Eve

Eve's spans are standard OpenTelemetry, built on the Vercel AI SDK telemetry conventions Harbor already reads. There is no adapter package to install — Eve's registerOTel setup takes a plain spanProcessors array, so HarborSpanProcessor plugs straight in.

Install

npm install harbor-ai-sdk

registerOTel comes from @vercel/otel, which Eve already depends on — nothing Eve-specific to add.

Set up

import { registerOTel } from '@vercel/otel'
import { defineInstrumentation } from 'eve/instrumentation'
import { HarborSpanProcessor } from 'harbor-ai-sdk'

export default defineInstrumentation({
  setup: ({ agentName }) =>
    registerOTel({
      serviceName: agentName,
      spanProcessors: [new HarborSpanProcessor()],
    }),
})

That is the whole integration. setup runs once at server startup, before Eve's first request, and agentName — passed through from defineAgent — becomes gen_ai.agent.name in Harbor.

What you get

invoke_agent {agent}          the turn
  chat {model}                 a model call (step)
    execute_tool {name}         a tool the agent invoked
  chat {model}

Turn, step and tool call spans arrive as one nested trace, carrying model, provider, token usage, messages and tool arguments — the same GenAI attributes as the Vercel AI SDK integration, since Eve's telemetry is built on it.

Add the user and session

Eve tracks its own session and turn lineage internally, but doesn't expose a Harbor-shaped user id, so capture supplies it — the same as with the Vercel AI SDK:

import { capture } from 'harbor-ai-sdk'

const result = await capture('agent-turn', () => agent.run(input), {
  userId: user.id,
  sessionId: session.id,
})

Reuse sessionId across turns and Harbor stitches them into one session.