@dynamicCallable: Unix Tools as Swift Functions

A new feature in Swift 5 are Dynamic Callable’s. We combine this with the related Dynamic Member Lookup feature to expose the filesystem and Unix shell commands as regular Swift objects and functions.

Wait what?! We want to call arbitrary commandline tools from within Swift, like so:

import Shell

for file in shell.ls("/Users/").split(separator: "\n") {
    print("dir:", file)
// dir: Guest
// dir: Shared
// dir: helge

let swiftVersion = shell.swift("--version")
print(swiftVersion.split(separator: "\n").first ?? "")
// Apple Swift version 5.0-dev (LLVM fe02928dd1, Clang 8836e4e85c, Swift 468f5b0530)

print(shell.usr.local.bin.python("-c", "'print(13 + 37)'"))
// 42

Inspired by the Python sh module, any executable tool, from at to xsltproc, is made available as a first class Swift function. Magic 🦄.

Dynamic Callable essentially allows you to turn any type into a regular Swift “function”. Let’s see how this works!

You can follow along, or you can go ahead and grab Shell from GitHub.

Important: Remember that you need to have Swift 5 via Xcode 10.2.

SE-0195: Dynamic Member Lookup

Before we jump into the new dynamic callable feature, let us revisit the Dynamic Member Lookup introduced in Swift 4.2. Environment variable lookup is the default example:

public struct EnvironmentTrampoline {
  public subscript(dynamicMember k: String) -> String? {
    return ProcessInfo.processInfo.environment[k]

let env = EnvironmentTrampoline()

let path = env.PATH ?? "" // retrieve an env variable

So what this does is instead of having to write:


you can directly use that key like:


When the compiler tries to lookup the PATH “member”, it doesn’t find that in our struct. Usually that would result in a compile time error. But if the compiler sees that the type is marked up as @dynamicMemberLookup, it will instead replace the env.PATH with this call:

env[dynamicMember: "PATH"]

The environment example isn’t very exciting, but this also allows you to traverse nested structures, for example a generic JSON dictionary:

// rewritten to:
json[dynamicMember: "person"][dynamicMember: "address"][dynamicMember: "street"]

In our case, we use this feature for two things:

  1. to navigate the filesystem
  2. to dynamically lookup tools in the $PATH

In the spirit of the Environment example above, let us create a trampoline which allows us to traverse the filesystem using just Swift a.b.c syntax:

public struct ShellPathTrampoline {
    let url : URL
    var fm  : FileManager { return FileManager.default }
    public subscript(dynamicMember key: String) 
           -> ShellPathTrampoline 
        let url = {
            let url = self.url.appendingPathComponent(key)
            if !isDirectory(url) { return url }
            return self.url.appendingPathComponent(key, isDirectory: true)
        return ShellPathTrampoline(url: url)
    func isDirectory(_ url: URL) -> Bool {
        var isDir  : ObjCBool = false
        let exists = fm.fileExists(atPath: url.path, isDirectory: &isDir)
        return exists && isDir.boolValue

let fsRoot = ShellPathTrampoline(url: URL(fileURLWithPath: "/"))

print(fsRoot.usr.local.bin.python) // <==
// ShellPathTrampoline(url: file:///usr/local/bin/python)

Note how we give the path using dot syntax: usr.local.bin.python which is translated to:

fsRoot[dynamicMember: "usr"]    // yields the usr trampoline
      [dynamicMember: "local"]  // appends "local" to "usr"
      [dynamicMember: "bin"]    // and then "bin"
      [dynamicMember: "python"] // .. you get it

The code of our subscript looks a little complicated, which is mainly due to the Foundation API to detect a directory being a little awkward (and we need that to append a proper ending “/” to the URL, e.g. “/usr/”).

In short: This is just a simple, but Swift-integrated, URL builder.

Do $PATH lookups

This is already quite nice, but we also want to find tools by traversing the $PATH - if necessary. I.e. instead of having to call shell.usr.bin.ls, we also want this shortcut to work: shell.ls.

