Building the NDK

The latest version of this document is available at

Both Linux and Windows NDKs are built on Linux machines. Windows host binaries are cross-compiled with MinGW.

Building the NDK for Mac OS X requires at least 10.13.


The first thing you need is the AOSP NDK repository. If you‘re new to using repo and gerrit, see for tips. If you’re already familiar with how to use repo and gerrit from other Android projects, you already know plenty :)

Check out the branch master-ndk. Do this in a new directory.

# For non-Googlers:
repo init -u -b master-ndk

# Googlers, follow http://go/repo-init/master-ndk (select AOSP in the Host menu,
# and uncheck the box for the git superproject). At time of writing, the correct
# invocation is:
repo init -u \
    sso:// -b master-ndk

If you wish to rebuild a given release of the NDK, the release branches can also be checked out. They're named ndk-release-r${RELEASE} for newer releases, but ndk-r{RELEASE}-release for older releases. For example, to check out the r19 release branch, use the -b ndk-release-r19 flag instad of -b master-ndk.

Linux dependencies are listed in the Dockerfile. You can use docker to build the NDK:

docker build -t ndk-dev infra/docker
docker run -it -u $UID -v `realpath ..`:/src -w /src/ndk ndk-dev ./

Building on Mac OS X has similar dependencies as Linux, but also requires Xcode.

Running tests requires that adb is in your PATH. This is provided as part of the Android SDK.

Python environment setup

To set up your environment to use the correct versions of Python and Python packages, install Poetry and then do the following.

Whenever you set up a new NDK tree (after a fresh repo init, for example), configure the project to use our prebuilt Python instead of your system's. If on Mac, be sure to use the darwin-x86 version instead.

poetry env use ../prebuilts/python/linux-x86/bin/python3

The first time, and also anytime you sync because there might be new or updated dependencies, install the NDK dependencies to the virtualenv managed by poetry.

poetry install

Spawn a new shell using the virtualenv that Poetry created. You could instead run NDK commands with the poetry run prefix (e.g. poetry run ./, but it‘s simpler to just spawn a new shell. Plus, if it’s in your environment your editor can use it.

poetry shell


For Linux or Darwin

$ python

If you get an error like the following:

Expected python to be $NDK_SRC/prebuilts/python/$HOST/bin/python3.9, but is ~/.cache/pypoetry/virtualenvs/$VENV/bin/python (/usr/bin/python3.9).

Your Poetry virualenv was misconfigured. It seems that poetry env use will not replace an existing virtualenv of the same major/minor version, so if you ran any poetry commands before poetry env use, your environment needs to be deleted and recreated.

$ poetry env remove $VENV
$ poetry env use ../prebuilts/python/linux-x86/bin/python3
$ poetry install

If you get an error like the following:

Expected python to be $NDK_SRC/prebuilts/python/linux-x86/bin/python3.9, but is /usr/bin/python (/usr/bin/python3.9).

You ran outside the poetry environment. Ensure that you've done the first time setup (poetry env use and poetry install, as above), then either run poetry shell to enter a new shell with the correct environment, or use poetry run

If you get errors from the pythonlint task but it appears to only affect your machine, one of the linters you have installed is probably not the correct version. Run poetry install to sync your environment with the expected versions.

For Windows, from Linux

$ python --system windows64 will also build all of the NDK tests. This takes about 3x as long as building the NDK itself, so pass --no-build-tests to skip building the tests if you're iterating on build behavior or plan to rebuild only specific tests. Tests can be built later with python --rebuild.

Note: The NDK's build and test scripts are implemented in Python 3 (currently 3.9). will use a prebuilt Python, but does not do this yet. also can be run outside of a complete development environment (as it is when it is run on Windows), so a Python 3.9 virtualenv is recommended.


Packaging uses zip -9 so is extremely time consuming and disabled by default. Use the --package flag to force packaging locally. This is not required for local development and only needs to be used when testing packaging behavior.