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