HUBFS is a file system for GitHub and Git. Git repositories and their contents are represented as regular directories and files and are accessible by any application, without the application having any knowledge that it is really accessing a remote Git repository. The repositories are writable and allow editing files and running build operations.
HUBFS is a command-line program with a usage documented below. On Windows there is also good desktop integration so that you can use HUBFS without the command line.
HUBFS supports both authenticated and non-authenticated access to repositories. When using HUBFS without authentication, only public repositories are available. When using HUBFS with authentication, both public and private repositories become available; an additional benefit is that the rate limiting that GitHub does for certain operations is relaxed.
In order to mount HUBFS issue the command hubfs MOUNTPOINT
. For example, hubfs H:
on Windows or hubfs mnt
on macOS and Linux.
The first time you run HUBFS you will be prompted to authorize it with GitHub:
> ./hubfs H:
First, copy your one-time code: XXXX-XXXX
Then press [Enter] to continue in the web browser...
HUBFS will then open your system browser where you will be able to authorize it with GitHub. HUBFS will store the resulting authorization token in the system keyring (Windows Credential Manager, macOS Keychain, etc.). Subsequent runs of HUBFS will use the authorization token from the system keyring and you will not be required to re-authorize the application.
To unmount the file system simply use Ctrl-C. On macOS and Linux you may also be able to unmount using umount
or fusermount -u
.
The full HUBFS command line usage is as follows:
usage: hubfs [options] [remote] mountpoint
-auth method
method is from list below; auth tokens are stored in system keyring
- force perform interactive auth even if token present
- full perform interactive auth if token not present (default)
- required auth token required to be present
- optional auth token will be used if present
- none do not use auth token even if present
- token=T use specified auth token T; do not use system keyring
-authkey name
name of key that stores auth token in system keyring
-authonly
perform auth only; do not mount
-d debug output
-filter rules
list of rules that determine repo availability
- list form: rule1,rule2,...
- rule form: [+-]owner or [+-]owner/repo
- rule is include (+) or exclude (-) (default: include)
- rule owner/repo can use wildcards for pattern matching
-o options
FUSE mount options
(default: uid=-1,gid=-1,rellinks,FileInfoTimeout=-1)
-version
print version information
(The default FUSE mount options depend on the OS. The uid=-1,gid=-1
option specifies that the owner/group of HUBFS files is determined by the user/group that launches the file system. This works on Windows, Linux and macOS.)
By default HUBFS presents the following file system hierarchy: / owner / repository / ref / path
Owner represents the owner of repositories under GitHub. It may be a user or organization. An owner is presented as a subdirectory of the root directory and contains repositories. However the root directory cannot be listed, because there are far too many owners to list.
Repository represents a repository owned by an owner. A repository is presented as a directory that contains refs.
Ref represents a git "ref". It may be a git branch, a git tag or even a commit hash. A ref is presented as a directory that contains repository content. However when listing a repository directory only branch refs are listed.
Path is a path to actual file content within the repository.
HUBFS interprets submodules as symlinks. These submodules can be followed if they point to other GitHub repositories. General repository symlinks should work as well. (On Windows you must use the FUSE option rellinks
for this to work correctly.)
With release 2022 Beta1 HUBFS ref directories are now writable. This is implemented as a union file system that overlays a read-write local file system over the read-only Git content. This scheme allows files to be edited and builds to be performed. A special file named .keep
is created at the ref root (full path: / owner / repository / ref / .keep
). When the edit/build modifications are no longer required the .keep
file may be deleted and the ref root will be garbage collected when not in use (i.e. when no files are open in it -- having a terminal window open with a current directory inside a ref root counts as an open file and the ref will not be garbage collected).
When you use the MSI installer under Windows there is better integration of HUBFS with the rest of the system:
There is a "Start Menu > HUBFS > Perform GitHub auth" shortcut that allows you to authorize HUBFS with GitHub without using the command line.
You can mount HUBFS drives using the Windows Explorer "Map Network Drive" functionality. To dismount use the "Disconnect Network Drive" functionality. (It is recommended to first authorize HUBFS with GitHub using the above mentioned shortcut.)
You can also mount HUBFS with the net use
command. The command net use H: \\hubfs\github.com
will mount HUBFS as drive H:
. The command net use H: /delete
will dismount the H:
drive.
In order to build HUBFS run build/make
. The build prerequisites for individual platforms are listed below:
Windows: Go 1.16, WinFsp, gcc (e.g. from Mingw-builds)
macOS: Go 1.16, FUSE for macOS, command line tools
Linux: Prerequisites: Go 1.16, libfuse-dev, gcc
HUBFS is a cross-platform file system written in Go. Under the hood it uses cgofuse over either WinFsp on Windows, macFUSE on macOS or libfuse on Linux. It also uses go-git for some git functionality.
HUBFS interfaces with GitHub using the REST API. The REST API is used to discover owners and repositories in the file system hierarchy, but is not used to access repository content. The REST API is rate limited (details).
HUBFS uses the git pack protocol to access repository content. This is the same protocol that git uses during operations like git clone
. HUBFS uses some of the newer capabilities of the pack protocol that allow it to fetch content on demand. HUBFS does not have to see all of the repository history or download all of the repository content. It will only download the commits, trees and blobs necessary to back the directories and files that the user is interested in. Note that the git pack protocol is not rate limited.
HUBFS caches information in memory and on local disk to avoid the need to contact the servers too often.
HUBFS uses the git pack protocol to fetch repository refs and objects. When HUBFS first connects to the Git server it fetches all of the server's advertised refs. HUBFS exposes these refs as subdirectories of a repository.
When accessing the content of a ref for the first time, the commit object pointed by the ref is fetched, then the tree object pointed by the commit is fetched. When fetching a tree HUBFS will also fetch all blobs directly referenced by the tree, this is required to compute proper stat
data (esp. size) for files.
HUBFS fetches objects with a depth of 1 and a filter of tree:0
. This ensures that the git server will only send objects whose hashes have been explicitly requested. This avoids sending extraneous information and speeds up communication with the server.
Consider a program that accesses files under /COMMON-NAME/DIR
. The owner of the COMMON-NAME
GitHub account could create a repository named DIR
and inject arbitrary file content into the program's process. This problem is particularly important when mounting the file system as a drive on Windows. To fix this problem:
Run HUBFS with the -filter
option. For example, running HUBFS with -filter ORG
will make available the repositories in ORG
only. Files within the file system will be accessible as / ORG
/ repository / ref / path. (The -filter
option allows the inclusion/exclusion of multiple repositories and supports wildcard syntax. Please see the HUBFS usage for more.)
Run HUBFS with a "prefix". For example, the (Windows) command line ./hubfs github.com/ORG H:
will place the root of the file system within ORG
and thus make available the repositories in ORG
only. Files within the file system will be accessible as / repository / ref / path.
The file system does not present a .git
subdirectory. It may be worthwhile to present a virtual .git
directory so that simple Git commands (like git status
) would work.
Additional providers such as GitHub Enterprise, BitBucket, GitLab, etc.
This project is licensed under the GNU Affero General Public License version 3.