Operations and Local Data¶
Agent Loop Guard has no remote control plane or product telemetry. Operational state lives on the machine where the daemon runs.
Health and status¶
doctor checks the Python runtime, configuration, storage path, and optional capabilities. status probes the running daemon. The health API is GET /api/health.
Storage¶
The default database is data/agent_loop_guard.db relative to the process working directory. Setup-generated data is kept under .agent-loop-guard. Use an absolute SQLite path when starting the daemon from different directories:
Database initialization applies the included Alembic revisions. Back up data before upgrading across versions.
Backup and restore¶
Restore refuses to overwrite existing local state unless --force is present. Stop the daemon before restoring so SQLite files are not being written concurrently.
Retention¶
Without --days, the configured storage.retention_days value is used. Pin important Replay runs before cleanup. Treat exports and backups as sensitive even though redaction is applied.
Binding and access¶
Keep the default loopback host:
The gateway key protects model proxy and MCP traffic, but the local admin and Replay interfaces are designed for loopback use and are not an internet-facing multi-user service. Do not bind to 0.0.0.0 without an authenticated reverse proxy, firewall rules, TLS, and a careful security review.
Logs and privacy¶
- Full request content logging is disabled by default.
- Common API key, authorization, password, token, and cookie fields are redacted recursively.
- Tool arguments are represented by fingerprints and selected metadata where possible.
- Chain-of-thought is neither requested nor stored.
- No user telemetry is sent by Agent Loop Guard.
Redaction lowers risk but cannot recognize every project-specific secret. Never use logs as a secret store, and inspect exported traces before sharing them.