What leaves your network

Your servers talk only to your hub. Your hub talks only to what you set up. Set up nothing and nothing leaves.

From your hub

To	Only if you	What
Anthropic (Claude)	add an API key	The alert, what the statistics found (with sample log lines), this alert's history, your runbook, and what the AI's own read-only queries return (200 log lines at most).
Teams, Slack, a webhook	set one	The incident card.
Jira, GitHub, Azure DevOps	link tickets	Read-only: is the ticket still open.
Your Prometheus, Loki, Alertmanager	connect them	Read-only queries; Whyfire's silences, in your Alertmanager.
Your websites	add a check	One request every 30 seconds.
Let's Encrypt	give the hub a domain	The certificate request.
Docker Hub, GitHub, npm	install or upgrade	Downloads: images, Grafana Alloy, packages.

Metrics and logs go to the hub only, over HTTPS, checked against the hub's certificate. Nothing listens on your servers.
The agent script comes from the hub over plain HTTP. The command checks its checksum, and the script checks the hub's certificate by fingerprint. The command's token works for 24 hours.
One thing is still readable on your network during an install: the key the server pushes with. Someone watching at that moment could send the hub data under a server's name, not read or change anything. HTTPS for the install is planned.
Visitors' addresses in web logs (--web-logs) are shortened on the server, before they leave it.

No phone-home: no usage statistics, update checks or licence calls. The licence is checked on the hub, by its signature.
We never receive your data. For support you send sudo whyfire support-bundle, secrets removed.

The AI and the cards quote log lines. Keep secrets out of the logs Whyfire collects, or run without an AI key: everything else still works.