hermes-otel

What hermes-otel does

Each turn becomes a tree of spans: an agent or cron root, skill.{name} for loaded skills, llm.{model} for each conversation run, api.{model} for each HTTP round-trip with token counts, cost and finish reason, tool.{name} for tool calls with arguments, result and outcome, approval.{pattern} for approval waits and subagent.{role} for delegated tasks, whose child trace nests inside the parent. Token, cost and latency metrics are emitted, with optional logs and structured events.

Attributes follow two conventions at once, OpenInference for Phoenix and OTel GenAI for Langfuse, Weave and generic dashboards, so backends need no custom mapping. The README lists tested backends including Phoenix, Langfuse, LangSmith, SigNoz, Jaeger, Grafana Tempo, Honeycomb, W&B Weave and Elastic, among others. A documentation site covers the quickstart, backend guides, configuration and references for span attributes, metrics and hooks.

Key features

• One OTel trace per Hermes turn, with nested sub-agent traces
• Spans for LLM calls, API round-trips, tools, skills and approvals
• Token, cost and latency metrics, plus optional logs and events
• Attributes in both OpenInference and OTel GenAI conventions
• Export to any OTLP/HTTP backend, with 19 listed as tested
• Tool outcome recorded as completed, error, timeout, blocked or cancelled

When to use it

• Finding why a Hermes turn was slow by reading its span tree in Jaeger or Grafana Tempo
• Tracking token cost per model in Langfuse or Phoenix
• Following a multi-agent run as a single connected trace

Who it is for: Developers and operators who run Hermes Agent and want standard observability data about its behavior.

How it fits with Hermes Agent

Built for Hermes Agent as a plugin that hooks into each turn, tool call, skill load, approval and delegated sub-agent.

FAQ