Skip to content

Runtime Monitoring

For the full runtime monitoring deployment guide, see the dedicated doc.

Overview

The runtime proxy (agent-bom proxy) intercepts MCP JSON-RPC messages between client and server, providing:

  • JSONL audit logging of all tool calls
  • 7-detector anomaly engine (tool drift, argument analysis, credential leak, rate limiting, sequence analysis, response inspector, vector DB injection)
  • Policy enforcement with block/allow rules
  • Prometheus metrics on port 8422
  • Optional visual leak detection for image and screenshot responses

Deployment modes

Mode Command Use case
Local sidecar agent-bom proxy -- npx server Dev/testing
Docker sidecar See Docker Production
K8s sidecar See Kubernetes Fleet
Optional node-wide monitor Helm monitor.enabled=true Broad runtime coverage only when a team explicitly accepts a DaemonSet
Config watcher agent-bom watch Drift alerting

The node-wide monitor is:

  • optional
  • off by default
  • not required for scan/discovery, fleet, gateway, or selected sidecar proxy rollout
  • the highest-trust runtime shape, so it should be enabled only when the operator wants per-node runtime coverage

Alert routing

The watch command supports webhook alerts to:

  • Slack
  • Microsoft Teams
  • PagerDuty
  • Custom webhook URLs
agent-bom watch --webhook-url https://hooks.slack.com/... --watch-interval 60

Operator guides