Your server crashed at 3 AM. You're asleep.

CoalHost detects, diagnoses, fixes and reboots while you sleep, then emails you what happened.

  1. Step 1

    Detect crash within seconds

    Agent on the host watches the JVM. The panel knows the moment a server dies — usually before the player rage-quits the lobby.

  2. Step 2

    AI analyses the crash report

    Haiku reads the stack trace, identifies the offending mod or config, and matches it against the cross-customer crash-pattern database.

  3. Step 3

    Apply the fix automatically

    Recipe DB applies the known-good remediation: disable broken recipe, lower a tick rate, swap a config flag, or roll back a bad mod update.

  4. Step 4

    Server back online in 60s

    Restart is gated on a smoke-test. If the fix doesn't hold, we email you with the root cause instead of looping forever.

Apex, Bisect and Shockbyte don't do this. They ticket you, wait for a tech, and bill you anyway.

Learn how it works
The four layers

Predict. Diagnose. Tune. Learn.

Most hosting providers stop at "we restarted it." CoalHost treats reliability as a four-stage engineering problem and runs an AI agent at every stage.

Predict

OOM-predictor — 30 min advance warning

Linear regression over heap-after-GC samples flags the trajectory toward OutOfMemoryError before the JVM panics.

Most providers find out about an OOM the same way you do: when players ping you in Discord. We sample heap usage every 30 seconds, fit a line, and the moment the projected ceiling crosses the cap inside the next 30 minutes we surface a banner in the panel and email the owner. Often the fix is a RAM bump or a leaking-mod swap — either way you make the call before the world is offline.

Diagnose

Memory-leak detector with heap-dump AI analysis

Heap retention by class is sampled hourly; if a mod's footprint grows monotonically over four hours, we triage it and tell you which mod is leaking.

On suspect growth the agent triggers a heap dump and feeds the dominant-tree class names to the AI. We've already correlated common leak signatures (entity caches, projectile holders, tile-entity collections) with the mods that ship them. You get a one-line root cause: "Mod XYZ leaks projectiles, version 2.0.3 has a fix" — not a 300 MB hprof and shrug.

Tune

Spark profiler + AI tuning suggestions

Spark runs on a schedule. The AI reads the report and says "Mod X consumes 30% of tick time — disable, or update, or live with it."

Spark is the standard, but the standard output is a 50-page profile most server owners don't read. We auto-install Spark, run profiled samples once per week, and ship the AI's plain-English summary into the panel. Common output: "AmbientSounds spends 18% of tick — set ambient_sound_volume=0 in the config." That's what tuning a modpack actually looks like.

Learn

Cross-customer crash-pattern matching

One customer's fix becomes every customer's fix. Recipe DB stores fingerprinted crash → remediation pairs and applies them on install.

Every crash gets fingerprinted (stack-trace head + mod IDs + Forge version). If a fingerprint matches a known pattern, we apply the recipe automatically. If it doesn't, we triage it manually once — and from that moment forward every customer with the same fingerprint gets the fix on install. After one month of running this on ZOMBIEMANIA we'd covered 12 modpacks. The network effect compounds.

How it works

From crash to fixed in four steps

  1. 01

    Crash detected

    Agent on the host watches the JVM and ships logs the moment a server exits non-zero.

  2. 02

    Fingerprint built

    Stack-trace head + mod IDs + Forge version → stable hash that survives line-number churn.

  3. 03

    AI matches & decides

    Recipe DB lookup first; on miss, Haiku reads the report and proposes a remediation.

  4. 04

    Fix applied, server back online

    Recipe applied (disable mod, tweak config, roll back update); smoke-test gates the restart.

The numbers

Real metrics from real packs

30 min

advance OOM warning before the JVM panics

100%

of recognised crashes have an automatic fix on install

12 packs

covered by recipe DB after the first month

<1 ms

ZGC pause vs G1's 50–300 ms — no lag-spike kills

vs the rest

The reliability gap, in one table

CoalHost is the only hosting provider in this segment that ships every layer. The others compete on price tier, panel polish or first-month coupons.

FeatureCoalHostApexBisectShockbyte
AI crash analysis
OOM prediction (30 min advance)
Memory-leak detection
ZGC sub-1ms GC pause by default
Recipe-DB cross-customer fixes
Spark profiler + AI tuning
One-click modpack install
Backups

Information about competitors gathered from their public pricing, documentation and feature pages as of 2026. Comparisons are made in good faith. Want a feature added? Email [email protected].

Try CoalHost from €2/GB

Modpacks that fix themselves. EU datacenter today, NA (Ashburn) coming Q4. Sign up free, pay only when you add a server.