Cadastre is a catalog service, not a Broker and not a deployment engine. The container has no Docker socket, no collector credentials, and no infrastructure write capability.

The profile is chosen with --profile on serve and security-check. There are five, and these are their exact names:

Every remote profile must specify endpoint identity, trust anchor, audience, principal/scope mapping, allowed operations, request limits, Host and Origin allowlists, and an explicit write decision. Missing identity or trust material fails closed. Run cadastre security-check before exposing an endpoint.

Agent access has two separate components: the Cadastre MCP server runs with the application stack, while an agent either uses a native remote MCP client or installs cadastre-mcp-remote as a local stdio bridge. The bridge is remote-only and must never open a local catalog fallback. The GUI is likewise an unprivileged HTTP/API client and never receives the SQLite volume or a Broker path.

The data directory contains two separate SQLite databases. Take a transaction- consistent cadastre backup before upgrades, retain encrypted/off-host copies, and test restoration into a fresh directory. DEPLOYMENT.md §7.1 is the procedure that follows that backup. Export bundles are review and recovery aids; they do not replace SQLite backups because they omit transaction history and may omit observed history by operator choice.

Fresh-user path:

1. Pull a published image by signed immutable digest.
2. Create the named persistent volume and run cadastre init --data-dir /var/lib/cadastre.
3. Choose a profile: loopback-development, direct-https, mtls, or trusted-proxy.
4. Verify readiness and run the first integrity check.
5. Configure optional collectors as separate jobs with separate read-only mounts.
6. Take and verify the first off-host backup.