Distlib is a pure-Python library. You should be able to install it using:

pip install distlib

for installing distlib into a virtualenv or other directory where you have write permissions. On POSIX platforms, you may need to invoke using sudo if you need to install distlib in a protected location such as your system Python’s site-packages directory.


A full test suite is included with distlib. To run it, you’ll need to download the source distribution, unpack it and run python tests/test_all.py in the top-level directory of the package.

If running the tests under Python >= 3.2.3, remember to first set the environment variable PYTHONHASHSEED=0 to disable hash randomisation, which is needed for the tests. (The environment variable also needs to be set if running Python 2.x with -R. which is only available in Python 2.6.8 and later.)

Continuous integration test results are available at:


Coverage results are available at:


Note that the actual coverage is higher than that shown, because coverage under Windows is not included in the above coverage figures.

Note that the index tests are configured, by default, to use a local test server, though they can be configured to run against PyPI itself. This local test server is not bundled with distlib, but is available from:


This is a slightly modified version of Ralf Schmitt’s pypiserver. To use, the script needs to be copied to the tests folder of the distlib distribution.

If the server script is not available, the tests which use it will be skipped. Naturally, this will also affect the coverage statistics.

PYPI availability

If PyPI is unavailable or slow, then some of the tests can fail or become painfully slow. To skip tests that might be sometimes slow, set the SKIP_SLOW environment variable:

$ SKIP_SLOW=1 PYTHONHASHSEED=0 python tests/test_all.py

on POSIX, or:

C:\> set SKIP_SLOW=1
C:\> python tests/test_all.py

on Windows.

First steps

For now, we just list how to use particular parts of the API as they take shape.

Using the database API

You can use the distlib.database package to access information about installed distributions. This information is available through the following classes:

Distribution paths

The Distribution and EggInfoDistribution classes are normally not instantiated directly; rather, they are returned by querying DistributionPath for distributions. To create a DistributionPath instance, you can do

>>> from distlib.database import DistributionPath
>>> dist_path = DistributionPath()

Querying a path for distributions

In this most basic form, dist_path will provide access to all non-legacy distributions on sys.path. To get these distributions, you invoke the get_distributions() method, which returns an iterable. Let’s try it:

>>> list(dist_path.get_distributions())

This may seem surprising if you’ve just started looking at distlib, as you won’t have any non-legacy distributions.

Including legacy distributions in the search results

To include distributions created and installed using setuptools or distribute, you need to create the DistributionPath by specifying an additional keyword argument, like so:

>>> dist_path = DistributionPath(include_egg=True)

and then you’ll get a less surprising result:

>>> len(list(dist_path.get_distributions()))

The exact number returned will be different for you, of course. You can ask for a particular distribution by name, using the get_distribution() method:

>>> dist_path.get_distribution('setuptools')
<EggInfoDistribution u'setuptools' 0.6c11 at '/usr/lib/python2.7/dist-packages/setuptools.egg-info'>

If you want to look at a specific path other than sys.path, you specify it as a positional argument to the DistributionPath constructor:

>>> from pprint import pprint
>>> special_dists = DistributionPath(['tests/fake_dists'], include_egg=True)
>>> pprint([d.name for d in special_dists.get_distributions()])

or, if you leave out egg-based distributions:

>>> special_dists = DistributionPath(['tests/fake_dists'])
>>> pprint([d.name for d in special_dists.get_distributions()])

Distribution properties

Once you have a Distribution instance, you can use it to get more information about the distribution. For example:

  • The metadata attribute gives access to the distribution’s metadata (see Using the metadata and markers APIs for more information).

  • The name_and_version attribute shows the name and version in the format name (X.Y).

  • The key attribute holds the distribution’s name in lower-case, as you generally want to search for distributions without regard to case sensitivity.

Exporting things from Distributions

Each distribution has a dictionary of exports. The exports dictionary is functionally equivalent to “entry points” in distribute / setuptools.

The keys to the dictionary are just names in a hierarchical namespace delineated with periods (like Python packages, so we’ll refer to them as pkgnames in the following discussion). The keys indicate categories of information which the distribution’s author wishes to export. In each such category, a distribution may publish one or more entries.

The entries can be used for many purposes, and can point to callable code or data. A common purpose is for publishing callables in the distribution which adhere to a particular protocol.

To give a concrete example, the Babel library for internationalisation support provides a mechanism for extracting, from a variety of sources, message text to be internationalised. Babel itself provides functionality to extract messages from e.g. Python and JavaScript source code, but helpfully offers a mechanism whereby providers of other sources of message text can provide their own extractors. It does this by providing a category 'babel.extractors', under which other software can register extractors for their sources. The Jinja2 template engine, for example, makes use of this to provide a message extractor for Jinja2 templates. Babel itself registers its own extractors under the same category, so that a unified view of all extractors in a given Python environment can be obtained, and Babel’s extractors are treated by other parts of Babel in exactly the same way as extractors from third parties.

Any installed distribution can offer up values for any category, and a set of distributions (such as the set of installed distributions on sys.path) conceptually has an aggregation of these values.

The values associated with a category are a list of strings with the format:

name = prefix [ ":" suffix ] [ "[" flags "]" ]

where name, prefix, and suffix are pkgnames. suffix and flags are optional and flags follow the description in Flag formats.

Any installed distribution can offer up values for any category, and a set of distributions (such as the set of installed distributions on sys.path) conceptually has an aggregation of these values.

For callables, the prefix is the package or module name which contains the callable, suffix is the path to the callable in the module, and flags can be used for any purpose determined by the distribution author (for example, the extras feature in distribute / setuptools).

This entry format is used in the The distlib.scripts package for installing scripts based on Python callables.

Distribution dependencies

You can use the distlib.locators package to locate the dependencies that a distribution has. The distlib.database package has code which allow you to analyse the relationships between a set of distributions:

  • make_graph(), which generates a dependency graph from a list of distributions.

  • get_dependent_dists(), which takes a list of distributions and a specific distribution in that list, and returns the distributions that are dependent on that specific distribution.

  • get_required_dists(), which takes a list of distributions and a specific distribution in that list, and returns the distributions that are required by that specific distribution.

The graph returned by make_graph() is an instance of DependencyGraph.

Using the locators API


To locate a distribution in an index, we can use the locate() function. This returns a potentially downloadable distribution (in the sense that it has a download URL – of course, there are no guarantees that there will actually be a downloadable resource at that URL). The return value is an instance of distlib.database.Distribution which can be queried for any distributions it requires, so that they can also be located if desired. Here is a basic example:

>>> from distlib.locators import locate
>>> flask = locate('flask')
>>> flask
<Distribution Flask (0.10.1) [https://pypi.org/packages/source/F/Flask/Flask-0.10.1.tar.gz]>
>>> dependencies = [locate(r) for r in flask.run_requires]
>>> from pprint import pprint
>>> pprint(dependencies)
[<Distribution Werkzeug (0.9.1) [https://pypi.org/packages/source/W/Werkzeug/Werkzeug-0.9.1.tar.gz]>,
 <Distribution Jinja2 (2.7) [https://pypi.org/packages/source/J/Jinja2/Jinja2-2.7.tar.gz]>,
 <Distribution itsdangerous (0.21) [https://pypi.org/packages/source/i/itsdangerous/itsdangerous-0.21.tar.gz]>]

The values in the run_requires property are just strings. Here’s another example, showing a little more detail:

>>> authy = locate('authy')
>>> authy.run_requires
set(['httplib2 (>= 0.7, < 0.8)', 'simplejson'])
>>> authy
<Distribution authy (1.0.0) [http://pypi.org/packages/source/a/authy/authy-1.0.0.tar.gz]>
>>> deps = [locate(r) for r in authy.run_requires]
>>> pprint(deps)
[<Distribution httplib2 (0.7.7) [http://pypi.org/packages/source/h/httplib2/httplib2-0.7.7.zip]>,
 <Distribution simplejson (3.3.0) [http://pypi.org/packages/source/s/simplejson/simplejson-3.3.0.tar.gz]>]

Note that the constraints on the dependencies were honoured by locate().

