Skip to main content
Export traces, metrics, and logs to SigNoz. On by default to local OTLP at http://127.0.0.1:4318 when SigNoz is running.

​
Quick start

​
Autostart on login (macOS / Linux)

After a one-time arka signoz setup -y, enable the local Docker stack to come up at login:
On macOS, also enable Docker Desktop → Settings → General → Start Docker Desktop when you sign in. Autostart waits for Docker, then runs foundryctl cast -f casting.yaml --no-gauge. Logs: ~/.cache/fish-agent/signoz-autostart.log (or your CACHE_DIR). Remove with arka signoz autostart uninstall. Pass --autostart to arka signoz setup -y to install in the same step. arka observability doctor is a read-only local check. It reports whether OpenTelemetry is enabled and configured, but does not claim that SigNoz has received data. After enabling telemetry, run arka signoz demo and verify service.name = arka in the SigNoz traces view. If the doctor reports a configured exporter but no traces appear, the problem is ingestion or the SigNoz endpoint—not Arka’s route instrumentation.

​
Configuration

​
What gets traced

  • LLM provider calls (model, latency, tokens, failover events)
  • Symbolic / deterministic routing — offline rule matches, matched route_* handler, sub-ms route latency, and arka.llm.used=false
  • Deterministic skill execution — local skills (password, calc, convert, help, etc.) with arka.execution.kind=deterministic
  • Skill dispatch and routing decisions (linked under arka.request for natural-language CLI input)
  • MCP client calls
  • Supermemory operations
Every dispatched skill records a bounded skill name, duration, exit status, execution kind (deterministic, llm_skill, llm_routed, fish, or shell), and whether an LLM was used. Secrets and full user prompts are not added to these attributes. This makes failures in long-running loops and MCP-backed skills visible without turning exported logs into a copy of the user’s code or credentials. Structured logs include trace_id and span_id when emitted inside an active span — use SigNoz log ↔ trace correlation to jump from a skill failure log to its trace. Metrics arka.routing.decisions, arka.skill.duration, and arka.llm.failover complement span attributes in the Metrics explorer. Useful SigNoz trace filters:
  • name = 'arka.request' — full NL request waterfall (route → skill)
  • name = 'arka.route' AND arka.route.decision = 'symbolic' — offline symbolic routes only
  • arka.execution.kind = 'deterministic' — local skills with no LLM
  • arka.llm.used = false — any span where no LLM participated

​
SigNoz dashboard

Install the bundled Arka agent observability dashboard (15 panels: service overview, skill dispatch latency/errors, routing, LLM failover, model/provider mix, token usage, correlated logs):
Requires SIGNOZ_API_KEY from SigNoz → Settings → Service Accounts. Template JSON: signoz/dashboards/arka-agent-observability.json. MCP-orchestrated install via self-build:

​
SigNoz MCP + Cursor

Bundled SigNoz assets (dashboards, alerts, demos) live in signoz/ in the repo.
OTLP export is on by default to http://127.0.0.1:4318. Set OTEL_SDK_DISABLED=true or OTEL_TRACES_ENABLED=0 to opt out.

​
Production-to-codebase connectors

Export JSON from SigNoz, Sentry, Datadog, or New Relic, then correlate stack paths with the current repository:
This is read-only: it reports event counts, providers, unmatched events, and frequent file/line hotspots without uploading exports, reading provider credentials, or editing code.

Related topics

Testing Arka with the pytest suiteHow Arka compares to other agentsMCP integration for Cursor and ClaudeTroubleshooting Arka: LLM, routing, voice, and RAG fixesOutput viewer