Eve

See what an Eve agent did, which model and tools it called, how long each step took, and how many tokens the model used.

Eve emits OpenTelemetry (OTel), the open industry standard for collecting traces, metrics, and logs. Eve discovers agent/instrumentation.ts automatically and runs it before the agent starts. You configure that file to send Eve’s traces to Logfire.

Before you start

You need:

• An Eve project. Run npx eve@latest init my-agent to create one.
• A Logfire project and write token. Copy the token from Project → Settings → Write tokens.
• Credentials for your chosen model, configured through eve link or the model provider’s environment variables.

Install the packages that connect Eve’s OpenTelemetry data to Logfire:

Terminal

npm install @vercel/otel @opentelemetry/exporter-trace-otlp-proto \
  @opentelemetry/sdk-trace-base

Send Eve traces to Logfire

Create agent/instrumentation.ts:

agent/instrumentation.ts

import { OTLPTraceExporter } from '@opentelemetry/exporter-trace-otlp-proto';
import { SimpleSpanProcessor } from '@opentelemetry/sdk-trace-base';
import { registerOTel } from '@vercel/otel';
import { defineInstrumentation } from 'eve/instrumentation';

const writeToken = process.env.LOGFIRE_WRITE_TOKEN;
if (!writeToken) {
  throw new Error('LOGFIRE_WRITE_TOKEN is required');
}

export default defineInstrumentation({
  setup: ({ agentName }) =>
    registerOTel({
      serviceName: agentName,
      spanProcessors: [
        new SimpleSpanProcessor(
          new OTLPTraceExporter({
            url: 'https://logfire-us.pydantic.dev/v1/traces',
            headers: {
              Authorization: writeToken,
            },
          }),
        ),
      ],
    }),
  recordInputs: true,
  recordOutputs: true,
});

The example sends each completed span (one unit of work: a single operation, with a name, a start, and a duration) immediately. This is important for Eve agents running on Vercel Workflow because a serverless instance can pause before a batch of completed spans is sent.

Use https://logfire-eu.pydantic.dev/v1/traces instead if your Logfire project is in the European Union region.

Prompts and responses leave your application

With recordInputs and recordOutputs set to true, Eve sends model inputs, outputs, tool arguments, and tool results to Logfire. This data is stored in your Logfire project. Set either option to false before sending sensitive data that you do not want Logfire to store.

Set the write token in your terminal, then start the agent:

Terminal

export LOGFIRE_WRITE_TOKEN='<your-write-token>'
npm run dev

Do not put the token in agent/instrumentation.ts or commit it to source control. For a Vercel deployment, add LOGFIRE_WRITE_TOKEN in the project’s environment-variable settings.

Verify the trace

Send the agent a message such as:

Reply with EVE_LOGFIRE_OK.

Open your Logfire project’s Live view. You should see an Eve turn containing the agent invocation, model step, and model call. Open the trace to inspect the input, EVE_LOGFIRE_OK response, duration, model, and token usage. The exact span names depend on your Eve and artificial intelligence (AI) software development kit (SDK) versions.

The large language model (LLM) call also appears on the LLMs page. Logfire currently detects Eve agent runs and their conversations, but aggregate Agent metrics can miss turns or count repeated token attributes more than once. Use the LLMs page for exact model-call usage while full Eve Agent analytics support is in progress. Eve runs also appear in the specialized Agents view; the support matrix shows which columns each view populates.

Troubleshoot missing traces

• No Eve service appears in Live: confirm the file is exactly agent/instrumentation.ts, default-exports defineInstrumentation(...), and is present in the deployed build. Run npx eve info to inspect Eve’s discovery diagnostics.
• The exporter returns 401: pass the raw write token as the Authorization header value. Do not add a Bearer prefix.
• Local traces work but some Vercel Workflow traces are missing: keep SimpleSpanProcessor as shown above. A timer-based batch can remain in memory when the serverless instance pauses.
• The service appears as the wrong name: Eve derives agentName from the package name, falling back to the application directory. Set a meaningful name in package.json.

Next steps

Add tools or subagents to your Eve project and repeat the verification. Their work appears as nested spans in the same trace, so you can see which step caused a slow response or error.