Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?


Failed to load latest commit information.
Latest commit message
Commit time
October 18, 2022 18:02
December 27, 2022 18:17
January 3, 2023 15:12
June 24, 2021 19:33
October 14, 2022 17:53
October 15, 2022 02:40
February 11, 2021 21:22
January 2, 2023 10:15
February 6, 2023 15:01
January 3, 2023 15:12
January 3, 2023 15:12


📦 Package Node.js applications into executable binaries 📦

Source Package Continuous Integration

Watch the Video Demonstration


Why Package Node.js Applications into Executable Binaries?

  • Simple deploys. Transfer the binary into a machine and run it.
  • Let users test an application even if they don’t have Node.js installed.
  • Simple installation story for command-line applications.
  • It’s like the much-praised distribution story of Go programs, but for Node.js.


  • Works on Windows, macOS (Intel & ARM), and Linux (Intel, ARM6, ARM7, ARM64).
  • Simple to use. npm install caxa and call caxa from the command line. No need to declare which files to include; no need to bundle the application into a single file.
  • Supports any kind of Node.js project, including those with native modules (for example, sharp, @leafac/sqlite (shameless plug!), and others).
  • Works with any Node.js version.
  • Packages in seconds.
  • Relatively small binaries. A “Hello World!” application is ~30MB, which is terrible if compared to Go’s ~2MB, and worse still if compared to C’s ~50KB, but best-in-class if compared to other packaging solutions for Node.js.
  • Produces .exes for Windows, simple binaries for macOS/Linux, and macOS Application Bundles (.app).
  • Based on a simple but powerful idea. Implemented in ~200 lines of code.
  • No magic. No traversal of require()s trying to find which files to include; no patches to Node.js source.


  • Doesn’t patch the Node.js source code.
  • Doesn’t build Node.js from source.
  • Doesn’t support cross-compilation (for example, building a Windows executable from a macOS development machine).
  • Doesn’t support packaging with a Node.js version different from the one that’s running caxa (for example, bundling Node.js 15 while running caxa with Node.js 14).
  • Doesn’t hide your JavaScript source code in any way.


$ npm install --save-dev caxa


Prepare the Project for Packaging

  • Install any dependencies with npm install or npm ci.
  • Build. For example, compile TypeScript with tsc, bundle with webpack, and whatever else you need to get the project ready to start. Typically this is the kind of thing that goes into an npm prepare script, so the npm ci from the previous point may already have taken care of this.
  • If there are files that shouldn’t be in the package, remove them from the directory. For example, you may wish to remove the .git directory.
  • You don’t need to npm dedupe --production, because caxa will do that for you from within the build directory. (Otherwise, if you tried to npm dedupe --production you’d uninstall caxa, which should probably be in devDependencies.)
  • It’s recommended that you run caxa on a Continuous Integration server. (GitHub Actions, for example, does a shallow fetch of the repository, so removing the .git directory becomes negligible—but you can always do that with the --exclude advanced option.)

Call caxa from the Command Line

$ npx caxa --help
Usage: caxa [options] <command...>

Package Node.js applications into executable binaries

  command                                The command to run and optional arguments to pass to the command every time the executable is called. Paths must be absolute. The ‘{{caxa}}’ placeholder is substituted for the folder from which the package runs. The ‘node’ executable is available at ‘{{caxa}}/node_modules/.bin/node’. Use double quotes to delimit the command and each argument.

