Important: This documentation covers modern versions of Yarn.
For 1.x docs, see
ConstraintsOffline CachePlug'n'PlayPluginsProtocolsRelease WorkflowWorkspacesZero-Installs


Edit this page on GitHub


Are you a library author trying to make your library compatible with the Plug'n'Play installation strategy? Do you want to use the PnP API for something awesome? If the answer to any of these questions is yes, make sure to visit the PnP API page after reading the introduction!

Unveiled in September 2018, Plug'n'Play is an innovative installation strategy for Node. Based on prior work in other languages (for example autoload for PHP), it presents interesting characteristics that build upon the regular CommonJS require workflow in an almost completely backward-compatible way.

The node_modules problem

The way installs used to work was simple: when running yarn install Yarn would generate a node_modules directory that Node was then able to consume thanks to its built-in Node Resolution Algorithm. In this context, Node didn't have to know the first thing about what a "package" was: it only reasoned in terms of files. "Does this file exist here? No: Ok, let's look in the parent node_modules then. Does it exist here? Still no: Ok ...", and it kept going until it found the right one. This process was vastly inefficient for several reasons:

  • The node_modules directories typically contained gargantuan amounts of files. Generating them could make up for more than 70% of the time needed to run yarn install. Even having preexisting installations wouldn't save you, as package managers still had to diff the contents of node_modules with what it should contain.

  • Because the node_modules generation was an I/O-heavy operation, package managers didn't have much leeway to optimize it beyond just doing a simple file copy - and even though it could have used hardlinks or copy-on-write when possible, it would still have needed to diff the current state of the filesystem before making a bunch of syscalls to manipulate the disk.

  • Because Node had no concept of packages, it also didn't know whether a file was meant to be accessed. It was entirely possible that the code you wrote worked one day in development but broke later in production because you forgot to list one of your dependencies in your package.json.

  • Even at runtime, the Node resolution had to make a bunch of stat and readdir calls to figure out where to load every single required file from. It was extremely wasteful and was part of why booting Node applications took so much time.

  • Finally, the very design of the node_modules folder was impractical in that it didn't allow package managers to properly de-duplicate packages. Even though some algorithms could be employed to optimize the tree layout (hoisting), we still ended up unable to optimize some particular patterns - causing not only the disk usage to be higher than needed, but also some packages to be instantiated multiple times in memory.

Fixing node_modules

Yarn already knows everything there is to know about your dependency tree - it even installs it on the disk for you. So, why is it up to Node to find where your packages are? Instead, it should be the package manager's job to inform the interpreter about the location of the packages on the disk and manage any dependencies between packages and even versions of packages. This is why Plug'n'Play was created.

In this install mode (the default starting from Yarn 2.0), Yarn generates a single .pnp.cjs file instead of the usual node_modules folder containing copies of various packages. The .pnp.cjs file contains various maps: one linking package names and versions to their location on the disk and another one linking package names and versions to their list of dependencies. With these lookup tables, Yarn can instantly tell Node where to find any package it needs to access, as long as they are part of the dependency tree, and as long as this file is loaded within your environment (more on that in the next section).

This approach has various benefits:

  • Installs are now nearly instantaneous. Yarn only needs to generate a single text file (instead of potentially tens of thousands). The main bottleneck becomes the number of dependencies in a project rather than disk performance.

  • Installs are more stable and reliable due to reduced I/O operations. Especially on Windows (where writing and removing files in batches may trigger various unintended interactions with Windows Defender and similar tools), I/O heavy node_modules operations were more prone to failure.

  • Perfect optimization of the dependency tree (aka perfect hoisting) and predictable package instantiations.

  • The generated .pnp.cjs file can be committed to your repository as part of the Zero-Installs effort, removing the need to run yarn install in the first place.

  • Faster application startup! The Node resolution doesn't have to iterate over the filesystem hierarchy nearly as much as before (and soon won't have to do it at all!).

Initializing PnP

Yarn generates a single .pnp.cjs file that needs to be installed for Node to know where to find the relevant packages. This registration is generally transparent: any direct or indirect node command executed through one of your scripts entries will automatically register the .pnp.cjs file as a runtime dependency. For the vast majority of use cases, the following will work just as you would expect:

  "scripts": {
    "start": "node ./server.js",
    "test": "jest"

For some remaining edge cases, a small setup may be required:

  • If you need to run an arbitrary Node script, use yarn node as the interpreter, instead of node. This will be enough to register the .pnp.cjs file as a runtime dependency.
yarn node ./server.js
  • If you operate on a system that automatically executes a Node script (for instance on Google Cloud Platform), simply require the PnP file at the top of your init script and call its setup function.

As a quick tip, all yarn node typically does is set the NODE_OPTIONS environment variable to use the --require option from Node, associated with the path of the .pnp.cjs file. You can easily apply this operation yourself if you prefer:

node -r ./.pnp.cjs ./server.js
NODE_OPTIONS="--require $(pwd)/.pnp.cjs" node ./server.js

PnP loose mode

Because the hoisting heuristics aren't standardized and predictable, PnP operating under strict mode will prevent packages from requiring dependencies that are not explicitly listed; even if other dependencies also depend on it. This may cause issues with some packages.

To address this problem, Yarn ships with a "loose" mode which will cause the PnP linker to work in tandem with the node-modules hoister - we will first generate the list of packages that would have been hoisted to the top level in a typical node_modules install, then remember this list as what we call the "fallback pool".

Note that because the loose mode directly calls the node-modules hoister, it follows the exact same implementation as the true algorithm used by the node-modules linker!

