Live demo scenarios5 reproducible security proofs: No provider API keys in agent pods • gVisor Python sandbox • Cryptographic tenant isolation • Gateway request circuit breakerExplore the 5 Proofs
Capability · Observability

One trace per agent run in self-hosted Langfuse—with zero SDKs in your code.

Zelkor links agent runs, AI gateway calls, guardrail checks, and MCP tool executions into a single trace inside self-hosted Langfuse, while keeping prompts and keys out of default logs.

Why Standard Container Logs Fail for Agents

Standard pod logs scatter a single agent run across multiple services, dump sensitive customer prompts or bearer tokens into cluster log pipelines, and leave gaps whenever a developer forgets to wrap a new graph node in a tracing library.

Zero-SDK Run Correlation

Without adding Langfuse or OpenTelemetry imports to your agent code, Zelkor automatically joins the agent run, gateway model calls, NeMo guardrail checks, and MCP tool calls under one trace.

Prompts and Keys Redacted by Default

Default logs and trace metadata do not record raw prompt bodies, bearer tokens, or upstream API keys, keeping sensitive data out of cluster log collectors.

Built-In Run Checks

Self-hosted Langfuse comes pre-configured with automated checks that score runs for tool call budgets, repeated tool loops, sandbox cleanliness, guardrail refusals, and tenant attribution.

In-Cluster Storage and Prometheus Metrics

All trace data stays inside your cluster on self-hosted Langfuse, ClickHouse, Postgres, and S3-compatible storage, alongside an optional Prometheus ServiceMonitor for cluster monitoring.

How Teams Use It

platform:
  telemetry:
    level: INFO
    langfuse:
      enabled: true
observability:
  serviceMonitor:
    enabled: true

Get complete flight-recorder visibility for every agent run inside your own cluster.