  -i, --input <input>                    [Required] The input directory to package.
  -o, --output <output>                  [Required] The path where the executable will be produced. On Windows, must end in ‘.exe’. In macOS and Linux, may have no extension to produce regular binary. In macOS and Linux, may end in ‘.sh’ to use the Shell Stub, which is a bit smaller, but depends on some tools being installed on the end-user machine, for example, ‘tar’, ‘tail’, and so forth. In macOS, may end in ‘.app’ to generate a macOS Application Bundle.
  -F, --no-force                         [Advanced] Don’t overwrite output if it exists.
  -e, --exclude <path...>                [Advanced] Paths to exclude from the build. The paths are passed to and paths that match will be excluded. [Super-Advanced, Please don’t use] If you wish to emulate ‘--include’, you may use ‘--exclude "*" ".*" "!path-to-include" ...’. The problem with ‘--include’ is that if you change your project structure but forget to change the caxa invocation, then things will subtly fail only in the packaged version.
  -D, --no-dedupe                        [Advanced] Don’t run ‘npm dedupe --production’ on the build directory.
  -p, --prepare-command <command>        [Advanced] Command to run on the build directory after ‘npm dedupe --production’, before packaging.
  -N, --no-include-node                  [Advanced] Don’t copy the Node.js executable to ‘{{caxa}}/node_modules/.bin/node’.
  -s, --stub <path>                      [Advanced] Path to the stub.
  --identifier <identifier>              [Advanced] Build identifier, which is part of the path in which the application will be unpacked.
  -B, --no-remove-build-directory        [Advanced] Remove the build directory after the build.
  -m, --uncompression-message <message>  [Advanced] A message to show when uncompressing, for example, ‘This may take a while to run the first time, please wait...’.
  -V, --version                          output the version number
  -h, --help                             display help for command

  > caxa --input "examples/echo-command-line-parameters" --output "echo-command-line-parameters.exe" -- "{{caxa}}/node_modules/.bin/node" "{{caxa}}/index.mjs" "some" "embedded arguments" "--an-option-thats-part-of-the-command"

  $ caxa --input "examples/echo-command-line-parameters" --output "echo-command-line-parameters" -- "{{caxa}}/node_modules/.bin/node" "{{caxa}}/index.mjs" "some" "embedded arguments" "--an-option-thats-part-of-the-command"

  macOS/Linux (Shell Stub):
  $ caxa --input "examples/echo-command-line-parameters" --output "" -- "{{caxa}}/node_modules/.bin/node" "{{caxa}}/index.mjs" "some" "embedded arguments" "--an-option-thats-part-of-the-command"

  macOS (Application Bundle):
  $ caxa --input "examples/echo-command-line-parameters" --output "Echo Command Line" -- "{{caxa}}/node_modules/.bin/node" "{{caxa}}/index.mjs" "some" "embedded arguments" "--an-option-thats-part-of-the-command"

Here’s a real-world example of using caxa. This example includes packaging for Windows, macOS, and Linux; distributing tags with GitHub Releases Assets; distributing Insiders Builds for every push with GitHub Actions Artifacts; and deploying a binary to a server with rsync (and publishing an npm package as well, but that’s beyond the scope of caxa).

Call caxa from TypeScript/JavaScript

Instead of calling caxa from the command line, you may prefer to write a program that builds your application, for example:

import caxa from "caxa";

(async () => {
  await caxa({
    input: "examples/echo-command-line-parameters",
    output: "echo-command-line-parameters",
    command: [
      "embedded arguments",

You may need to inspect process.platform to determine in which operating system you’re running and come up with the appropriate parameters.

Fine Points

Calling an Executable That Isn’t node

If you wish to run a command that isn’t node, for example, ts-node, you may do so by extending the PATH. For example, you may run the following on macOS/Linux:

$ caxa --input <input> --output <output> -- "env" "PATH={{caxa}}/node_modules/.bin/:\$PATH" "ts-node" "{{caxa}}/index.ts"

Preserving the Executable Mode of the Binary

This is only an issue on macOS/Linux. In these operating systems a binary must have the executable mode enabled in order to run. You may check the mode from the command line with ls -l: on an output that reads like -rwxr-xr-x [...]/bin/node, the xs represent that the file is executable.

Here’s what you may do when you distribute the binary to ensure that the file mode is preserved:

  1. Create a tarball or zip. The file mode is preserved through compression/decompression, and macOS/Linux (most distributions, anyway) come out of the box with software to uncompress tarballs and zips—the user can just double-click on the file.

    You may generate a tarball with, for example, the following command:

    $ tar -czf <caxa-output>.tgz <caxa-output>

    Fun fact: Windows 10 also comes with the tar executable, so the command above works on Windows as well. Unfortunately the File Explorer on Windows doesn’t support uncompressing the .tgz with a double-click (it supports uncompressing .zip, however). Fortunately, Windows doesn’t have issues with file modes to begin with (it simply looks for the .exe extension) so distributing the caxa output directly is appropriate.

