Crucible v0.13.14 on npm
Bare metal server monitoring.
Catch hardware failures before they cascade. SMART, IPMI, ZFS, and ECC monitoring with honest fix commands. $3 per node, first 3 free.
Runs as glassmkr user, never root. MIT licensed. Source: github.com/glassmkr/crucible
eno2 has logged 160 hardware errors (CRC + frame) and 90 discards: the error ratio crossed 0.1% of packets, usually a failing cable or SFP.
ethtool -S eno2 | grep -iE 'err|crc'interface_errorsFOUNDING PRINCIPLES
Three principles.
Opinionated thresholds, not “configure your own monitoring.”
You shouldn’t have to figure out what to monitor. We’ve watched bare-metal servers fail for ten years. The 62 alert rules that ship with Glassmkr fire when something is genuinely wrong, not when CPU briefly spikes. Tune per-server thresholds or disable rules that don’t apply to your fleet. You don’t have to.
An agent you can read.
Crucible is MIT licensed and on npm. Read every line of code that runs on your servers. No proprietary blob, no telemetry you can’t inspect. Audit the agent before you install it. We’d be suspicious of any monitoring tool that asks you not to.
AI where it helps, not as the headline.
Furnace is our AI assistant. It reads the alert, looks at the evidence, suggests a fix. It doesn’t autocomplete shell commands at you. It hedges interpretive claims. It says “I don’t know” when it doesn’t know. The AI is a value-add for the alert remediation workflow, not the product.
Connect a server in sixty seconds.
One command on systemd, or Docker Compose. The agent auto-detects your hardware.
$ curl -sf https://glassmkr.com/install.sh | sudo bash✓ Connected to app.glassmkr.com (eu-ams-1) ✓ Detected: 2 disks (Samsung MZ7L3480), IPMI v2.0, ECC RAM ✓ 62 alert rules active ✓ Streaming telemetry. View at app.glassmkr.com/h/abc123
What this does: downloads the Crucible binary, verifies the signature against our published key, installs as a systemd service, and registers the node with your account. Source: bash script at install.sh, ~150 lines.
Or read the install docs →Detection that matters.
62 alert rules tuned for bare-metal infrastructure. Disk SMART degradation, RAID failures, NVMe wear bands, IPMI sensor critical, ECC errors, ZFS pool health, network interface saturation, ZFS scrub errors, conntrack exhaustion. Not generic “host is unreachable” pings. Real failure modes detected before they cascade.
- 1.1 Disk SMART, NVMe wear bands, RAID + ZFS pool state
- 1.2 IPMI sensors, ECC errors, BMC accessibility
- 1.3 Network interface saturation, conntrack exhaustion
- 1.4 ZFS scrub errors, filesystem capacity, FD exhaustion
Storage
SMART Health
Notifications routed where you live.
Telegram, Slack, Discord, PagerDuty, email, generic webhook. Group alerts by server, by severity, by team. Suppress noise during planned maintenance windows.
The alerting layer is unopinionated; route alerts wherever your team already pays attention. Configure once, then forget about it.
- 2.1 Telegram, Slack, Discord, PagerDuty, email, webhook
- 2.2 Group by server, by severity, by team
- 2.3 Suppress noise during planned maintenance windows
- 2.4 Unopinionated routing: bring your own workflow
AI-assisted remediation, hedged honestly.
Every alert page shows what we recommend: the command to run, what to verify after. Furnace (our AI assistant) annotates with context-specific notes: “On Ubuntu 22.04, the binary path is X; on RHEL 8 it’s Y.”
RAID md126 degraded
md126 (raid1) is degraded. Failed disks: sda1. One more failure means data loss.
Sustained just now. (raid_degraded)
Replace the failed drive immediately. Check cat /proc/mdstat and mdadm --detail /dev/md126 for status.
→ Storage overview
# Identify the degraded array and the failed member (alert evidence has the device names)
cat /proc/mdstat
sudo mdadm --detail /dev/mdX # replace mdX with the array from the alert
lsblk
# After replacing the failed disk, re-add the new member
sudo mdadm --manage /dev/mdX --add /dev/sdYN # sdYN = partition on the new disk
# Watch the rebuild progress
watch -n 5 cat /proc/mdstatPricing without surprises.
$3 per node per month. First three nodes free. No agent-tier upcharges. No “log volume” billing. No 90th-percentile pricing tricks. You pay $3 per node, billed monthly, cancellable any time.
- 4.1 $3 per node per month after 3 free nodes
- 4.2 No telemetry volume billing, no log ingestion charges
- 4.3 No tier upcharges; every Pro account gets every feature
- 4.4 Cancel any time, no exit fee
YOUR TELEMETRY STAYS YOURS
Your alert text never reaches OpenAI, Anthropic, or Google.
The Crucible agent is MIT licensed and open source. Only metrics and alert state leave your server: no logs, no command output, no secrets. Dashboard runs in Amsterdam on dedicated hardware, operated by an EU team under GDPR. Read the full trust posture →
Pricing
$3 per node per month. First 3 nodes free. No credit card. Cancel anytime.
Honest pricing. $3 per node per month. First three nodes are free. Cancel any time. No card required to start.
WHAT YOU GET ON EVERY PLAN
- No per-feature paywalls on alert rules. All 62 rules in every plan, free or Pro.
- The full read+write API, all six notification channels, and predictive trend warnings: all free.
- Built in the open. MIT-licensed Crucible agent on npm.
- Opinionated thresholds from a decade of bare-metal operations.
- Self-host Crucible forever for free. Pay for the interpretation.
PRO INCLUDES
Unlimited AI analysis.
Self-hosted Gemma 4 narrates your alerts and suggests root causes. No third-party APIs. Free gets one trial analysis per server; Pro is unlimited.
90-day history.
Free tier retains 7 days. Pro retains 90.
More than 3 nodes.
The first three nodes are free forever. Past three, it is a flat $3 per node per month, with no other gate on node count. The same 62 rules, the full API, every notification channel, and predictive trend warnings apply whether you run one node or four hundred.
Every cloud, VM, and serverless workload runs on bare metal underneath. We build the opinionated early-warning system that keeps that base running, so the layers above can focus on the app.
ABOUT GLASSMKR’S INFRASTRUCTURE
Glassmkr runs on dedicated bare-metal infrastructure, not shared cloud VMs.
The Crucible agent runs on every Glassmkr server. The AI assistant Furnace runs on a self-hosted Gemma 4 26B model on a single NVIDIA L4 GPU in Amsterdam. No OpenAI, no Anthropic, no Google. Your alert data never leaves European jurisdiction.
SvelteKit for the dashboard, llama.cpp running Gemma for the AI, pgBackRest for backups. Boring, well-understood, debuggable.
More about how Glassmkr is built →Set up your dashboard in sixty seconds.
Start free for 3 nodesNo credit card. No email blasts. Cancel anytime.
3 nodes free. Install in under 2 minutes. Cancel any time with no friction.