Skip to main content

Install with an AI agent

Crumb ships an agent-runnable install runbook in the repository at docs/AI-INSTALL.md. If you're using a tool like Claude Code, point it at a clone of the repository on the host where Crumb will run and ask it to follow that file. It is not a replacement for the manual path, it's the same steps, written so an agent can execute them safely with a Verify check after each one.

Why a separate runbook​

The manual path in Install with Docker Compose is written for a person reading along. The agent runbook encodes the same steps as explicit, checkable instructions: generate secrets with the provided script rather than inventing them, confirm a health check before moving to the next step, and stop and report rather than guessing forward when something fails.

Ground rules the runbook enforces​

These apply whether a person or an agent is doing the install, and are worth knowing either way:

  • Secure by default, LAN-only. The default install never exposes Crumb to the public internet. An agent following the runbook will not open WAN firewall ports, set up port-forwarding, or stand up a public reverse proxy on its own initiative, even if asked to "make it work from anywhere," without first confirming TLS and a strong admin password are in place.
  • Never invents or prints secrets. Secrets come from scripts/setup-env.sh, which generates strong random values. An agent should never hardcode a password or secret, and should avoid echoing them into chat history if you ask for the admin password back; the script's --print flag is the intended path for that.
  • Confirms before privileged or destructive actions. Installing system packages, changing firewall rules, deleting data, or overwriting an existing .env are all things the runbook asks the agent to show you and confirm first, not run silently.
  • One step at a time, verified. Crumb is a recorder people rely on; the runbook is written to prefer correctness over speed, with an explicit Verify check after every step before moving on.

What the runbook covers​

The same ground as the manual path, plus the machinery for a fully hands-off setup: host prerequisites, secret generation, choosing a storage path, optionally enabling hardware decode, bringing up the stack, and then first-run configuration. For that last step, the runbook documents two paths: handing off to the web setup wizard (the simplest), or driving the entire first-run flow through the REST API for a fully scripted, headless install, including camera discovery, adding cameras in bulk, and setting up notifications and additional users without opening a browser.

It also covers what a fresh install needs after the first boot: confirming the nightly database backup is landing somewhere durable (see Backups), and, if you want it, setting up external monitoring for the API process itself, since Crumb's own alerting runs inside the API and can't page you if the API itself is down.

Using it​

Clone the repository on the target host, start your agent there, and ask it to read docs/AI-INSTALL.md and follow it. The file is self-contained; it does not assume the agent has read anything else in the repository first, though it links out to deeper docs (motion recording, backups, TLS) for background on specific decisions along the way.