node package manager



npm version Status Windows build status Join the chat at Twitter Follow

Fast, disk space efficient package manager

pnpm is:

  1. Fast. Faster than npm and Yarn. This is not a fair battle as we organize dependencies in a completely different way. Nevertheless, dependencies installed with pnpm are Node.js-compatible!
  2. Efficient. One version of a package is saved only ever once on a disk.
  3. Simple and human-deterministic. Flattening is hard. A package manager can be deterministic but no human can easily predict how will a flattened dependency tree look like. Pnpm not only has a lockfile, the files in its node_modules folder are organized always the same way, the way they are described in package.json files.

Like this project? Let people know with a tweet.

Table of Contents


pnpm uses hard links and symlinks to save one version of a module only ever once on a disk. When using npm or Yarn for example, if you have 100 projects using the same version of lodash, you will have 100 copies of lodash on disk. With pnpm, lodash will be saved in a single place on the disk and a hard link will put it into the node_modules where it should be installed.

As a result, you save gigabytes of space on your disk and you have a lot faster installations! If you'd like more details about the unique node_modules structure that pnpm creates and why it works fine with the Node.js ecosystem, read this small article: Why should we use pnpm?


Install it via npm.

npm install -g pnpm

Do you wanna use pnpm on CI servers? See: Continuous Integration.


pnpm install lodash

Use pnpm in place of npm. It overrides install, update, uninstall, link, prune, list and install-test. Also, it has some custom commands, like dislink for unlinking a package (like yarn unlink but pnpm reinstalls the dependency after removing the external link). The rest of the commands pass through to npm.

For using the programmatic API, use pnpm's engine: supi.


pnpm uses npm's programmatic API to read configs. Hence, you should set configs for pnpm the same way you would for npm.

Furthermore, pnpm uses the same configs that npm uses for doing installations. If you have a private registry and npm is configured to work with it, pnpm should be able to authorize requests as well, with no additional configuration.

However, pnpm has some unique configs as well:


  • Default: ~/.pnpm-store
  • Type: path

The location where all the packages are saved on the disk.

Currently if installations are done not on the same disk on which the store is, packages are copied, not linked, from the global store (see issue #712). You can use this config to set a store location that is on the same disk where installation is done.


  • Default: false
  • Type: Boolean

If true, pnpm will use only packages already available in the store. If a package won't be found locally, the installation will fail.


  • Default: 16
  • Type: Number

Controls the maximum number of HTTP requests that can be done simultaneously.


  • Default: 5
  • Type: Number

Controls the number of child processes run parallelly to build node modules.


  • Default: true
  • Type: Boolean

Dangerous! If false, the store is not locked. It means that several installations using the same store can run simultaneously.

Can be passed in via a CLI option. --no-lock to set it to false. E.g.: pnpm install --no-lock.

If you experience issues similar to the ones described in #594, use this option to disable locking. In the meanwhile, we'll try to find a solution that will make locking work for everyone.


  • Default: false
  • Type: Boolean

If true, symlinks leaf dependencies directly from the global store. Leaf dependencies are packages that have no dependencies of their own. Setting this config to true might break some packages that rely on location but gives an average of 8% installation speed improvement.


  • Default: true
  • Type: Boolean

If false, doesn't check whether packages in the store were mutated.


pnpm is faster than npm and Yarn. See this benchmark which compares the three package managers on different types of applications.

time npm i babel-preset-es2015 browserify chalk debug minimist mkdirp
    66.15 real        15.60 user         3.54 sys
time pnpm i babel-preset-es2015 browserify chalk debug minimist mkdirp
    11.04 real         6.85 user         2.85 sys


  1. npm-shrinkwrap.json and package-lock.json are ignored. Unlike pnpm, npm can install the same name@version multiple times and with different sets of dependencies. npm's shrinkwrap file is designed to reflect the node_modules layout created by npm. pnpm cannot create a similar layout, so it cannot respect npm's lockfile format.
  2. You can't publish npm modules with bundleDependencies managed by pnpm.
  3. Binstubs (files in node_modules/.bin) are always shell files not symlinks to JS files. The shell files are created to help pluggable CLI apps in finding their plugins in the unusual node_modules structure. This is very rarely an issue and if you expect the file to be a js file, just reference the original file instead, as described in #736.
  4. Node.js doesn't work with the --preserve-symlinks flag when executed in a project that uses pnpm.
  5. The package store should be on the same disk on which installations are done, otherwise packages will be copied, not linked. So if you install on disk D:, you have to have the store on disk D:. Currently pnpm cannot automatically detect whether the disk is the same or different (see #712). To avoid copying in cases like that, specify the store location manually via the store config.

Got an idea for workarounds for these issues? Share them.

Other Node.js package managers

  • npm. The oldest and most widely used. See pnpm vs npm.
  • ied. Built on a very similar premise as pnpm. pnpm takes huge inspiration from it.
  • Yarn. The first Node.js package manager that invented lockfiles and offline installations.

Frequently Asked Questions

Why does my node_modules folder use disk space if packages are stored in a global store?

pnpm creates hard links from the global store to project's node_modules folders. Hard links point to the same place on the disk where the original files are. So, for example, if you have foo in your project as a dependency and it occupies 1MB of space, then it will look like it occupies 1MB of space in the project's node_modules folder and the same amount of space in the global store. However, that 1MB is the same space on the disk addressed from two different locations. So in total foo occupies 1MB, not 2MB.

For more on this subject:

Does it work on Windows? It is harder to create symlinks on Windows

Using symlinks on Windows is problematic indeed. That is why pnpm uses junctions instead of symlinks on Windows OS.

Does it work on Windows? Nested node_modules approach is basically incompatible with Windows

Early versions of npm had issues because of nesting all node_modules (see Node's nested node_modules approach is basically incompatible with Windows). However, pnpm does not create deep folders, it stores all packages flatly and uses symlinks to create the dependency tree structure.

What about circular symlinks?

Although pnpm uses symlinks to put dependencies into node_modules folders, circular symlinks are avoided because parent packages are placed into the same node_modules folder in which their dependencies are. So foo's dependencies are not in foo/node_modules but foo is in node_modules/foo, together with its own dependencies.

Why have hard links at all? Why not symlink directly to the global store?

One package can have different sets of dependencies on one machine.

In project A foo@1.0.0 can have dependency resolved to bar@1.0.0 but in project B the same dependency of foo might resolve to bar@1.1.0. So pnpm hard links foo@1.0.0 to every project where it is used, in order to create different sets of dependencies for it.

Direct symlinking to the global store would work with Node's --preserve-symlinks flag. But --preserve-symlinks comes with a bunch of different issues, so we decided to stick with hard links. For more details about why this decision was made, see: