> ## Documentation Index
> Fetch the complete documentation index at: https://docs.aperium.apps.hillspire.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Security and observability

> TLS, NetworkPolicy, audit logs, and the dashboards and alerts an on-prem deployment must expose.

## Security requirements

* All ingress must use TLS.
* Upstream credentials for every MCP connector must be stored in the approved secret manager and mounted only into the corresponding `aperium-mcp-<connector>` pod — never into the frontend, the backend, or other MCP pods.
* MCP auth tokens must not be shared with the frontend.
* NetworkPolicy must allow the backend to call each `aperium-mcp-<connector>` service and deny direct public access to `/mcp`.
* MCP write tools must be restricted by Aperium MCP permissions and by the upstream system's service-account permissions.
* Audit logs must identify user, tenant, agent, MCP server, tool name, request id, and write/read classification.
* The local model endpoint must not be reachable from user networks or the public internet.
* Egress to cloud LLM APIs must be blocked or explicitly exceptioned.

## Observability requirements

### Required dashboards (or equivalent views)

* Backend request latency, websocket health, and error rate.
* Per MCP service: `/healthz`, `/readyz`, tool count, request latency, and error rate.
* Per MCP service: tool-call success/failure broken down by tool name and read/write classification.
* Local model request latency, queue depth, tokens/sec, GPU memory, GPU utilization, and OOM count.
* PostgreSQL connection pool usage, locks, migration status, and backup status.
* Redis availability when multi-pod mode is enabled.
* Qdrant availability when retrieval or memory features are enabled.

### Required alerts

* Backend unavailable.
* Any MCP service readiness failing.
* Any MCP service discovery status not OK.
* Any MCP service auth failures.
* Write-tool errors above threshold on any MCP service.
* Local model readiness failing or GPU unavailable.
* Local model latency above the agreed SLO.
* PostgreSQL replication, backup, disk, or connection saturation issue.
* Redis unavailable in multi-pod mode.