Under the hood

Under the hood, locate() uses locators. Locators are a mechanism for finding distributions from a range of sources. Although the pypi subpackage has been copied from distutils2 to distlib, there may be benefits in a higher-level API, and so the distlib.locators package has been created as an experiment. Locators are objects which locate distributions. A locator instance’s get_project() method is called, passing in a project name: The method returns a dictionary containing information about distribution releases found for that project. The keys of the returned dictionary are versions, and the values are instances of distlib.database.Distribution.

The following locators are provided:

  • DirectoryLocator – this is instantiated with a base directory and will look for archives in the file system tree under that directory. Name and version information is inferred from the filenames of archives, and the amount of information returned about the download is minimal. The locator searches all subdirectories by default, but can be set to only look in the specified directory by setting the recursive keyword argument to False.

  • PyPIRPCLocator. – This takes a base URL for the RPC service and will locate packages using PyPI’s XML-RPC API. This locator is a little slow (the scraping interface seems to work faster) and case-sensitive. For example, searching for 'flask' will throw up no results, but you get the expected results when searching from 'Flask'. This appears to be a limitation of the underlying XML-RPC API. Note that 20 versions of a project necessitate 41 network calls (one to get the versions, and two more for each version – one to get the metadata, and another to get the downloads information).

  • PyPIJSONLocator. – This takes a base URL for the JSON service and will locate packages using PyPI’s JSON API. This locator is case-sensitive. For example, searching for 'flask' will throw up no results, but you get the expected results when searching from 'Flask'. This appears to be a limitation of the underlying JSON API. Note that unlike the XML-RPC service, only non-hidden releases will be returned.

  • SimpleScrapingLocator – this takes a base URL for the site to scrape, and locates packages using a similar approach to the PackageFinder class in pip, or as documented in the setuptools documentation as the approach used by easy_install.

  • DistPathLocator – this takes a DistributionPath instance and locates installed distributions. This can be used with AggregatingLocator to satisfy requirements from installed distributions before looking elsewhere for them.

  • AggregatingLocator – this takes a list of other aggregators and delegates finding projects to them. It can either return the first result found (i.e. from the first aggregator in the list provided which returns a non-empty result), or a merged result from all the aggregators in the list.

There is a default locator, available at distlib.locators.default_locator.

The locators package also contains a function, get_all_distribution_names(), which retrieves the names of all distributions registered on PyPI:

>>> from distlib.locators import get_all_distribution_names
>>> names = get_all_distribution_names()
>>> len(names)

This is implemented using the XML-RPC API.

Apart from JSONLocator,

None of the locators currently returns enough metadata to allow dependency resolution to be carried out, but that is a result of the fact that metadata relating to dependencies are not indexed, and would require not just downloading the distribution archives and inspection of contained metadata files, but potentially also introspecting setup.py! This is the downside of having vital information only available via keyword arguments to the setup() call: hopefully, a move to fully declarative metadata will facilitate indexing it and allowing the provision of improved features.

The locators will skip binary distributions other than wheels. (.egg files are currently treated as binary distributions).

The PyPI locator classes don’t yet support the use of mirrors, but that can be added in due course – once the basic functionality is working satisfactorily.

Using the index API

You can use the distlib.index package to perform operations relating to a package index compatible with PyPI. This includes things like registering a project, uploading a distribution or uploading documentation.


You access index functionality through an instance of the PackageIndex class. This is instantiated with the URL of the repository (which can be omitted if you want to use PyPI itself):

>>> from distlib.index import PackageIndex
>>> index = PackageIndex()
>>> index.url

To use a local test server, you might do this:

>>> index = PackageIndex('http://localhost:8080/')

Registering a project

Registering a project can be done using a Metadata instance which holds the index metadata used for registering. A simple example:

>>> from distlib.metadata import Metadata
>>> metadata = Metadata()
>>> metadata.name = 'tatterdemalion'
>>> metadata.version = '0.1'
>>> # other fields omitted
>>> response = index.register(metadata)

The register() method returns an HTTP response, such as might be returned by a call to urlopen. If an error occurs, a HTTPError will be raised. Otherwise, the response.code should be 200.

Uploading a source distribution

To upload a source distribution, you need to do the following as a minimum:

>>> metadata = ... # get a populated Metadata instance
>>> response = index.upload_file(metadata, archive_name)

The upload_file() method returns an HTTP response or, in case of error, raises an HTTPError.

Uploading binary distributions

When uploading binary distributions, you need to specify the file type and Python version, as in the following example:

>>> response = index.upload_file(metadata, archive_name,
...                              filetype='bdist_dumb',
...                              pyversion='2.6')

Signing a distribution

To sign a distribution, you will typically need GnuPG. The default implementation looks for gpg or gpg2 on the path, but if not available there, you can can explicitly specify an absolute path indicating where the signing program is to be found:

>>> index.gpg = '/path/to/gpg'

Once this is set, you can sign the archive before uploading, as follows:

>>> response = index.upload_file(metadata, archive_name,
...                              signer='Test User',
...                              sign_password='secret',

As an alternative to passing the keystore with each call, you can specify that in an instance attribute:

>>> index.gpg_home = '/path/to/keys'

The keystore is a directory which contains the GnuPG key database (files like pubring.gpg, secring.gpg, and trustdb.gpg).

When you sign a distribution, both the distribution and the signature are uploaded to the index.

Downloading files

The PackageIndex class contains a utility method which allows you to download distributions (and other files, such as signatures):

>>> index.download_file(url, destfile, digest=None, reporthook=None)

This is similar in function to urlretrieve() in the standard library. Provide a digest if you want the call to check that the has digest of the downloaded file matches a specific value: if not provided, no matching is done. The value passed can just be a plain string in the case of an MD5 digest or, if you want to specify the hashing algorithm to use, specify a tuple such as ('sha1', '0123456789abcdef...'). The hashing algorithm must be one that’s supported by the hashlib module.

Benefits to using this method over plain urlretrieve() are:

  • It will use the ssl_verifier, if set, to ensure that the download is coming from where you think it is (see Verifying HTTPS connections).

  • It will compute the digest as it downloads, saving you from having to read the whole of the downloaded file just to compute its digest.

Note that the url you download from doesn’t actually need to be on the index – in theory, it could be from some other site. Note that if you have an ssl_verifier set on the index, it will perform its checks according to whichever url you supply – whether it’s a resource on the index or not.

Verifying signatures

For any archive downloaded from an index, you can retrieve any signature by just appending .asc to the path portion of the download URL for the archive, and downloading that. The index class offers a verify_signature() method for validating a signature. If you have files ‘good.bin’, ‘bad.bin’ which are different from each other, and ‘good.bin.asc’ has the signature for ‘good.bin’, then you can verify signatures like this:

>>> index.verify_signature('good.bin.asc', 'good.bin', '/path/to/keys')
>>> index.verify_signature('good.bin.asc', 'bad.bin', '/path/to/keys')

The last argument, which is optional, specifies a directory which holds the GnuPG keys used for verification – the keystore. Instead of specifying the keystore location in each call, you can specify the location in an instance attribute:

>>> index.gpg_home = '/path/to/keys'

If you do this, you don’t need to pass the keystore location.