Let’s add another trampoline which can do the dynamic lookup:

extension ShellPathTrampoline {
    var doesExist : Bool {
        return fm.fileExists(atPath: url.path)

public struct ShellTrampoline {
    public let root : ShellPathTrampoline
    public var url  : URL { return root.url }
    public init(url: URL = URL(fileURLWithPath: "/")) {
        self.root = ShellPathTrampoline(url: url)
    public let environment = EnvironmentTrampoline()
    public subscript(dynamicMember key: String) 
           -> ShellPathTrampoline 
        let trampoline = root[dynamicMember: key]
        if trampoline.doesExist { return trampoline }
        return lookupInPATH(key) ?? trampoline
    func lookupInPATH(_ k: String) -> ShellPathTrampoline? {
        let searchPath = (environment.PATH ?? "/usr/bin")
                         .components(separatedBy: ":")
        let testURLs = searchPath.lazy.map { 
            ( path: String ) -> URL in
            let testDirURL = URL(fileURLWithPath: path, relativeTo: self.url)
            return testDirURL.appendingPathComponent(k)
        let fm = FileManager.default
        for testURL in testURLs {
            let testPath = testURL.path
            var isDir    : ObjCBool = false
            if fm.fileExists(atPath: testPath, isDirectory: &isDir) {
                if !isDir.boolValue && fm.isExecutableFile(atPath: testPath) {
                    return ShellPathTrampoline(url: testURL)
        return nil

public let shell = ShellTrampoline()

// ShellPathTrampoline(url: file:///usr/bin/python)

Notice how the simple shell.python is expanded to the full path: /usr/bin/python.

This trampoline accepts absolute pathes in the dynamicMember subscript. If the path passed-in does not exist, it will search the lookup pathes stored in the $PATH environment variable (usually looks like /usr/local/bin:/usr/bin:/sbin), using the lookupInPATH function (notice how we use our environment trampoline from above to access PATH).

Summary: Dynamic Member Lookup

We’ve shown how you can use Swift dot syntax to lookup values dynamically. And by that, we built a way to construct and even lookup filesystem pathes. But so far, we can’t really do anything with those ShellPathTrampoline values we get.

Remember that the compiler just rewrites this:


into this:

shell[dynamicMember: "python"]

SE-0216: Dynamic Callable

So let’s approach Dynamic Callable. Again, there is actually very little magic involved. Let us assume our shell.swift returns us a ShellPathTrampoline struct. That struct is obviously not a function and if we try to do this:


The compiler will rightfully complain:

Cannot call value of non-function type 'ShellPathTrampoline'

This is where Dynamic Callable steps in. It allows us to turn a non-function type into a function type by adding the @dynamicCallable attribute to the struct. Let’s do this:

@dynamicCallable     // <== add this!
public struct ShellPathTrampoline {
    ... code from above ...

    func dynamicallyCall(withArguments arguments: [ String ])
         -> Process.FancyResult
        // some error handling in the real module here
        return Process.launch(at: url.path, with: arguments)

// Apple Swift version 4.2.1 (swiftlang-1000.11.42 clang-1000.11.45.1)
// Targ...

You can’t add @dynamicCallable or @dynamicMemberLookup in an extension, it has to be defined in the basetype. The code also omits a helper extension on Process, which you can find in the GitHub repo.

When the Swift compiler sees this:


it is about to emit the cannot-call error above, but because our type is marked as @dynamicCallable, it is instead going to rewrite this into:

shell.swift.dynamicallyCall(withArguments: [ "--version" ])

And since .swift is looked up dynamically, the whole thing looks like this:

shell[dynamicMember: "swift"]
     .dynamicallyCall(withArguments: [ "--version" ])

Our implementation of dynamicallyCall gets the URL to the tool as part of the lookup process (file:///usr/bin/swift).
It then just forks this tool using Process.launch(at: url, with: args) and returns with the result (another helper object containing the tools exit status, plus the command output and error data emitted).

