tracelet-instrument

tracelet-instrument

Goal: make an agent's execution visible — every LLM call, tool call, prompt, token count and latency — without sending anything to a cloud service. tracelet is a local OTLP collector + UI; any OpenTelemetry exporter pointed at http://127.0.0.1:4318 shows up live.

Procedure

1. Start the collector (keep it running in a terminal):

   npx @jnmetacode/tracelet        # ingests OTLP on :4318, UI on :4321
   

2. Point the app's exporter at it. For most OpenTelemetry SDKs the zero-code route is environment variables:

   export OTEL_EXPORTER_OTLP_ENDPOINT=http://127.0.0.1:4318
   export OTEL_SERVICE_NAME=my-agent
   

3. Framework-specific wiring (pick the one that matches the codebase):

   Vercel AI SDK — pass telemetry on each call and register a Node tracer:

   const result = await generateText({
     model, prompt,
     experimental_telemetry: { isEnabled: true },
   });
   

   OpenTelemetry JS (manual spans):

   import { NodeSDK } from '@opentelemetry/sdk-node';
   import { OTLPTraceExporter } from '@opentelemetry/exporter-trace-otlp-http';
   new NodeSDK({ traceExporter: new OTLPTraceExporter({ url: 'http://127.0.0.1:4318/v1/traces' }) }).start();
   

   Python: pip install opentelemetry-sdk opentelemetry-exporter-otlp-proto-http, then the same OTEL_EXPORTER_OTLP_ENDPOINT env var works.

4. Name spans so the waterfall reads like the agent's plan: one root span per run (agent.run), one child per tool call (tool.<name>) and per model call. Attach gen_ai.request.model, input.value, output.value attributes — tracelet recognizes the GenAI/OpenInference conventions and renders prompts, tokens and tool I/O.

5. Verify: trigger one agent run, open http://127.0.0.1:4321, and confirm the trace shows the expected span tree with prompts/tokens. If nothing arrives, the exporter is usually pointed at the wrong port (4318) or batching hasn't flushed — force a flush/shutdown on process exit.