OPS RUNBOOK

When it goes silent at 2am,
this is the page you want open.

Standing up a self-hosted AI assistant is the easy part. Keeping it running is the part that decides whether you still have an assistant in a month. This is the field guide for that part — 7 real production failure modes, each with the exact fix.

Get the runbook — $12
PDF · 7 pages
Instant download · 30-day money-back · tax handled at checkout

Seven ways it breaks — diagnosed and solved

Every entry follows the same shape: symptom → diagnosis → fix → prevent. All from real boxes, not theory.

#1 · THE SILENT ONE

"Healthy but silent"

The service reads active (running) while your assistant is quietly switched off. Triage it in 30 seconds — ping first, always.

#2 · GONE DARK

The box disappears

No ping, no SSH. Is it the machine, the network, or DNS? The decision tree that saves you an hour of poking at the wrong layer.

#3 · DEAD AIR

Empty replies

"The model seems broken." When it's the provider, when it's your config, and when it's an empty wallet wearing a "rate limit" mask.

#4 · LOCKED OUT

Auth & key failures

Expired tokens, rotated keys, the 401 that looks like an outage. Where to look and what to rotate, in order.

#5 · THE BOOT RACE

Auto-paused on reboot

The boot-time race that silently pauses your chat platform after a restart — and the one ordering fix that ends it.

#6 · THE BILL

Runaway costs

A loop or a wrong model can turn pennies into a surprise. How to cap it, catch it early, and right-size what's running.

#7 · THE ONE YOU SKIPPED

The backup you never tested

A backup you've never restored from isn't a backup. How to prove yours works before the day a dead machine makes it the only thing that matters.

THE METHOD

Ping first. Read the actual symptom. Apply the known fix. Then close the door so it can't happen again.

Who it's for

Anyone running an LLM assistant on a home server, mini-PC, Pi, or VPS who wants the "when it breaks" answers before it breaks. If you can SSH in and read a log, this is written for you.

PART OF THE BIGGER BUILD

This is Module 10 of Build Your Own Self-Hosted AI Assistant — the chapter people come back to, so it stands alone. Want the whole build (stand it up → personality → files, calendar & email → scheduling → backup → security)? The Guide or the Guide + Kit are in the shop, and this runbook's price comes off if you upgrade.

Questions

What exactly do I get?

A single PDF, 7 pages, ~485 KB. Seven failure modes, each as symptom → diagnosis → fix → prevent. Download it the moment you buy.

Do I need to be a developer?

No. If you're comfortable in a terminal — ssh, reading a service log, editing a config — you'll be at home here. No AI background required.

What if it's not for me?

30-day, no-questions money-back. And tax is handled automatically at checkout — Gumroad is the merchant of record, so there's nothing for you to file.

Is this tied to one model or platform?

The failure modes are general to self-hosted LLM assistants — service health, networking, providers, keys, boot ordering, cost, backups. The fixes translate across stacks.

Get the answers before you need them.

The 2am version of you will be glad this is already on the drive.

Get the runbook — $12

PDF · 7 pages · instant download · 30-day money-back