In short: We just turned a Unix commandline tool into a Swift function with minimal effort. Now we can do all the fancy stuff we wanted above, like:

for file in shell.ls("/Users/").split(separator: "\n") {
    print("dir:", file)
// dir: Guest
// dir: Shared
// dir: helge

As you can see @dynamicCallable and @dynamicMemberLookup combine beautifully, and doing so increases their usefulness a lot.
Imagine how your SQL library could dynamically lookup a stored procedure (or EOFetchSpecification) and run it, just by using db.processPendingOrders().

Finished Shell Swift Package

A few words of warning: This is intended as a demo. It should work just fine, but in the name of error handling and proper Swift beauty, you might want to approach forking processes differently 🤓 (BTW: PRs are welcome!)

Sample tool using the Shell package

The regular Swift Package Manager setup process:

mkdir ShellConsumerTest && cd ShellConsumerTest
swift package init --type executable

Sample main.swift, calling the host tool (located in /usr/bin):

import Shell


Sample Package.swift:

// swift-tools-version:5.0

import PackageDescription

let package = Package(
    name: "ShellConsumerTest",
    dependencies: [
        .package(url: "https://github.com/helje5/Shell.git",
                 from: "0.1.0"),
    targets: [
        .target(name: "ShellConsumerTest",
                dependencies: [ "Shell" ]), // <= do not forget!

Remember to add the dependency in two places. WET is best!

swift run and swift test patch the $PATH to just /usr/bin. You may want to run the binary directly to make lookup work properly.

Origins: Python

Congratulations! Now that you understand @dynamicCallable and @dynamicMemberLookup you essentially turned into a Python-Pro! It is roughly how Python implements functions, methods and objects in its runtime.

It is common knowledge that the features were added so that Google could integrate Python machine learning libraries neatly into Swift. But there is more to the feature, in fact Python itself has “always” had the same feature: ___getattr___ and ___call___.

Python is the first time we’ve seen the Callable concept. In Python every call like


is essentially a “property get” which returns a (potentially self-bound) “callable”, which is then being called. This is pretty different to other languages which often directly invoke a method, or pass it through a message dispatcher. (In Objective-C you could always do the same shown here using forwardInvocation: and friends.)


An obvious limitation is that both features are statically typed. You can’t lookup one function thats returns an Int, and another function which returns a String. You have to tell the compiler in advance what type you expect.
In practice this is probably going to end up in a lot of Anys / as? / is when this feature is being used. Time will show.

Another limitation is that the reverse is not possible, i.e. you cannot lookup a Callable for a Swift function and dynamically invoke it via m.dynamicallyCall(withArguments:). Aka reflection.

At least the current implementation doesn’t seem to support overloading, i.e. you can’t have this:

public struct MyCallable {
  func dynamicallyCall(withArguments arguments: [ Int ]) -> Int {
    return arguments.reduce(0, +)

  func dynamicallyCall(withArguments arguments: [ Any ]) -> String {
    return arguments.map { "\($0)" }.joined(separator: ",")
let call = MyCallable()
call.joined(1, "5", [2,3,4])

You can’t mix types, all arguments have to be the same type. I.e. this is not possible:

func dynamicallyCall(arg1: Int, arg2: String) -> String 

Or this, which is specifically annoying for APIs (though this goes away a little with async/await):

func dynamicallyCall<T>(withArguments arguments: [ Any ], yield: ( T ) -> Void)

Streaming and Async I/O

Using this library in server side code is not recommended, it is blocking and the stdout/err/in processing is not streaming. (Also: do I have to talk about the security implications of doing such stuff on the server? I hope not 😎)

An example on how to do this properly can be found in Noze.io, which provides piping, backpressure aware streams, etc:

let s = spawn("git", "log", "-100", "--pretty=format:%H|%an|<%ae>|%ad")
  | readlines
  | through2(linesToRecords)
  | through2(recordsToHTML)
  | response


Those two are pretty exciting features and we are looking forward what people are going to do with them!

The code didn’t have any cows, so let’s at least have this one: 🐄

Written on December 21, 2018