Note that if you don’t have the gpg or gpg2 programs on the path, you may need to specify the location of the verifier program explicitly:

>>> index.gpg = '/path/to/gpg'
Some caveats about verified signatures

In order to be able to perform signature verification, you’ll have to ensure that the public keys of whoever signed those distributions are in your key store. However, having these keys shouldn’t give you a false sense of security; unless you can be sure that those keys actually belong to the people or organisations they purport to represent, the signature has no real value, even if it is verified without error. For you to be able to trust a key, it would need to be signed by someone you trust, who vouches for it – and this requires there to be either a signature from a valid certifying authority (e.g. Verisign, Thawte etc.) or a Web of Trust around the keys that you want to rely on.

An index may itself countersign distributions (so it deals with the keys of the distribution publishers, but you need only deal with the public signing key belonging to the index). If you trust the index, you can trust the verified signature if it’s signed by the index.

Uploading documentation

To upload documentation, you need to specify the metadata and the directory which is the root of the documentation (typically, if you use Sphinx to build your documentation, this will be something like <project>/docs/_build/html):

>>> response = index.upload_documentation(metadata, doc_dir)

The upload_documentation() method returns an HTTP response or, in case of error, raises an HTTPError. The call will zip up the entire contents of the passed directory doc_dir and upload the zip file to the index.


Operations which update the index (all of the above) will require authenticated requests. You can specify a username and password to use for requests sent to the index:

>>> index.username = 'test'
>>> index.password = 'secret'

For your convenience, these will be automatically read from any .pypirc file which you have; if it contains entries for multiple indexes, a repository key in .pypirc must match index.url to identify which username and password are to be read from .pypirc. Note that to ensure compatibility, distlib uses distutils code to read the .pypirc configuration. Thus, given the .pypirc file:

index-servers =

username: me
password: my_strong_password

repository: http://localhost:8080/
username: test
password: secret

you would see the following:

>>> index = PackageIndex()
>>> index.username
>>> index.password
>>> index = PackageIndex('http://localhost:8080/')
>>> index.username
>>> index.password

Verifying HTTPS connections

Although Python has full support for SSL, it does not, by default, verify SSL connections to servers. That’s because in order to do so, a set of certificates which certify the identity of the server needs to be provided (see the relevant Python documentation for details).

Support for verifying SSL connections is provided in distlib through a handler, distlib.util.HTTPSHandler. To use it, set the ssl_verifier attribute of the index to a suitably configured instance. For example:

>>> from distlib.util import HTTPSHandler
>>> verifier = HTTPSHandler('/path/to/root/certs.pem')
>>> index.ssl_verifier = verifier

By default, the handler will attempt to match domains, including wildcard matching. This means that (for example) you access foo.org or www.foo.org which have a certificate for *.foo.org, the domains will match. If the domains don’t match, the handler raises a CertificateError (a subclass of ValueError).

Domain mismatches can, however, happen for valid reasons. Say a hosting server bar.com hosts www.foo.org, which we are trying to access using SSL. If the server holds a certificate for www.foo.org, it will present it to the client, as long as both support Server Name Indication (SNI). While distlib supports SNI where Python supports it, Python 2.x does not include SNI support. For this or some other reason , you may wish to turn domain matching off. To do so, instantiate the verifier like this:

>>> verifier = HTTPSHandler('/path/to/root/certs.pem', False)
Ensuring that only HTTPS connections are made

You may want to ensure that traffic is only HTTPS for a particular interaction with a server – for example:

  • Deal with a Man-In-The-Middle proxy server which listens on port 443 but talks HTTP rather than HTTPS

  • Deal with situations where an index page obtained via HTTPS contains links with a scheme of http rather than https.

To do this, instead of using HTTPSHandler as shown above, use the HTTPSOnlyHandler class instead, which disallows any HTTP traffic. It’s used in the same way as HTTPSHandler:

>>> from distlib.util import HTTPSOnlyHandler
>>> verifier = HTTPSOnlyHandler('/path/to/root/certs.pem')
>>> index.ssl_verifier = verifier

Note that with this handler, you can’t make any HTTP connections at all - it will raise URLError if you try.

Getting hold of root certificates

At the time of writing, you can find a file in the appropriate format on the cURL website. Just download the cacert.pem file and pass the path to it when instantiating your verifier.

Saving a default configuration

If you don’t have a .pypirc file but want to save one, you can do this by setting the username and password and calling the save_configuration() method:

>>> index = PackageIndex()
>>> index.username = 'fred'
>>> index.password = 'flintstone'
>>> index.save_configuration()

This will use distutils code to save a default .pypirc file which specifies a single index – PyPI – with the specified username and password.

Searching PyPI

You can use the search() method of PackageIndex to search for distributions on PyPI:

>>> index = PackageIndex()
>>> from pprint import pprint
>>> pprint(index.search('tatterdema'))
[{'_pypi_ordering': 0,
  'name': 'tatterdemalion',
  'summary': 'A dummy distribution',
  'version': '0.1.0'}]

If a string is specified, just the name is searched for. Alternatively, you can specify a dictionary of attributes to search for, along with values to match. For example:

>>> pprint(index.search({'summary': 'dummy'}))
[{'_pypi_ordering': 5,
  'name': 'collective.lorem',
  'summary': 'A package that provides dummy content generation.',
  'version': '0.2.3'},
 {'_pypi_ordering': 7,
  'name': 'collective.loremipsum',
  'summary': 'Creates dummy content with populated Lorem Ipsum.',
  'version': '0.8'},
 {'_pypi_ordering': 1,
  'name': 'cosent.dummypackage',
  'summary': 'A dummy package for buildtools testing',
  'version': '0.4'},
 {'_pypi_ordering': 0,
  'name': 'django-dummyimage',
  'summary': 'Dynamic Dummy Image Generator For Django!',
  'version': '0.1.1'},
 {'_pypi_ordering': 1,
  'name': 'django-plainpasswordhasher',
  'summary': 'Dummy (plain text) password hashing for Django.',
  'version': '0.2'},
 {'_pypi_ordering': 2,
  'name': 'django-plainpasswordhasher',
  'summary': 'Dummy (plain text) password hashing for Django.',
  'version': '0.3'},
 {'_pypi_ordering': 1,
  'name': 'dummycache',
  'summary': 'A dummy in-memory cache for development and testing. (Not recommended for production use.)',
  'version': '0.0.2'},
 {'_pypi_ordering': 0,
  'name': 'dummy-txredis',
  'summary': 'Dummy txRedis client and factory.',
  'version': '0.5'},
 {'_pypi_ordering': 7,
  'name': 'eea.eggmonkeytesttarget',
  'summary': 'A dummy package to test eea.eggmonkey',
  'version': '5.7'},
 {'_pypi_ordering': 8,
  'name': 'invewrapper',
  'summary': 'dummy/transitional package that depends on "pew"',
  'version': '0.1.8'},
 {'_pypi_ordering': 0,
  'name': 'monoprocessing',
  'summary': 'A dummy implementation of multiprocessing.Pool',
  'version': '0.1'},
 {'_pypi_ordering': 0,
  'name': 'myFun',
  'summary': 'This is a dummy function which prints given list data.',
  'version': '1.0.0'},
 {'_pypi_ordering': 0,
  'name': 'ReadableDict-a-dict-without-brackets',
  'summary': 'provides a dummy implementation of a dict without brackets',
  'version': '0.0'},
 {'_pypi_ordering': 4,
  'name': 'setuptools_dummy',
  'summary': 'Setuptools Dummy Filefinder',
  'version': ''},
 {'_pypi_ordering': 0,
  'name': 'tatterdemalion',
  'summary': 'A dummy distribution',
  'version': '0.1.0'}]