At runtime, packages that require unlisted dependencies will still be allowed to access them if any version of the dependency ended up in the fallback pool (which packages exactly are allowed to rely on the fallback pool can be tweaked with pnpFallbackMode).

Note that the content of the fallback pool is undetermined. If a dependency tree contains multiple versions of the same package, there is no means to determine which one will be hoisted to the top-level. Therefore, a package accessing the fallback pool will still generate a warning (via the process.emitWarning API).

This mode provides a compromise between the strict PnP linker and the node_modules linker.

In order to enable loose mode, make sure that the nodeLinker option is set to pnp (the default) and add the following into your local .yarnrc.yml file:

pnpMode: loose

More information about the pnpMode option.


Because we emit warnings (instead of throwing errors) on resolution errors, applications can't catch them. This means that the common pattern of trying to require an optional peer dependency inside a try/catch block will print a warning at runtime if the dependency is missing, even though it shouldn't. The only runtime implication is that such a warning can cause confusion, but it can safely be ignored.

For this reason, PnP loose mode won't be the default starting with version 2.1 (as we originally planned). It will continue to be supported as an alternative, hopefully easing the transition to the default and recommended workflow: PnP strict mode.


In the years leading up to Plug'n'Play being ratified as the main install strategy, other projects came up with alternative implementations of the Node Resolution Algorithm - usually to circumvent shortcomings of the require.resolve API. Examples include Webpack (enhanced-resolve), Babel (resolve), Jest (jest-resolve), and Metro (metro-resolver). These alternatives should be considered as superseded by proper integration with Plug'n'Play.

Compatibility Table

The following compatibility table gives you an idea of the integration status with various tools from the community. Note that only CLI tools are listed there, as frontend libraries (such as react, vue, lodash, ...) don't reimplement the Node resolution and as such don't need any special logic to take advantage of Plug'n'Play:

Suggest an addition to this table

Native support

Many common frontend tools now support Plug'n'Play natively!

Project name
AngularStarting from 13+
BabelStarting from resolve 1.9
Create-React-AppStarting from 2.0+
DocusaurusStarting from 2.0.0-beta.14
ESLintSome compatibility issues w/ shared configs (fixable using @rushstack/eslint-patch)
GatsbySupported with version ≥2.15.0, ≥3.7.0
GulpSupported with version 4.0+
HuskyStarting from 4.0.0-1+
JestStarting from 24.1+
Next.jsStarting from 9.1.2+
ParcelStarting from 2.0.0-nightly.212+
Preact CLIStarting from 3.1.0+
PrettierStarting from 1.17+
🐊PutoutStarting from 22.8
RollupStarting from resolve 1.9+
StorybookStarting from 6.0+
TypeScriptVia plugin-compat (enabled by default)
TypeScript-ESLintStarting from 2.12+
VSCode-StylelintStarting from 1.1+
WebStormStarting from 2019.3+; See Editor SDKs
WebpackStarting from 5+ (plugin available for 4.x)

Support via plugins

Project name
ESBuildVia @yarnpkg/esbuild-plugin-pnp
VSCode-ESLintFollow Editor SDKs
VSCodeFollow Editor SDKs
Webpack 4.xVia pnp-webpack-plugin (native starting from 5+)


The following tools cannot be used with pure Plug'n'Play install (even under loose mode).

Important: Even if a tool is incompatible with Plug'n'Play, you can still enable the node-modules plugin. Just follow the instructions and you'll be ready to go in a minute 🙂

Project name
FlowFollow yarnpkg/berry#634
React NativeFollow react-native-community/cli#27
PulumiFollow pulumi/pulumi#3586
VSCode Extension Manager (vsce)Use the vsce-yarn-patch fork with the node-modules plugin enabled. The fork is required until microsoft/vscode-vsce#493 is merged, as vsce currently uses the removed yarn list command
HugoHugo pipes expect a node-modules dir. Enable the node-modules plugin
ReScriptFollow rescript-lang/rescript-compiler#3276

This list is kept up-to-date based on the latest release we've published starting from v2. In case you notice something off in your own project please try to upgrade Yarn and the problematic package first, then feel free to file an issue. And maybe a PR? 😊

Frequently Asked Questions

Why not use import maps?

Yarn Plug'n'Play provides semantic errors (explaining you the exact reason why a package isn't reachable from another) and a sensible JS API to solve various shortcomings with require.resolve. These are features that import maps wouldn't solve by themselves. This is answered in more detail in this thread.

A main reason we're in this mess today is that the original node_modules design tried to abstract packages away in order to provide a generic system that would work without any notion of packages. This became a challenge that prompted many implementers to come up with their own interpretations. Import maps suffer from the same flaw.

Packages are stored inside Zip archives: How can I access their files?

When using PnP, packages are stored and accessed directly inside the Zip archives from the cache. The PnP runtime (.pnp.cjs) automatically patches Node's fs module to add support for accessing files inside Zip archives. This way, you don't have to do anything special:

const {readFileSync} = require(`fs`);

// Looks similar to `/path/to/.yarn/cache/`
const lodashCeilPath = require.resolve(`lodash/ceil`);


Fallback Mode

Back when PnP was first implemented, the compatibility wasn't as good as it is now. To help with the transition, we designed a fallback mechanism: if a package tries to access an unlisted dependency, it's still allowed to resolve it if the top-level package lists it as a dependency. We allow this because there's no resolution ambiguity, as there's a single top-level package in any project. Unfortunately, this may cause confusing behaviors depending on how your project is set up. When that happens, PnP is always right, and the only reason it works when not in a workspace is due to some extra lax.

This behavior was just a patch, and will eventually be removed to clear up any confusion. You can prepare for that now by setting pnpFallbackMode to none, which will disable the fallback mechanism altogether.