Visual Tour

These captures show real AgentGuard documentation and deterministic product output. They use repository-provided synthetic fixtures, contain no external agent or model activity, and make no claim about broader production adoption or universal detection quality. See the source and sanitization record for reproduction details and limitations.

Product overview

The public AgentGuard documentation puts the local-first evaluation model, current release, package identity, and first installation commands on one screen. The visible navigation leads into the architecture, benchmark, evidence, artifact, and integration guides.

Evaluation dashboard

The generated static report site summarizes six deterministic showcase scenarios, their suite record, and one separate audit-mode incident run. The single safe fixture scores 100 and passes; unsafe fixtures fail through their configured checks. These counts describe this small demo corpus only.

Guard incident detail

This real online guard incident comes from the deterministic showcase_unsafe_command fixture in audit mode. The page shows category, mode, status, severity, timing, and a fixed evidence summary without rendering the raw command payload, environment data, diff, or an absolute path.

Evaluation evidence

The static trend view aggregates the same generated records. It connects the eight represented records to one sanitized command-guard incident, sixteen failed checks, seven expected unsafe failures, and one safe pass. The showcase metrics separately retain the committed 5/5 unsafe-detection and 1/1 safe-allowance evidence for the six-scenario showcase.

Interpretation boundary

These screenshots demonstrate reporting and evidence presentation, not a perfect security boundary or a statistical effectiveness claim. Local agent execution uses host-user permissions and is not inherently sandboxed. Docker can provide configured containment for Docker-backed evaluations.