If you specify multiple attributes, then the search returns the intersection of matches – an and operation:

>>> pprint(index.search({'summary': 'dummy', 'name': 'ta'}))
[{'_pypi_ordering': 7,
  'name': 'eea.eggmonkeytesttarget',
  'summary': 'A dummy package to test eea.eggmonkey',
  'version': '5.7'},
 {'_pypi_ordering': 0,
  'name': 'tatterdemalion',
  'summary': 'A dummy distribution',
  'version': '0.1.0'}]

If you want a union of matches – an or operation – specify a second argument to the PackageIndex.search() method with the value 'or':

>>> pprint(index.search({'version': '2013.9', 'name': 'pytzp'}, 'or'))
[{'_pypi_ordering': 65,
  'name': 'pytz',
  'summary': 'World timezone definitions, modern and historical',
  'version': '2013.9'},
 {'_pypi_ordering': 2,
  'name': 'pytzpure',
  'summary': 'A pure-Python version of PYTZ (timezones).',
  'version': '0.2.4'}]

The search functionality makes use of PyPI’s XML-RPC interface, so it will only work for indexes which supply a compatible implementation. The following search attributes are currently supported:

  • name

  • version

  • stable_version

  • author

  • author_email

  • maintainer

  • maintainer_email

  • home_page

  • license

  • summary

  • description

  • keywords

  • platform

  • download_url

  • classifiers (list of classifier strings)

  • project_url

  • docs_url (URL of the pythonhosted.org docs if they’ve been supplied)

Using the metadata and markers APIs

The metadata API is exposed through a Metadata class. This class can read and write metadata files complying with any of the defined versions: 1.0 (PEP 241), 1.1 (PEP 314), 1.2 (PEP 345), 2.1 (PEP 566) and 2.2 (PEP 643). It implements methods to parse and write metadata files.

Instantiating metadata

You can simply instantiate a Metadata instance and start populating it:

>>> from distlib.metadata import Metadata
>>> md = Metadata()
>>> md.name = 'foo'
>>> md.version = '1.0'

An instance so created may not be valid unless it has some minimal properties which meet certain constraints, as specified in the Core metadata specifications.

These constraints aren’t applicable to legacy metadata. Therefore, when creating Metadata instances to deal with such metadata, you can specify the scheme keyword when creating the instance:

>>> legacy_metadata = Metadata(scheme='legacy')

The term ‘legacy’ refers to the version scheme.

Whether dealing with current or legacy metadata. an instance’s validate() method can be called to ensure that the metadata has no missing or invalid data. This raises a DistlibException (either MetadataMissingError or MetadataInvalidError) if the metadata isn’t valid.

You can initialise an instance with a dictionary using the following form:

>>> metadata = Metadata(mapping=a_dictionary)

Reading metadata from files and streams

The Metadata class can be instantiated with the path of the metadata file. Here’s an example with legacy metadata:

>>> from distlib.metadata import Metadata
>>> metadata = Metadata(path='PKG-INFO')
>>> metadata.name
>>> metadata.version
>>> metadata.run_requires

Instead of using the path keyword argument to specify a file location, you can also specify a fileobj keyword argument to specify a file-like object which contains the data.

Writing metadata to paths and streams

Writing metadata can be done using the write method:

>>> metadata.write(path='/to/my/pydist.json')

You can also specify a file-like object to write to, using the fileobj keyword argument.

Using markers

Environment markers are implemented in the distlib.markers package and accessed via a single function, interpret().

See PEP 508 for more information about environment markers. The interpret() function takes a string argument which represents a Boolean expression, and returns either True or False:

>>> from distlib.markers import interpret
>>> interpret('python_version >= "1.0"')

You can pass in a context dictionary which is checked for values before the environment:

>>> interpret('python_version >= "1.0"', {'python_version': '0.5'})

You won’t normally need to work with markers in this way – they are dealt with by the Metadata and Distribution logic when needed.

Using the resource API

You can use the distlib.resources package to access data stored in Python packages, whether in the file system or .zip files. Consider a package which contains data alongside Python code:

├── bar
│   ├── bar_resource.bin
│   ├── baz.py
│   └── __init__.py
├── foo_resource.bin
├── __init__.py
└── nested
    └── nested_resource.bin

Access to resources in the file system

You can access these resources like so:

>>> from distlib.resources import finder
>>> f = finder('foofoo')
>>> r = f.find('foo_resource.bin')
>>> r.is_container
>>> r.size
>>> r.bytes
>>> s = r.as_stream()
>>> s.read()
>>> s.close()
>>> r = f.find('nested')
>>> r.is_container
>>> r.resources
>>> r = f.find('nested/nested_resource.bin')
>>> r.size
>>> r.bytes
b'nested data\n'
>>> f = finder('foofoo.bar')
>>> r = f.find('bar_resource.bin')
>>> r.is_container
>>> r.bytes

Access to resources in the .zip files

It works the same way if the package is in a .zip file. Given the zip file foo.zip:

$ unzip -l foo.zip
Archive:  foo.zip
  Length      Date    Time    Name
---------  ---------- -----   ----
       10  2012-09-20 21:34   foo/foo_resource.bin
        8  2012-09-20 21:42   foo/__init__.py
       14  2012-09-20 21:42   foo/bar/baz.py
        8  2012-09-20 21:42   foo/bar/__init__.py
        5  2012-09-20 21:33   foo/bar/bar_resource.bin
---------                     -------
       45                     5 files

You can access its resources as follows:

>>> import sys
>>> sys.path.append('foo.zip')
>>> from distlib.resources import finder
>>> f = finder('foo')
>>> r = f.find('foo_resource.bin')
>>> r.is_container
>>> r.size
>>> r.bytes

and so on.

Iterating over resources

You can iterate over resources as shown in the following example:

>>> from distlib.resources import finder
>>> f = finder('foofoo')
>>> iterator = f.iterator('')
>>> for r in iterator: print('%-20s %s' % (r.name, r.is_container))
foo_resource.bin     False
__init__.py          False
bar                  True
bar/bar_resource.bin False
bar/baz.py           False
bar/__init__.py      False
nested               True
nested/nested_resource.bin False

It works with zipped resources, too:

>>> import sys
>>> sys.path.append('foo.zip')
>>> from distlib.resources import finder
>>> f = finder('foo')
>>> iterator = f.iterator('')
>>> for r in iterator: print('%-20s %s' % (r.name, r.is_container))
foo_resource.bin     False
__init__.py          False
bar                  True
bar/bar_resource.bin False
bar/baz.py           False
bar/__init__.py      False

Using the scripts API

You can use the distlib.scripts API to install scripts. Installing scripts is slightly more involved than just copying files:

  • You may need to adjust shebang lines in scripts to point to the interpreter to be used to run scripts. This is important in virtual environments (venvs), and also in other situations where you may have multiple Python installations on a single computer.

