Supervise any process in any language — Node.js, Python, Go, Rust, Ruby, PHP, or anything else that runs on Linux. Proxy HTTP/WS/TCP/UDP, rate-limit, and monitor everything from a single static binary.
Linux · Any language · Zero config to start
No config files required to get started. Everything you need is built in.
localhost:6789 with real-time CPU, memory, connection stats, and per-instance LED event indicators.TCP_USER_TIMEOUT, no client support needed, slow links never cut.setPing), round-trip latency kept per device (pingStats), pings only to silent connections.apm healconf checks a config and everything it includes, explains each problem and offers the fix: a missing ;, a key brought from PM2, a boolean the daemon reads as false, a setting a new version reads differently. Runs as you, backs up before it edits, and converts PM2 ecosystem files.apm_network answers request() from any of them. No connector changes.memory_limit via cgroups v2 (OOM-kill, then your restart policy), pull or push health checks, connection drain, and an on_crash webhook with HMAC-signed payload and the last log lines.start_concurrency caps the width; depends_on is honoured whatever order the file declares.apm reload works out what actually changed and applies only that — log paths, watcher, Vanguard rules, listeners — without bouncing children. Only changes the process snapshots at start (exec, env, user, path) restart it.Small binary, small blast radius, small surprises. APM is the process manager most teams reach for when PM2 stops fitting the shape of their stack.
APM holds the listening socket and cycles instances behind it one at a time. A deploy finishes without a single user seeing a 502. No gateway in front, no blue/green orchestration — just restart. This is the piece teams most often wire up by hand on top of PM2 cluster mode; in APM it's how restarts work by default, for any language, not just Node.
The supervisor itself is a single statically-linked Go binary — 2.6 MB, no runtime, no npm tree. Goroutines and channels keep scheduling off your app's CPU, so a busy daemon doesn't steal cycles from your workers. Contrast this with Node-based managers where the supervisor competes with your app for the same event loop.
Workers in any language talk to each other through the daemon: fire-and-forget channels, request/reply with first-reply-wins, and persistent bidirectional streams. A Node worker can hand a live console stream to a Python job; a PHP request can query a Go service and get a reply in milliseconds. You don't run Redis, you don't run RabbitMQ, you don't run a sidecar. The broker is the process manager you already have.
One native inotify tree inside the daemon, per-worker glob include/exclude, debounced restarts. The "watcher eats my CPU" failure mode that plagues chokidar-based watchers — including PM2's --watch on large repos — does not happen here, because there is no chokidar, no polling, no per-worker watcher process.
If it runs on Linux and writes to stdout, APM can supervise it. No adapter. No "PM2 interpreter" wrapper script. Node, Python, Go, Rust, PHP-FPM workers, Bash pipelines, a compiled C++ daemon — all first-class citizens, all supervised identically, all with the same crash recovery, logs, metrics, and IPC.
The binary goes to /usr/sbin/apm. Your config lives in /etc/apm/apm.conf. That's it — no ~/.pm2 directory full of JSON state that can corrupt, no dump.pm2 to port between hosts. Migrate a server by copying two files.
ps -eo pid,rss,command.
