Confined, auditable system diagnostics for AI agents
maat exposes read-only Ubuntu diagnostics (journal, systemd units, processes,
filesystems, AppArmor denials, network interface counters) to AI agents over
the Model Context Protocol, with the limits enforced by snap confinement
rather than by the model's good behaviour. It records typed, reviewable plans that a human applies
with the separate, unconfined maatctl tool after authorizing them through
polkit.
This is a development preview. The server is strictly confined and plugs only
read-oriented interfaces; it has no network access.