  2. Fix the file mode after downloading. Tell your users to run the following command:

    $ chmod +x <path-to-downloaded-application>

    In some contexts this may make more sense, but it requires your users to use the command line.

Detect Whether the Application Is Running from the Packaged Version

caxa doesn’t do anything special to your application, so there’s no built-in way of telling whether the application is running from the packaged version. It’s part of caxa’s ethos of being as out of the way as possible. Also, I consider it to be a bad practice: an application that is so self-aware is more difficult to reason about and test.

That said, if you really need to know whether the application is running from the packaged versions, here are some possible workarounds in increasing levels of badness:

  1. Set an environment variable in the command, for example, "env" "CAXA=true" "{{caxa}}/node_modules/.bin/node" "...".
  2. Have a different entrypoint for the packaged application, for example, "{{caxa}}/node_modules/.bin/node" "caxa-entrypoint.js".
  3. Receive a command-line argument that you embed in the packaging process, for example, "{{caxa}}/node_modules/.bin/node" "application.js" "--caxa".
  4. Check whether __dirname.startsWith(path.join(os.tmpdir(), "caxa")).

The Current Working Directory

Even though the code for the application is in a temporary directory, the current working directory when calling the packaged application is preserved, and you may inspect it with process.cwd(). This is probably not something you have to think about—caxa just gets it right.

How It Works

The Issue

As far as I can understand, the root of the problem with creating binaries for Node.js projects is native modules. Native modules are libraries written at least partly in C/C++, for example, sharp, @leafac/sqlite (shameless plug!), and others. There are at least three issues with native modules that are relevant here:

  1. You must have a working C/C++ build system to install these libraries (C/C++ compiler, make, Python, and so forth). On Windows, you must install windows-build-tools. On macOS, you must install the Command-Line Tools (CLT) with xcode-select --install. On Linux, it depends on the distribution, but on Ubuntu sudo apt install build-essential is enough.

  2. The installation of native modules isn’t cross-platform. Unlike JavaScript dependencies, which you may copy from an operating system to another, native modules produce compiled C/C++ code that’s specific to the operating system on which the dependency is installed. This compiled code appears in your node_modules directory in the form of .node files.

  3. As far as I understand, Node.js insists on loading native modules from files in the disk. Other Node.js packaging solutions get around this limitation in one of two ways: They either patch Node.js to trick it into loading native modules differently; or they put .node files somewhere before starting your program.

The Solution

caxa builds on the idea of putting .node files in a temporary location, but takes it to ultimate consequence: a caxa executable is a form of self-extracting archive containing your whole project along with the node executable. When you first run a binary produced by caxa, it extracts the source the whole project (and the bundled node executable) into a temporary location. From there, it simply calls whatever command you told it to run when you packaged the project.

At first, this may seem too costly, but in practice it’s mostly okay: It doesn’t take too long to uncompress a project in the first place, and caxa doesn’t clean the temporary directory after running your program, so subsequent calls are effectively cached and run without overhead.

This idea is simple, but it’s super powerful! caxa supports any kind of project, including those with native dependencies, because running a caxa executable amounts to the same as installing Node.js on the user’s machine. caxa produces packages fast, because generating a self-extracting archive is a simple matter of concatenating some files. caxa supports any version of Node.js, because it simply copies the node executable with which it was called into the self-extracting archive.

Fun fact: By virtue of compressing the archive, caxa produces binaries that are naturally smaller when compared to other packaging solutions. Obviously, you could achieve the same outcome by compressing the output of these other tools, which may want to do anyway to preserve the file mode (see § Preserving the Executable Mode of the Binary).

