@agentmug/otel

OpenTelemetry tracing adapter for @agentmug/runtime. Emits LLM-call + transcription spans into your existing OTel SDK pipeline — Honeycomb, Datadog, Tempo, Jaeger, Sentry, anything that consumes OpenTelemetry.

npm install @agentmug/otel @opentelemetry/api
# plus whichever OTel SDK + exporter your stack uses

Usage

Set up your OTel SDK once at process start (you almost certainly already have this in your app). Then plug the adapter into runAgent():

import { runAgent } from "@agentmug/runtime";
import { OtelTracingAdapter } from "@agentmug/otel";

const result = await runAgent({
  agentId: "...",
  userId: "...",
  userInput: "...",
  adapters: {
    persistence: yourPersistence,
    llm: yourLlmClient,
    tracing: new OtelTracingAdapter(),  // ← that's it
  },
  onEvent: () => {},
});

Every LLM call becomes a span with these attributes:

• agentmug.run.id
• agentmug.agent.id
• agentmug.llm.model
• agentmug.llm.tokens
• agentmug.llm.cost_cents
• agentmug.llm.latency_ms
• agentmug.llm.input_preview (truncated to 512 chars)
• agentmug.llm.output_preview (truncated to 512 chars)

Transcription calls get their own spans with provider + audio token counts.

As a plugin

createOtelPlugin() packages the adapter as an AgentMugPlugin, which is handy if you already pass plugins around. It does not register tracing by itself — InMemoryToolRegistry.loadPlugin only reads plugin.tools, and this plugin ships no tools. You still have to hand the adapter to runAgent:

import { createOtelPlugin } from "@agentmug/otel";

const otelPlugin = createOtelPlugin();

await runAgent({
  // ...
  adapters: { tracing: otelPlugin.adapters.tracing },
});

If you are not already routing plugins through your own wiring, construct OtelTracingAdapter directly as shown above — it is the same instance.

Custom tracer

Pass your own Tracer if you need per-tenant tracing, test isolation, or want to attach resource attributes:

import { trace } from "@opentelemetry/api";
new OtelTracingAdapter({
  tracer: trace.getTracer("my-app/agentmug", "2.1.0"),
});

What this doesn't do

• SDK setup — every OTel pipeline is different. Configure your NodeSDK / WebSDK / exporter as you normally would.
• Metrics — token counts live as span attributes. If you want histograms, derive them from spans in your backend.
• Logs — the runtime's onEvent callback gives you token-by-token output. Wire that to your logging pipeline directly.

License

Apache-2.0. Existing versions published under MIT remain MIT.