Security model
A tool that reaches your servers has to earn trust. Here is how.
Least privilege
The per-server agent exposes a narrow HTTP API — no shell. It enforces its own command allowlist, so a hub compromise cannot become arbitrary shell.
Approval gates
Commit, push, deploy, rollback, and migrations are approval-gated. High-risk actions need an exact confirmation phrase.
Audit
Every request, command, and approval is written to an append-only audit log, exportable as JSON or Markdown. Secrets are never logged.