pm2 listPM2 is a great Node-first process manager. APM is what you reach for when your stack stops being Node-only, when you're tired of wiring nginx in front for rate limiting, or when you want cross-language IPC without running Redis. Here's the honest, feature-by-feature picture:
| APM v2.12.0 | PM2 v5.x | |
|---|---|---|
| Runtime | Static Go binary — 2.6 MB, zero deps | Node.js + npm package — 26 MB, 5,000+ files |
| Supervised languages | Any executable — Node, Python, Go, Rust, PHP, Ruby, C/C++, Bash, Deno, Bun… | Node-first; other languages need a wrapper interpreter script |
| Daemon footprint (idle) | ~16 MB RSS supervising 5 workers | ~86 MB RSS supervising 2 workers |
| Reverse proxy (HTTP/WS/TCP/UDP) | Built in — round-robin, session migration on restart | External — requires nginx or HAProxy |
| Rate limiting & firewall | Built in (Vanguard): token bucket, IP banlists, CDN trust, path bans | External — requires nginx, fail2ban, Cloudflare, or middleware |
| TLS termination | Built in — HTTPS/WSS, mTLS, bring your own certs | External — requires nginx or app-level |
| Cross-language IPC | Channels + persistent streams, routed by daemon, identical API in Node/Python/PHP/Perl/Lua | pm2.launchBus — Node-only, axon-based |
| File watcher | Native inotify, per-worker glob patterns, debounced, one tree | chokidar — historically CPU-heavy on large repos (issue history) |
| Zero-downtime rolling restart | Default for any worker type with a shared listener | Node cluster mode only (pm2 reload) |
| GUI / dashboard | Local live GUI shipped in-binary at :6789 |
PM2+ SaaS (paid) or third-party dashboards |
| Config format | apm.conf (nginx-style) or CLI flags |
ecosystem.config.js or CLI flags |
| Persistent state | One config file. No hidden dump file. | ~/.pm2/dump.pm2 + pm2 save / resurrect |
| Install | curl … | bash — one file to /usr/sbin/apm |
npm i -g pm2 — requires Node.js + npm already installed |
| License | Free to use (binary); source proprietary | AGPL-3.0 (core); PM2+ is a paid SaaS |
For supervising Node apps in fork mode — effectively yes. You'll re-express your ecosystem.config.js as an apm.conf block (5–10 lines per worker), install the one binary, and run. For PM2 cluster mode specifically, APM's -server flag is the equivalent: APM holds the listener and round-robins connections to instances in any language, not only Node. Features you will not find in APM: PM2+ SaaS, pre-packaged ecosystem modules (pm2-logrotate, etc.) — APM ships rotation and metrics as core, so most of those are simply obviated.
Run apm convert ecosystem.config.js out.conf. APM evaluates the JavaScript ecosystem (using your local node — same one PM2 uses), or reads JSON directly (ecosystem.json), and emits one APM worker { } block per apps[] entry. Fields that have direct equivalents — name, script, cwd, instances, watch, env, max_memory_restart, log_date_format, kill_timeout, and the rest — convert silently. Anything that doesn't map cleanly (PM2 cluster mode, cron_restart, instances: 'max', multi-profile env_production / env_staging) surfaces as an inline # WARN: comment above the block it applies to, so you can review and decide. The converter is intentionally a starting point, not a behaviour-bit-for-bit clone — review the output, then cat it into your main apm.conf and apm reload.
APM doesn't care what stack you use. If it compiles, interprets, or just runs on Linux — APM can start it, supervise it, proxy it, and restart it when it crashes. No plugins. No adapters. No wrapper scripts.
If your app runs on Linux and talks to stdout — APM has you covered.
One line. Detects your arch, downloads the right binary, asks for sudo only when it needs to copy itself to /usr/sbin/apm.
$ curl -fsSL https://processmanager.dev/install.sh | bashwget alternative
$ wget -qO- https://processmanager.dev/install.sh | bashPrefer to inspect first? Download manually →
Already running APM? sudo apm upgrade installs the new version and then checks your config with apm healconf. Since v2.8 a root-owned path needs a user and the CLI needs root or group apm. Since v2.11 the web GUI always asks for a password — with no gui_password set, a generated one is printed in the daemon log and by apm gui — and a listener behind a proxy can name its trusted_proxies so forwarded client IPs are only believed from them.
# Simple start
$ apm run node server.js
# With proxy + multi-instance
$ apm run node server.js \
-instances 4 \
-server http://0.0.0.0:3000# /etc/apm/apm.conf
worker {
name myapp;
exec node;
params server.js;
instances 4;
restart true;
watch *.js;
server http://0.0.0.0:3000;
}$ apm load /etc/apm/apm.conf
$ apm list
$ apm restart myapp
$ apm reload /etc/apm/apm.conf
$ apm monitor
$ apm gui
$ apm stop myappWacatac.B!ml. This is a known false positive: its ML heuristic misfires on statically compiled Go binaries and UPX-compressed executables. It has been reported to Microsoft for review. All other major antivirus engines report these files as clean — see the VirusTotal links below for full scan results.