How the Self-Extracting Archive Works

Did you know that you may append anything to a binary and it’ll continue to work? This is true of binaries for Windows, macOS, and Linux. Here’s an example to try out on macOS/Linux:

$ cp $(which ls) ./ls  # Copy the ‘ls’ binary into the current directory to play with it
$ ./ls                 # List the files, proving the that the binary works
$ echo ANYTHING >> ls  # Append material to the binary
$ tail ./ls            # You should see ‘ANYTHING’ at the end of the output
$ ./ls                 # The output should be same as before!
$ rm ls                # Okay, the test is over

The caxa self-extracting archives work by putting together three parts: 1. a stub; 2. an archive; and 3. a footer. This is the layout of these parts in the binary produced by caxa:


The STUB and the ARCHIVE are separated by the CAXACAXACAXA string. And the ARCHIVE and the FOOTER are separated by a newline. This layout allows caxa to find the footer by simply looking backward from the end of the file until it reaches a newline. And if this is the first time you’re running the caxa executable and the archive needs to be uncompressed, then caxa may find the beginning of the ARCHIVE by looking forward from the beginning until it reaches the CAXACAXACAXA separator.

Build a binary with caxa and inspect it yourself in a text editor (Visual Studio Code asks you to confirm that you want to open a binary, but works fine after that). You should be able to find the CAXACAXACAXA separator between the STUB and the ARCHIVE, as well as the FOOTER at the end.

Let’s examine each of the parts in detail:

Part 1: Stub

This is a program written in Go that:

  1. Reads itself as a file.
  2. Finds the footer.
  3. Determines whether it’s necessary to extract the archive.
    1. If so, finds the archive.
    2. Extracts it.
  4. Runs whatever command it’s told in the footer.

You may find the source code for the stub in stubs/stub.go. You may build the stub with npm run prepare:stubs. You will need a Go compiler, but the stub has no dependencies beyond the Go standard library, so there’s no need to setup Go modules or configure a $GOPATH. There are pre-compiled stubs for the major platforms in the npm package. If you wish to verify that the stubs really were compiled from stubs/stub.go, you may recompile it yourself, because the Go compiler appears to be deterministic and always produce the same binaries given the same source (at least that’s what happened in my tests).

This is beautiful in a way: We’re using Go’s ability to produce binaries to bootstrap Node.js’s ability to produce binaries.

Part 2: Archive

This is a tarball of the directory with your project.

Part 3: Footer

This is JSON containing the extra information that caxa needs to run your project: Most importantly, the command that you want to run, but also an identifier for where to uncompress the archive.

Using the Self-Extracting Archive without caxa

Fun fact: There’s nothing Node.js-specific about the stubs. You may use them to uncompress any kind of archive and run any arbitrary command on the output! And it’s relatively straightforward to build a self-extracting archive from scratch. For example, you may run the following in macOS:

$ cp stub an-ls-caxa
$ tar -czf - >> an-ls-caxa
$ printf "\n{ \"identifier\": \"an-ls-caxa/AN-ARBITRARY-STRING-THAT-SHOULD-BE-DIFFERENT-EVERY-TIME\", \"command\": [\"ls\", \"{{caxa}}\"] }" >> an-ls-caxa
$ ./an-ls-caxa

To Where Are the Packages Uncompressed at Runtime?

It depends on the operating system. You may find the location on your system with:

$ node -p "require(\"os\").tmpdir()"

Look for a directory named caxa in there.

Why No Cross-Compilation? Why No Different Versions of Node.js besides the Version with Which caxa Was Called?

Two reasons:

  1. I believe you should have environments to work with all the operating systems you plan on supporting. They may not be your main development environment, but they should be able to build your project and let you test things. At the very least, you should use a service like GitHub Actions which lets you run build tasks and tests on Windows, macOS, and Linux.

    (I, for one, bought a PC to work on caxa. Yet another reason to support my work!)

  2. The principle of least surprise. When cross-compiling (for example, building a Windows executable from a macOS development machine), or when bundling different versions of Node.js (for example, bundling Node.js 15 while running caxa with Node.js 14), there’s no straightforward way to guarantee that the packaged project will run the same as the unpackaged version. If you aren’t using any native modules then things may work, but as soon as you introduce a new dependency that you didn’t know was native your application may break. Not only are native dependencies different on the operating systems, but they may also be different between different versions of Node.js if these versions aren’t ABI-compatible (which is why sometimes when you update Node.js you must run npm install again).

