generate a user lockfile via pip freeze:
#!/usr/bin/env bash

set -euo pipefail

# You can change these constants.

"${PYTHON_BIN}" -m venv "${VIRTUALENV}"
"${PIP}" install pip --upgrade
# Install all our requirements.txt, and also any 3rdparty
# dependencies specified outside requirements.txt, e.g. via a
# handwritten python_requirement_library target.
"${PIP}" install \
-r <(./pants dependencies :: |
xargs ./pants filter --target-type=python_requirement |
xargs ./pants peek |
jq -r '.[]["requirements"][]')
echo "# Generated by build-support/ on $(date)" > "${CONSTRAINTS_FILE}"
"${PIP}" freeze --all >> "${CONSTRAINTS_FILE}"

# If you are using multiple resolves, you will need to use JQ to filter to all
# requirements from a single resolve. For most resolves, use this JQ snippet:
# '.[] | select(.resolve == "my-resolve") | .["requirements"][]'
# If the resolve is the default, you must also add `or .resolve == null`, like this:
# '.[] | select(.resolve == "python-default" or .resolve == null) | .["requirements"][]'

Users have usually had more success using the generate-lockfiles goal to generate tool lockfiles, so no one has yet written a script to manually generate tool lockfiles. You can grab the requirements used by Pants, though, by inspecting ./pants help-advanced $tool. Or you can use ./pants help-all to get JSON that you can query with JQ, e.g. ./pants help-all | jq -r '.scope_to_help_info.isort.advanced'.

Advanced usage

Requirements with undeclared dependencies

Sometimes a requirement does not properly declare in its packaging metadata the other dependencies it depends on, so those will not be installed. It's especially common to leave off dependencies on setuptools, which results in import errors like this:

import pkg_resources
ModuleNotFoundError: No module named 'pkg_resources'

To work around this, you can use the dependencies field of python_requirement, so that anytime you depend on your requirement, you also bring in the undeclared dependency.

# First, make sure you have a `python_requirement` target for
# the undeclared dependency.


If you are using the python_requirements and poetry_requirements target generators, you can use the overrides field to do the same thing:

"mongomock": {"dependencies": [":reqs#setuptools"]},

Version control and local requirements

You might be used to using pip's proprietary VCS-style requirements for this, like git+ However, this proprietary format does not work with Pants.

Instead of pip VCS-style requirements:


Use direct references from PEP 440:

Django@ git+
Django@ git+
Django@ git+

You can also install from local files using PEP 440 direct references. You must use an absolute path to the file, and you should ensure that the file exists on your machine.

Django @ file:///Users/pantsbuild/prebuilt_wheels/django-3.1.1-py3-none-any.whl

Pip still works with these PEP 440-compliant formats, so you won't be losing any functionality by switching to using them.

Version control via SSH

When using version controlled direct references hosted on private repositories with SSH access:

target@ git+ssh://[email protected]:/myorg/myrepo.git@myhash may see errors like:

 Complete output (5 lines):
[email protected]: Permission denied (publickey).
fatal: Could not read from remote repository.
Please make sure you have the correct access rights
and the repository exists.

To fix this, Pants needs to be configured to pass relevant SSH specific environment variables to processes by adding the following to pants.toml:

env_vars.add = [

Custom repositories

If you host your own wheels at a custom index (aka "cheese shop"), you can instruct Pants to use it with the option indexes in the [python-repos] scope.

indexes.add = [""]

To exclusively use your custom index—i.e. to not use PyPI—use indexes = [..] instead of indexes.add = [..].

You can also add Python repositories with the option repos in the [python-repos] scope.

repos = ["https://your/repo/here"]

Indexes are assumed to have a nested structure (like, whereas repos are flat lists of packages.

Tip: use ./pants export to create a virtual environment for IDEs

See Setting up an IDE for more information on ./pants export. This will create a virtual environment for your user code for compatibility with the rest of the Python ecosystem, e.g. IDEs like Pycharm.