  • On Windows, on systems where the PEP 397 launcher isn’t installed, it is not easy to ensure that the correct Python interpreter is used for a script. You may wish to install native Windows executable launchers which run the correct interpreter, based on a shebang line in the script.

Specifying scripts to install

To install scripts, create a ScriptMaker instance, giving it the source and target directories for scripts:

>>> from distlib.scripts import ScriptMaker
>>> maker = ScriptMaker(source_dir, target_dir)

You can then install a script foo.py like this:

>>> maker.make('foo.py')

The string passed to make can take one of the following forms:

  • A filename, relative to the source directory for scripts, such as foo.py or subdir/bar.py.

  • A reference to a callable, given in the form:

    name = some_package.some_module:some_callable [flags]

    where the flags part is optional.

    For more information about flags, see Flag formats.

    Note that this format is exactly the same as for export entries in a distribution (see Exporting things from Distributions).

    When this form is passed to the ScriptMaker.make() method, a Python stub script is created with the appropriate shebang line and with code to load and call the specified callable with no arguments, returning its value as the return code from the script.

    You can pass an optional options dictionary to the make() method. This is meant to contain options which control script generation. There are two options currently in use:

    gui: This Boolean value, if True, indicates on Windows that a Windows executable launcher (rather than a launcher which is a console application) should be used. (This only applies if add_launchers is true.)

    interpreter_args: If provided, this should be a list of strings which are added to the shebang line following the interpreter. If there are values with spaces, you will need to surround them with double quotes.


    Use of this feature may affect portability, since POSIX does not standardise how these arguments are passed to the interpreter (see https://en.wikipedia.org/wiki/Shebang_line#Portability for more information).

    For example, you can pass {'gui': True} to generate a windowed script.

Wrapping callables with scripts

Let’s see how wrapping a callable works. Consider the following file:

 $ cat scripts/foo.py
   def main():
     print('Hello from foo')

 def other_main():
     print('Hello again from foo')

we can try wrapping ``main`` and ``other_main`` as callables::

   >>> from distlib.scripts import ScriptMaker
   >>> maker = ScriptMaker('scripts', '/tmp/scratch')
   >>> maker.make_multiple(('foo = foo:main', 'bar = foo:other_main'))
   ['/tmp/scratch/foo', '/tmp/scratch/bar']

we can inspect the resulting scripts. First, ``foo``::

 $ ls /tmp/scratch/
 bar  foo
 $ cat /tmp/scratch/foo

 if __name__ == '__main__':
     import sys, re

     def _resolve(module, func):
         mod = sys.modules[module]
         parts = func.split('.')
         result = getattr(mod, parts.pop(0))
         for p in parts:
             result = getattr(result, p)
         return result

         sys.argv[0] = re.sub('-script.pyw?$', '', sys.argv[0])

         func = _resolve('foo', 'main')
         rc = func() # None interpreted as 0
     except Exception as e:  # only supporting Python >= 2.6
         sys.stderr.write('%s\n' % e)
         rc = 1

The other script, bar, is different only in the essentials:

$ diff /tmp/scratch/foo /tmp/scratch/bar
<         func = _resolve('foo', 'main')
>         func = _resolve('foo', 'other_main')

Specifying a custom executable for shebangs

You may need to specify a custom executable for shebang lines. To do this, set the executable attribute of a ScriptMaker instance to the absolute Unicode path of the executable which you want to be written to the shebang lines of scripts. If not specified, the executable running the ScriptMaker code is used. If the value has spaces, you should surround it with double quotes. You can use the enquote_executable() function for this.

Changed in version 0.3.1: The enquote_executable() function was an internal function _enquote_executable in earlier versions.

For relocatable .exe files under Windows, you can specify the location of the python executable relative to the script by putting <launcher_dir> as the beginning of the executable path. Since windows places python.exe in the root install directory and the application scripts in the Scripts subdirectory, setting maker.executable = r"<launcher_dir>\..\python.exe" will allow you to move a python installation which is installed together with an application to a different path or a different machine and the .exe files will still run.

Generating variants of a script

When installing a script foo, it is not uncommon to want to install version-specific variants such as foo3 or foo-3.2. You can control exactly which variants of the script get written through the ScriptMaker instance’s variants attribute. This defaults to set(('', 'X.Y')), which means that by default a script foo would be installed as foo and foo-3.2 under Python 3.2. If the value of the variants attribute were set(('', 'X', 'X.Y')) then the foo script would be installed as foo, foo3 and foo-3.2 when run under Python 3.2.


If you need to generate variants for a different version of Python than the one running the ScriptMaker code, set the version_info attribute of the ScriptMaker instance to a 2-tuple holding the major and minor version numbers of the target Python version.

New in version 0.3.1.

Avoiding overwriting existing scripts

In some scenarios, you might overwrite existing scripts when you shouldn’t. For example, if you use Python 2.7 to install a distribution with script foo in the user site (see PEP 370), you will write (on POSIX) scripts ~/.local/bin/foo and ~/.local/bin/foo-2.7. If you then install the same distribution with Python 3.2, you would write (on POSIX) scripts ~/.local/bin/foo and ~/.local/bin/foo-3.2. However, by overwriting the ~/.local/bin/foo script, you may prevent verification or removal of the 2.7 installation to fail, because the overwritten file may be different (and so have a different hash from what was computed during the 2.7 installation).

To control overwriting of generated scripts this way, you can use the clobber attribute of a ScriptMaker instance. This is set to False by default, which prevents overwriting; to force overwriting, set it to True.

Generating windowed scripts on Windows

The make() and make_multiple() methods take an optional second options argument, which can be used to control script generation. If specified, this should be a dictionary of options. Currently, only the value for the gui key in the dictionary is inspected: if True, it generates scripts with .pyw extensions (rather than .py) and, if add_launchers is specified as True in the ScriptMaker instance, then (on Windows) a windowed native executable launcher is created (otherwise, the native executable launcher will be a console application).

Using the version API


The NormalizedVersion class implements a PEP 440 compatible version:

>>> from distlib.version import NormalizedVersion
>>> v1 = NormalizedVersion('1.0')
>>> v2 = NormalizedVersion('1.0a1')
>>> v3 = NormalizedVersion('1.0b1')
>>> v4 = NormalizedVersion('1.0c1')
>>> v5 = NormalizedVersion('1.0.post1')

These sort in the expected order:

>>> v2 < v3 < v4 < v1 < v5

You can’t pass any old thing as a version number:

>>> NormalizedVersion('foo')
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "distlib/version.py", line 49, in __init__
 self._parts = parts = self.parse(s)
File "distlib/version.py", line 254, in parse
 def parse(self, s): return normalized_key(s)
File "distlib/version.py", line 199, in normalized_key
 raise UnsupportedVersionError(s)
distlib.version.UnsupportedVersionError: foo

Matching versions against constraints

The NormalizedMatcher is used to match version constraints against versions:

>>> from distlib.version import NormalizedMatcher
>>> m = NormalizedMatcher('foo (1.0b1)')
>>> m
NormalizedMatcher('foo (1.0b1)')
>>> [m.match(v) for v in v1, v2, v3, v4, v5]
[False, False, True, False, False]

Specifying 'foo (1.0b1)' is equivalent to specifying 'foo (==1.0b1)', i.e. only the exact version is matched. You can also specify inequality constraints:

>>> m = NormalizedMatcher('foo (<1.0c1)')
>>> [m.match(v) for v in v1, v2, v3, v4, v5]
[False, True, True, False, False]

and multiple constraints:

>>> m = NormalizedMatcher('foo (>= 1.0b1, <1.0.post1)')
>>> [m.match(v) for v in v1, v2, v3, v4, v5]
[True, False, True, True, False]

You can do exactly the same thing as above with setuptools/ distribute version numbering (use LegacyVersion and LegacyMatcher) or with semantic versioning (use SemanticVersion and SemanticMatcher). However, you can’t mix and match versions of different types:

>>> from distlib.version import SemanticVersion, LegacyVersion
>>> nv = NormalizedVersion('1.0.0')
>>> lv = LegacyVersion('1.0.0')
>>> sv = SemanticVersion('1.0.0')
>>> lv == sv
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "distlib/version.py", line 61, in __eq__
File "distlib/version.py", line 58, in _check_compatible
raise TypeError('cannot compare %r and %r' % (self, other))
TypeError: cannot compare LegacyVersion('1.0.0') and SemanticVersion('1.0.0')
>>> nv == sv
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "distlib/version.py", line 61, in __eq__
File "distlib/version.py", line 58, in _check_compatible
raise TypeError('cannot compare %r and %r' % (self, other))
TypeError: cannot compare NormalizedVersion('1.0.0') and SemanticVersion('1.0.0')

Using the wheel API

You can use the distlib.wheel package to build and install from files in the Wheel format, defined in PEP 427.

Building wheels

Building wheels is straightforward:

from distlib.wheel import Wheel

wheel = Wheel()

# Set the distribution's identity
wheel.name = 'name_of_distribution'
wheel.version = '0.1'

# Indicate where the files to go in the wheel are to be found
paths = {
    'prefix': '/path/to/installation/prefix',
    'purelib': '/path/to/purelib',  # only one of purelib
    'platlib': '/path/to/platlib',  # or platlib should be set
    'scripts': '/path/to/scripts',
    'headers': '/path/to/headers',
    'data': '/path/to/data',

wheel.dirname = '/where/you/want/the/wheel/to/go'
# Now build

If the 'data', 'headers' and 'scripts' keys are absent, or point to paths which don’t exist, nothing will be added to the wheel for these categories. The 'prefix' key and one of 'purelib' or 'platlib' must be provided, and the paths referenced should exist.

Customising tags during build

By default, the build() method will use default tags depending on whether or not the build is a pure-Python build:

  • For a pure-Python build, the pyver will be set to pyXY where XY is the version of the building Python. The abi tag will be none and the arch tag will be any.

  • For a build which is not pure-Python (i.e. contains C code), the pyver will be set to e.g. cpXY, and the abi and arch tags will be set according to the building Python.

If you want to override these default tags, you can pass a tags parameter to the build() method which has the tags you want to declare. For example, for a pure build where we know that the code in the wheel will be compatible with the major version of the building Python:

from wheel import PYVER
tags = {
    'pyver': [PYVER[:-1], PYVER],
wheel.build(paths, tags)

This would set the pyver tags to be pyX.pyXY where X and Y relate to the building Python. You can similarly pass values using the abi and arch keys in the tags dictionary.

Specifying a wheel’s version

You can also specify a particular “Wheel-Version” to be written to the wheel metadata of a wheel you’re building. Simply pass a (major, minor) tuple in the wheel_version keyword argument to build(). If not specified, the most recent version supported is written.

Installing from wheels

Installing from wheels is similarly straightforward. You just need to indicate where you want the files in the wheel to be installed:

from distlib.wheel import Wheel
from distlib.scripts import ScriptMaker

wheel = Wheel('/path/to/my_dist-0.1-py32-none-any.whl')

# Indicate where the files in the wheel are to be installed to.
# All the keys should point to writable paths.
paths = {
    'prefix': '/path/to/installation/prefix',
    'purelib': '/path/to/purelib',
    'platlib': '/path/to/platlib',
    'scripts': '/path/to/scripts',
    'headers': '/path/to/headers',
    'data': '/path/to/data',

maker = ScriptMaker(None, None)
# You can specify a custom executable in script shebang lines, whether
# or not to install native executable launchers, whether to do a dry run
# etc. by setting attributes on the maker, either when creating it or
# subsequently.

# Now install. The method accepts optional keyword arguments:
# - A ``warner`` argument which, if specified, should be a callable that
#   will be called with (software_wheel_version, file_wheel_version) if
#   they differ. They will both be in the form (major_ver, minor_ver).
# - A ``lib_only`` argument which indicates that only the library portion
#   of the wheel should be installed - no scripts, header files or
#   non-package data.

wheel.install(paths, maker)

Only one of the purelib or platlib paths will actually be written to (assuming that they are different, which isn’t often the case). Which one it is depends on whether the wheel metadata declares that the wheel contains pure Python code.

Verifying wheels

You can verify that a wheel’s contents match the declared contents in the wheel’s RECORD entry, by calling the verify() method. This will raise a DistlibException if a size or digest mismatch is found.

Modifying wheels


In an ideal world one would not need to modify wheels, but in the short term there might be a need to do so (for example, to add dependency information which is missing). If you are working with wheels on your own projects, you shouldn’t use the method described here, as you will have full control of the wheels you build yourself. However, if you are working with third party wheels which you don’t build yourself but you need to modify in some way, then the approach described below might be useful.

You can update existing wheels with distlib by calling the update() method of a wheel. This is called as follows:

modified = wheel.update(modifier, dest_dir, **kwargs)

where the modifier is a callable which you specify, and kwargs are options you want to pass to it (currently, the update() method passes kwargs unchanged to the modifier). The dest_dir argument indicates where you want any new wheel to be written - it is optional and if not specified, the existing wheel will be overwritten.

The update() method extracts the entire contents of the wheel to a temporary location, and then calls modifier as follows:

modified = modifier(path_map, **kwargs)

where path_map is a dictionary mapping archive paths to the location of the corresponding extracted archive entry, and kwargs is whatever was passed to the update method. If the modifier returns True, a new wheel is built from the (possibly updated) contents of path_map and its path name. The passed path_map will contain all of the wheel’s entries other than the RECORD entry (which will be recreated if a new wheel is built).

For example, if you wanted to add numpy as a dependency in a scipy wheel, you might do something like this:

def add_numpy_dependency(path_map, **kwargs):
    mdpath = path_map['scipy-0.11.dist-info/pydist.json']
    md = Metadata(path=mdpath)
    return True

wheel = Wheel('scipy-0.11-py27-abi3-linux_x86_64.whl')

In the above example, the modifier doesn’t actually use kwargs, but you could pass useful information which can be used to control the modifier’s operation. For example, you might make the function work with other distributions than scipy, or other versions of scipy:

def add_numpy_dependency(path_map, **kwargs):
    name = kwargs.get('name', 'scipy')
    version = kwargs.get('version', '0.11')
    key = '%s-%s.dist-info/pydist.json' % (name, version)
    mdpath = path_map[key]
    md = Metadata(path=mdpath)
    return True

Mounting wheels

One of Python’s perhaps under-used features is zipimport, which gives the ability to import Python source from .zip files. Since wheels are .zip files, they can sometimes be used to provide functionality without needing to be installed. Whereas .zip files contain no convention for indicating compatibility with a particular Python, wheels do contain this compatibility information. Thus, it is possible to check if a wheel can be directly imported from, and the wheel support in distlib allows you to take advantage of this using the mount() and unmount() methods. When you mount a wheel, its absolute path name is added to sys.path, allowing the Python code in it to be imported. (A DistlibException is raised if the wheel isn’t compatible with the Python which calls the mount() method.)