Fun fact: The gold-standard for easy cross-compilation these days is Go. But even in Go cross-compilation goes out the window as soon as you introduce C dependencies (something called CGO). It appears that many people in the Go community try to solve the issue by avoiding CGO dependencies, sometimes going to great lengths to reinvent everything in pure Go. On the one hand, this sounds like fun when it works out. On the other hand, it’s a huge case of not-invented-here syndrome. In any case, native modules seem to be much more prevalent in Node.js than CGO is in Go, so I think that cross-compilation in caxa would be a fool’s errand.

If you still insist on cross-compiling or compiling for different versions of Node.js, you can still use the stub to build a self-extracting archive by hand (see § Using the Self-Extracting Archive without caxa). You may even use to more easily bundle different versions of Node.js.

How the macOS Application Bundles (.app) Work

An macOS Application Bundle is just a folder with a particular structure and an executable at a particular place. When creating a macOS Application Bundle caxa doesn’t build a self-extracting archive, instead it just copies the application to the right place and creates an executable bash script to start the process.

The macOS Application Bundle may be run by simply double-clicking on it from Finder. It opens a window with your application. If you’re running an application that wasn’t built on your machine (which is most likely the case for your users, who probably downloaded the application from the internet), then the first time you run it macOS will probably complain about the lack of a signature. The solution is to go to System Preferences > Security & Privacy > General and click on Allow. You must instruct your users on how to do this.

How the Shell Stub (.sh) Works

It’s equivalent to the Go stub, except that it is smaller, because it’s just a dozen lines of Bash, but it depends on some things being installed on the end-user machine, for example, tar, tail, and so forth.

Features to Consider Implementing in the Future

If you’re interested in one of these features, please send a Pull Request if you can, or at least reach out to me and mention your interest, and I may get to them.

  1. Other compression algorithms. Currently caxa uses tarballs, which are ubiquitous and reasonably efficient in terms of compression/uncompression times and archive size. But there are better algorithms out there… (See #1.)

  2. Add support for signing the executables. There are limitations on the kinds of executables that are signable, and a self-extracting archive of the kind that caxa produces may be unsignable (I know very little about this…). A solution could be use Go’s support for embedding data in the binary (which landed in Go 1.16). Of course this would require the person packaging a project to have a working Go build system. Another solution would be to manipulate the executables as data structures, instead of just appending stuff at the end. Go has facilities for this in the standard library, but then the packager itself (not only the stubs) would have to be written in Go, and creating packages on the command line by simply concatenating files would be impossible.

  3. Add support for custom icons and other package metadata. This should be relatively straightforward by using rcedit for .exes and by adding .plist files to .apps (we may copy whatever Electron is doing here as well).

Prior Art

Here’s my preliminary research: vercel/pkg#837 (comment)

Below follows the extended version with everything I learned along the way of building caxa.


Deno has experimental support for producing binaries. I haven’t tried it myself, but maybe one day it catches on and caxa becomes obsolete. Let’s hope for that!

pkg is great, and it’s where I first learned that you could think about compiling Node.js projects this way. It’s the most popular packaging solution for Node.js by a long shot.

It works by patching the Node.js executable with a proxy around fs. This proxy adds the ability to look into something called a snapshot file system, which is where your project is stored. Also, it doesn’t store your source JavaScript directly. It runs your JavaScript through the V8 compiler and produces a V8 snapshot, which has two nice consequences: 1. Your code will start marginally faster, because all the work of parsing the JavaScript source and so forth is already done; and 2. Your code doesn’t live in the clear in the binary, which may be advantageous if you want to hide it.

Unfortunately, this approach has a few issues:

  1. The Node.js patches must be kept up-to-date. For example, when fs/promises became a thing, the fs proxy didn’t support it. It was a subtle and surprising issue that only arises in the packaged version of the application. (For the fix, see my fork of pkg, @leafac/pkg (which has been deprecated now that caxa has been released).)

  2. The patched Node.js distributions must be updated with each new Node.js release. At the time of this writing they’re lagging behind by half an year (v14.4.0, while the latest LTS is v14.16.0). That’s new features and security updates you may not be getting. (See for a seemingly abandoned attempt at automating the patching process that could improve on this situation. Of course, manual intervention would still be required every time the patches become incompatible with Node.js upstream.)

