18x faster warm installs than Composer. Byte-identical vendor/.
An extremely fast, Composer-compatible PHP installer, written in Rust. It reads your
composer.json and composer.lock unchanged and writes the same vendor/,
byte for byte, in a fraction of the time on every install that is not bound by the network.
Early, built in the open. phpm install is byte-identical to Composer
on 98.1% of a 312-project nightly sweep, and its warm installs beat Composer's on every project
measured on Linux and macOS. Windows isn't supported yet
(open issues,
help welcome). It installs from a lockfile only: no
update or require yet.
Every phase, decision and benchmark behind those numbers is public: start at
phases.
vendor/, bytes and file modes, on every pinned fixtureWarm install, Laravel skeleton
109 packages, M1 Pro, APFS, Composer 2.10.3, hyperfine, every tool measured in the same session. Lower is faster.
Real apps, not just a skeleton
Same machine, with each fixture's own scripts and plugins on (results).
| Fixture | Composer warm | phpm warm | Speed-up | Composer no-op | phpm no-op | Speed-up |
|---|---|---|---|---|---|---|
| Bedrock (WordPress) | 4.496 s | 319 ms | 14.1x | 1.204 s | 4.7 ms | 256x |
| drupal/recommended-project | 15.119 s | 1.973 s | 7.7x | 1.260 s | 5.4 ms | 233x |
Why a second tool, when Composer is good
Composer is good: a lockfile, parallel downloads, one blessed registry, a respected maintainer team. The question was never "is Composer slow" — it is "where is it slow enough that someone would install a second tool."
It never breaks a project
Every competitor either refuses plugins, emulates a few, or breaks silently. phpm installs
natively when it can prove the result is identical to Composer's, and hands the rest to real
Composer when it cannot — --explain says which. A user should never be worse off for
having tried it.
Compatibility is published, not claimed
A nightly sweep diffs phpm's vendor/ against Composer's across hundreds of real
projects and publishes the number, the way Ruff published ">99.9% Black-compatible." See
Compatibility.
It is built for where installs multiply now
Agents run installs in CI, in containers, and in parallel git worktrees, each wanting its own
vendor/. With a shared store, a new worktree's vendor/ costs a fraction
of a second and almost no disk.
Published nightly, not claimed
A nightly sweep installs phpm and Composer side by side across 330 pinned
open-source PHP projects and diffs vendor/ byte for byte, modes included —
live results and methodology.
| Mode | Identical | Ratio | Gate |
|---|---|---|---|
| Plugins and scripts off | 306 / 312 | 98.1% | 95% — passed |
| Plugins and scripts on (Composer fallback included) | 267 / 281 | 95.0% | — |
The sweep is also how real bugs get found before users hit them: it caught a case where, on filesystems without copy-on-write cloning, a plugin writing into a hard-linked file could corrupt the shared package store for every later install. phpm now copies instead of hard-linking whenever scripts or plugins will run.
vendor/ is also byte-identical, bytes and file modes, on every one of the project's
own pinned fixtures (Laravel, Symfony demo, Monica, Bedrock, drupal/recommended-project, and two
production apps), including the optimised class map.
Install
Prebuilt binaries for macOS and Linux are on every GitHub release. Windows isn't supported yet; the gaps are filed as issues.
Then run it against a real project:
--explain prints, per package, whether it was installed natively or handed to Composer
and why. Plain phpm install otherwise behaves like composer install:
--no-dev, --no-scripts, --no-plugins,
--optimize-autoloader / -o, --classmap-authoritative / -a
all work.