The mount() method takes an optional keyword parameter append which defaults to False, meaning the a mounted wheel’s pathname is added to the beginning of sys.path. If you pass True, the pathname is appended to sys.path.

The mount() method goes further than just enabling Python imports – any C extensions in the wheel are also made available for import. For this to be possible, the wheel has to be built with additional metadata about extensions – a JSON file called EXTENSIONS which serialises an extension mapping dictionary. This maps extension module names to the names in the wheel of the shared libraries which implement those modules.

Running unmount() on the wheel removes its absolute pathname from sys.path and makes its C extensions, if any, also unavailable for import.


The C extension mounting functionality may not work in all cases, though it should work in a useful subset of cases. Use with care. Note that extension information is currently only available in wheels built using distil – for wheels built using e.g. pip, this note will not apply, because C extensions will never be available for import.

  • There might be subtle differences in binary compatibility between the extension and the running Python, because the compatibility tag framework currently does not capture all the relevant ABI information. This is a situation which can be expected to improve over time.

  • If the extension uses custom dynamically linked libraries which are bundled with the extension, it may not be found by the dynamic loading machinery, for reasons that are platform-dependent. In such cases, you should have a good understanding of how dynamic loading works on your platforms, before taking advantage of this feature.

Using vanilla pip to build wheels for existing distributions on PyPI

Although work is afoot to add wheel support to pip, you don’t need this to build wheels for existing PyPI distributions if you use distlib. The following script shows how you can use an unpatched, vanilla pip to build wheels:

#!/usr/bin/env python
# -*- coding: utf-8 -*-
# Copyright (C) 2013 Vinay Sajip. License: MIT

import logging
import optparse     # for 2.6
import os
import re
import shutil
import subprocess
import sys
import tempfile

logger = logging.getLogger('wheeler')

from distlib.compat import configparser, filter
from distlib.database import DistributionPath, Distribution, make_graph
from distlib.locators import (JSONLocator, SimpleScrapingLocator,
                              AggregatingLocator, DependencyFinder)
from distlib.manifest import Manifest
from distlib.metadata import Metadata
from distlib.util import parse_requirement, get_package_data
from distlib.wheel import Wheel

EGG_INFO_RE = re.compile(r'(-py\d\.\d)?\.egg-info', re.I)

INSTALLED_DISTS = DistributionPath(include_egg=True)

def get_requirements(data):
    lines = []
    for line in data.splitlines():
        line = line.strip()
        if not line or line[0] == '#':
    reqts = []
    extras = {}
    result = {'install': reqts, 'extras': extras}
    for line in lines:
        if line[0] != '[':
            i = line.find(']', 1)
            if i < 0:
                raise ValueError('unrecognised line: %r' % line)
            extra = line[1:i]
            extras[extra] = reqts = []
    return result

def convert_egg_info(libdir, prefix, options):
    files = os.listdir(libdir)
    ei = list(filter(lambda d: d.endswith('.egg-info'), files))[0]
    olddn = os.path.join(libdir, ei)
    di = EGG_INFO_RE.sub('.dist-info', ei)
    newdn = os.path.join(libdir, di)
    os.rename(olddn, newdn)
    if options.compatible:
        renames = {}
        renames = {
            'entry_points.txt': 'EXPORTS',
    excludes = set([
        'SOURCES.txt',          # of no interest in/post WHEEL
        'installed-files.txt',  # replaced by RECORD, so not needed
        'requires.txt',         # added to METADATA, so not needed
        'PKG-INFO',             # replaced by METADATA
        'not-zip-safe',         # not applicable
    files = os.listdir(newdn)
    metadata = mdname = reqts = None
    for oldfn in files:
        pn = os.path.join(newdn, oldfn)
        if oldfn in renames:
            os.rename(pn, os.path.join(newdn, renames[oldfn]))
            if oldfn == 'requires.txt':
                with open(pn, 'r') as f:
                    reqts = get_requirements(f.read())
            elif oldfn == 'PKG-INFO':
                metadata = Metadata(path=pn)
                pd = get_package_data(metadata.name, metadata.version)
                metadata = Metadata(mapping=pd['index-metadata'])
                mdname = os.path.join(newdn, 'pydist.json')
            if oldfn in excludes or not options.compatible:
    if metadata:
        # Use Metadata 1.2 or later
        metadata.provides += ['%s (%s)' % (metadata.name,
        # Update if not set up by get_package_data
        if reqts and not metadata.run_requires:
            metadata.dependencies = reqts
    manifest = Manifest(os.path.dirname(libdir))
    paths = manifest.allfiles
    dp = DistributionPath([libdir])
    dist = next(dp.get_distributions())
    dist.write_installed_files(paths, prefix)

def install_dist(distname, workdir, options):
    pfx = '--install-option='
    purelib = pfx + '--install-purelib=%s/purelib' % workdir
    platlib = pfx + '--install-platlib=%s/platlib' % workdir
    headers = pfx + '--install-headers=%s/headers' % workdir
    scripts = pfx + '--install-scripts=%s/scripts' % workdir
    data = pfx + '--install-data=%s/data' % workdir
    # Use the pip adjacent to sys.executable, if any (for virtualenvs)
    d = os.path.dirname(sys.executable)
    files = filter(lambda o: o in ('pip', 'pip.exe'), os.listdir(d))
    if not files:
        prog = 'pip'
        prog = os.path.join(d, next(files))
    cmd = [prog, 'install',
           '--no-deps', '--quiet',
           '--index-url', 'http://pypi.org/simple/',
           '--timeout', '3', '--default-timeout', '3',
           purelib, platlib, headers, scripts, data, distname]
    result = {
        'scripts': os.path.join(workdir, 'scripts'),
        'headers': os.path.join(workdir, 'headers'),
        'data': os.path.join(workdir, 'data'),
    print('Pipping %s ...' % distname)
    p = subprocess.Popen(cmd, shell=False, stdout=sys.stdout,
    stdout, _ = p.communicate()
    if p.returncode:
        raise ValueError('pip failed to install %s:\n%s' % (distname, stdout))
    for dn in ('purelib', 'platlib'):
        libdir = os.path.join(workdir, dn)
        if os.path.isdir(libdir):
            result[dn] = libdir
    convert_egg_info(libdir, workdir, options)
    dp = DistributionPath([libdir])
    dist = next(dp.get_distributions())
    md = dist.metadata
    result['name'] = md.name
    result['version'] = md.version
    return result

def build_wheel(distname, options):
    result = None
    r = parse_requirement(distname)
    if not r:
        print('Invalid requirement: %r' % distname)
        dist = INSTALLED_DISTS.get_distribution(r.name)
        if dist:
            print('Can\'t build a wheel from already-installed '
                  'distribution %s' % dist.name_and_version)
            workdir = tempfile.mkdtemp()    # where the Wheel input files will live
                paths = install_dist(distname, workdir, options)
                paths['prefix'] = workdir
                wheel = Wheel()
                wheel.name = paths.pop('name')
                wheel.version = paths.pop('version')
                wheel.dirname = options.destdir
                result = wheel
    return result

def main(args=None):
    parser = optparse.OptionParser(usage='%prog [options] requirement [requirement ...]')
    parser.add_option('-d', '--dest', dest='destdir', metavar='DESTDIR',
                      default=os.getcwd(), help='Where you want the wheels '
                      'to be put.')
    parser.add_option('-n', '--no-deps', dest='deps', default=True,
                      help='Don\'t build dependent wheels.')
    options, args = parser.parse_args(args)
    options.compatible = True   # may add flag to turn off later
    if not args:
        # Check if pip is available; no point in continuing, otherwise
            with open(os.devnull, 'w') as f:
                p = subprocess.call(['pip', '--version'], stdout=f, stderr=subprocess.STDOUT)
        except Exception:
            p = 1
        if p:
            print('pip appears not to be available. Wheeler needs pip to '
                  'build  wheels.')
            return 1
        if options.deps:
            # collect all the requirements, including dependencies
            u = 'http://pypi.org/simple/'
            locator = AggregatingLocator(JSONLocator(),
                                         SimpleScrapingLocator(u, timeout=3.0),
            finder = DependencyFinder(locator)
            wanted = set()
            for arg in args:
                r = parse_requirement(arg)
                if not r.constraints:
                    dname = r.name
                    dname = '%s (%s)' % (r.name, ', '.join(r.constraints))
                print('Finding the dependencies of %s ...' % arg)
                dists, problems = finder.find(dname)
                if problems:
                    print('There were some problems resolving dependencies '
                          'for %r.' % arg)
                    for _, info in problems:
                        print('  Unsatisfied requirement %r' % info)
                wanted |= dists
            want_ordered = True     # set to False to skip ordering
            if not want_ordered:
                wanted = list(wanted)
                graph = make_graph(wanted, scheme=locator.scheme)
                slist, cycle = graph.topological_sort()
                if cycle:
                    # Now sort the remainder on dependency count.
                    cycle = sorted(cycle, reverse=True,
                                   key=lambda d: len(graph.reverse_list[d]))
                wanted = slist + cycle

                # get rid of any installed distributions from the list
                for w in list(wanted):
                    dist = INSTALLED_DISTS.get_distribution(w.name)
                    if dist or w.name in ('setuptools', 'distribute'):
                        s = w.name_and_version
                        print('Skipped already-installed distribution %s' % s)

            # converted wanted list to pip-style requirements
            args = ['%s==%s' % (dist.name, dist.version) for dist in wanted]

        # Now go build
        built = []
        for arg in args:
            wheel = build_wheel(arg, options)
            if wheel:
        if built:
            if options.destdir == os.getcwd():
                dest = ''
                dest = ' in %s' % options.destdir
            print('The following wheels were built%s:' % dest)
            for wheel in built:
                print('  %s' % wheel.filename)

if __name__ == '__main__':
    logging.basicConfig(format='%(levelname)-8s %(name)s %(message)s',
                        filename='wheeler.log', filemode='w')
        rc = main()
    except Exception as e:
        print('Failed - sorry! Reason: %s\nPlease check the log.' % e)
        rc = 1

This script, wheeler.py, is also available here. Note that by default, it downloads dependencies of any distribution you specify and builds separate wheels for each distribution. It’s smart about not repeating work if dependencies are common across multiple distributions you specify:

$ python wheeler.py sphinx flask
Finding the dependencies of sphinx ...
Finding the dependencies of flask ...
Pipping Jinja2==2.6 ...
Pipping docutils==0.10 ...
Pipping Pygments==1.6 ...
Pipping Werkzeug==0.8.3 ...
Pipping Sphinx==1.1.3 ...
Pipping Flask==0.9 ...
The following wheels were built:

Note that the common dependency – Jinja2 – was only built once.

You can opt to not build dependent wheels by specifying --no-deps on the command line.

Note that the script also currently uses an http: URL for PyPI – this may need to change to an https: URL in the future.


It can’t be used to build wheels from existing distributions, as pip will either refuse to install to custom locations (because it views a distribution as already installed), or will try to upgrade and thus uninstall the existing distribution, even though installation is requested to a custom location (and uninstallation is not desirable). For best results, run it in a fresh venv:

$ my_env/bin/python wheeler.py some_dist

It should use the venv’s pip, if one is found.

Using the manifest API

You can use the distlib.manifest API to construct lists of files when creating distributions. This functionality is an improved version of the equivalent functionality in distutils, where it was not a public API.

You can create instances of the Manifest class to work with a set of files rooted in a particular directory:

>>> from distlib.manifest import Manifest
>>> manifest = Manifest('/path/to/my/sources')

This sets the base attribute to the passed in root directory. You can add one or multiple files using names relative to the base directory:

>>> manifest.add('abc')
>>> manifest.add_many(['def', 'ghi'])

As a result of the above two statements, the manifest will consist of '/path/to/my/sources/abc', '/path/to/my/sources/def' and '/path/to/my/sources/ghi'. No check is made regarding the existence of these files.

You can get all the files below the base directory of the manifest:

>>> manifest.findall()

This will populate the allfiles attribute of manifest with a list of all files in the directory tree rooted at the base. However, the manifest is still empty:

>>> manifest.files
>>> set()

You can populate the manifest – the files attribute – by running a number of directives, using the process_directive() method. Each directive will either add files from allfiles to files, or remove files from allfiles if they were added by a previous directive. A directive is a string which must have a specific syntax: malformed lines will result in a DistlibException being raised. The following directives are available: they are compatible with the syntax of MANIFEST.in files processed by distutils.

Consider the following directory tree:

├── keep
│   └── keep.txt
├── README.txt
└── subdir
    ├── lose
    │   └── lose.txt
    ├── somedata.txt
    └── subsubdir
        └── somedata.bin

This will be used to illustrate how the directives work, in the following sections.

The include directive

This takes the form of the word include (case-sensitive) followed by a number of file-name patterns (as used in MANIFEST.in in distutils). All files in allfiles matching the patterns (considered relative to the base directory) are added to files. For example:

>>> manifest.process_directive('include R*.txt LIC* keep/*.txt')

This will add README.txt, LICENSE and keep/keep.txt to the manifest.

The exclude directive

This takes the form of the word exclude (case-sensitive) followed by a number of file-name patterns (as used in MANIFEST.in in distutils). All files in files matching the patterns (considered relative to the base directory) are removed from files. For example:

>>> manifest.process_directive('exclude LIC*')

This will remove ‘LICENSE’ from the manifest, as it was added in the section above.

The global-include directive

This works just like include, but will add matching files at all levels of the directory tree:

>>> manifest.process_directive('global-include *.txt')

This will add subdir/somedata.txt and subdir/lose/lose.txt from the manifest.

The global-exclude directive

This works just like exclude, but will remove matching files at all levels of the directory tree:

>>> manifest.process_directive('global-exclude l*.txt')

This will remove subdir/lose/lose.txt from the manifest.

The recursive-include directive

This directive takes a directory name (relative to the base) and a set of patterns. The patterns are used as in global-include, but only for files under the specified directory:

>>> manifest.process_directive('recursive-include subdir l*.txt')

This will add subdir/lose/lose.txt back to the manifest.

The recursive-exclude directive

This works like recursive-include, but excludes matching files under the specified directory if they were already added by a previous directive:

>>> manifest.process_directive('recursive-exclude subdir lose*')

This will remove subdir/lose/lose.txt from the manifest again.

The graft directive

This directive takes the name of a directory (relative to the base) and copies all the names under it from allfiles to files.

The prune directive

This directive takes the name of a directory (relative to the base) and removes all the names under it from files.

Next steps

You might find it helpful to look at information about Distlib’s design – or peruse the API Reference.