  3. Native modules work by the way of a self-extracting archive.

Also, pkg traverses the source code for your application and its dependencies looking for things like require()s to prune code that isn’t used. This is good if you want to optimize for small binaries with little effort. But often this process goes wrong, specially when something like TypeScript produces JavaScript that throws off pkg’s heuristics. In that case you have to intervene and list the files that should be included by hand.

Not to mention that the maintainers of pkg haven’t been super responsive this past year. (And who can blame them? Open-source is hard. No shade thrown here; pkg is awesome! And speaking of “open-source is hard,” support my work!)

The second most popular packaging solution in Node.js. nexe works by a similar strategy, and suffers from some of the same issues. But fs/promises work, newer Node.js versions are available, and the project seems to be maintained more actively.

Native modules don’t work, but there’s a workaround based on the idea of self-extracting archives:

This works with a different strategy. Node.js has a part of the standard library written in JavaScript itself, and when Node.js is built, this JavaScript ends up embedded as part of the node executable. boxednode works by recompiling Node.js from source with your project embedded as if it were part of the standard library. On the upside, this supports native extensions and whatever new fs/promises situation comes up in the future. The down side is that compiling Node.js takes hours (the first time, and still a couple minutes after the subsequent times) and 10+GB of disk(!) Also, boxednode only works with a single JavaScript file, so you must bundle with something like ncc or webpack before packaging. And I don’t think it handles assets like images along with the code, which would be essential when packaging a web application.

This works with an idea of a snapshot file system (à la pkg), but it follows a more principled approach for that, using something called Squashfs. To the best of my knowledge the native-extensions story in node-packer is the same self-extracting archive from most packaging solutions. The downside of node-packer is that installing and setting it up is a bit more involved than a simple npm install. For that reason I ended up not really giving it a try, so I’ll say no further…

This should work with a strategy similar to boxednode, but with a pre-compiled binary including some pre-allocated space to save you from having to compile Node.js from source. Like boxednode, it should handle only a single JavaScript file, requiring a bundler like ncc or webpack. I tried js2bin and it produced binaries that didn’t work at all. I have no idea why…

The predecessor of pkg. Worked with the same idea. I believe it has been deprecated in favor of pkg. To the best of my knowledge it was closed source and paid.

This is the project that gave me the idea for caxa! It’s more obscure, so at first I payed it little attention in my investigation. But then it handled native extensions and the latest Node.js versions out-of-the-box despite haven’t been updated in 4 years! I was delighted and intrigued!

In principle, nar works the same as caxa, using the idea of a self-extracting archive. There are some important differences, though:

  1. nar doesn’t support Windows. That’s because nar’s stub is a bash script instead of the Go binary used in caxa.
  2. nar gets some small details wrong. For example, it changes your current working directory to the temporary directory in which the archive is uncompressed. This breaks some assumptions about how command-line tools should work; for example, if you’re project implements ls in Node.js, then when running it from nar it’d always list the files in the temporary directory.
  3. It’s no longer maintained. They recommend pkg instead.
  4. It was written in LiveScript, which is significantly more obscure than TypeScript/Go, in which caxa is implemented.

Similar to nar. Hasn’t seen activity in 8 years.

Other Packages

If you dig through npm, GitHub, and Google, you’ll find other projects in this space, but I couldn’t find one that had a good combination of working well, being well documented, being well maintained, and so forth.

A similar idea to caxa from our friends in Elixir community.

References on Self-Extracting Archives

Creating a self-extracting archive with a bash script for the stub (only works on macOS/Linux, and depends on things like tar being available—which they probably are) (this is the inspiration for the Shell Stub):

Creating a self-extracting batch file for Windows (an idea I didn’t pursue, going for the Go stub instead):

Other tools that create self-extracting archives:

References on Building the Stub in C

Besides Go, I also considered writing the stub in C. Ultimately Go won because it’s less prone to errors and has a better cross-compilation/standard-library story. But C has the advantage of being setup in the machines of Node.js developers because of native dependencies. You could leverage that to use the linker (ld) to embed the archive, instead of crudely appending it to the end of the stub. This could be necessary to handle signing…

