Agent Protocol
How agents connect to Vigilhold, and how commands and updates reach them.
Connection
The agent opens one outbound, encrypted WebSocket to vigilhold.com on TCP 443, falling back to the Vigilhold nodes on TCP 8443, and proves who it is with its own certificate, issued at install. Nothing connects in to your server.
Each connection is also signed with the agent's secret and the current time. Vigilhold refuses a signature more than 30 seconds off, or one it has already seen, so a server whose clock has drifted cannot connect: turn on NTP (timedatectl set-ntp true).
What the agent sends
Metrics (CPU, memory, disk, network), what is on the host (processes, services, packages, connections, containers) and what its checks find, plus new lines from the system's authentication, system and kernel logs, and the text of a short list of system configuration files with values that look secret blanked. The Trust page has the full list. Outbound destinations leave as fingerprints only the server can make. Cortex, the rules-based engine that decides what to do about a threat, runs in Vigilhold's backend, not on your server.
Commands
Commands travel from Vigilhold to the agent over the same connection, each signed and accepted once. Stopping a process, quarantining a file, stopping a container or Kubernetes pod, ending SSH sessions, locking down or isolating a server, undoing a containment, updating the agent or changing its keys or servers also needs a second signature from Vigilhold's own signing key, which covers the whole command. The agent knows that key from its signed release, so a stolen agent secret alone cannot issue these commands.
The agent refuses anything outside the host's own limits (/etc/vigilhold/host-limits, see Installing the Agent → --limits), whoever asks.
Updates
Agent updates are signed releases. The agent downloads and checks a release; a separate updater running as root checks the signature and checksum again before installing, refuses downgrades, and puts the previous version back if the new one fails to start. You choose when each host updates in Settings → Fleet.
Tips
- If an agent cannot connect, check outbound TCP 443 and the server's clock first, then read journalctl -u sentinel-agent -n 50 (on Alpine: tail -n 50 /var/log/sentinel-agent/agent.log).