Anyway, here’s what you could use to build a stub in C:

References on Creating Self-Extracting Archives in Node.js

References on the Structure of Executables

A more principled way of building the self-extracting archive is to not append data at the end of the file, but manipulate the stub binary as a data structure. It’s actually three data structures: Portable Executables (Windows), Mach-O (macOS), and ELF (Linux). This idea was abandoned because it’s more work for the packager and for the stub—the CAXACAXACAXA separator is a hack that works well enough. But we may have to revisit this to make the executables signable. You can even manipulate binaries with Go standard libraries…

Anyway, here are some references on the subject:

References on Just Appending Data to an Executable Works

The data that you append is sometimes called an overlay.

References on Cross-Compilation of CGO

References on Building macOS Application Bundles (.app)

References on How to Untar in Go

The Go standard library has low-level utilities for handling tarballs. I could have used a higher-level library, but I couldn’t get them to work with an archive that’s in memory (having been extracted from the binary). Besides, relying only on the standard library is good for an easy compilation story. In the end, the solution was to copy and paste a bunch.

References on How to Execute a Command from Go

It’d have been nice to use syscall.Exec(), which replaces the currently running binary (the stub) with another one (the command you want to run for your application), but syscall.Exec() is macOS/Linux-only. So we use os.Exec() instead, paying attention to wiring stdin/stdout/stderr between the processes, and forwarding the command-line arguments on the way and the status code on the way out. The downside is that there’s an extra process in the process tree.

References on the Layout of the Data in the Self-Extracting Archive

What’s up with This Name?

caxa is a misspelling of caixa, which is Portuguese for box. I find it amusing to say that you’re putting an application in the caxa 📦 🙄


As you see from this long README, despite being simple in spirit, caxa is the result of a lot of research and hard work. Simplicity is hard. So support my work.



v3.0.1 · 2022-10-15

Hotfix of v3.0.0 whose binary detection didn’t work when executed from npx 🤦‍♂️

v3.0.0 · 2022-10-15

This version modernizes the codebase in preparation to add new features, for example, the much anticipated #60.

It doesn’t add or remove features, yet it’s a major version bump because it includes the following backwards-incompatible changes:

  • caxa is now ESM-only.
  • We dropped support for Node.js 14.


  • Added a new stub strategy: Shell Stub. It’s a simple Bash script that does the same as the Go stub, except that it takes less space (about 10 lines as opposed to a 2MB Go binary), but it depends on some tools being installed on the end-user machine, for example, tar, tail, and so forth.
  • Simplified the build/distribution of stubs.
    • Cross-compile the stubs, to simplify the GitHub Actions architecture. We were already cross-compiling for macOS ARM, so that isn’t a big loss. Most bugs that may arise from this decision would be bugs in the Go cross-compiler, which is unlikely.
    • Distribute the stubs with the npm package to avoid issues like: #26, #28, #31, and #32. If you wish to verify that the stubs really were compiled from stubs/stub.go, you may run npm run prepare:stubs, because the Go compiler appears to be deterministic and always produce the same binaries given the same source (at least that’s what happened in my tests).
    • Check the stubs in version control, to simplify distribution and the workflow of people who want to help in the JavaScript part of caxa and who may not want to setup Go.
    • Distribute only one version of the stub for Linux ARMv6 & Linux ARMv7. They happened to be the same binary, anyway—it seems that the Go compiler doesn’t differentiate between these architectures.
  • Added the --stub advanced option to specify a custom stub.
  • Added documentation on how to emulate an --include option, and why that’s probably a bad idea.
  • Added the --uncompression-message advanced option to print a message when uncompressing.
  • Fixed the --exclude advanced option in Windows.


  • Added support for ARM, both on Linux and macOS. (Thanks @maxb2!)

  • Fixed inconsistent application directory state that would happen if you stopped caxa in the middle of the extraction or if multiple extractions are attempted at once.

  • Added several command-line parameters to customize the build.

  • [BREAKING]: Changed the command-line parameters:

    From To
    --directory --input
    --command The last arguments passed to caxa. Use -- to separate if your command includes something that starts with -.