Observe - Phoenix

Trace an OpenAI call and send it to Arize Phoenix. Background: What are traces. You need a running Phoenix; see Start Phoenix. The example calls OpenAI, so set OPENAI_API_KEY too. Have an existing app? Let your coding agent set it up.

Instrument OpenAI

Register Phoenix as the trace destination and turn on auto-instrumentation. Phoenix instruments every installed OpenInference package it finds, OpenAI here.

Python
TypeScript

pip install -U arize-phoenix-otel openai openinference-instrumentation-openai

from phoenix.otel import register

register(
    project_name="tracing-quickstart",
    protocol="http/protobuf",
    auto_instrument=True,
)

npm install openai @arizeai/phoenix-otel @arizeai/openinference-instrumentation-openai

Put this in instrumentation.ts and import it before anything else:

import OpenAI from "openai";
import { OpenAIInstrumentation } from "@arizeai/openinference-instrumentation-openai";
import { register, registerInstrumentations } from "@arizeai/phoenix-otel";

export const provider = register({ projectName: "tracing-quickstart" });

const instrumentation = new OpenAIInstrumentation();
instrumentation.manuallyInstrument(OpenAI);
registerInstrumentations({ instrumentations: [instrumentation] });

Make a call

Any OpenAI call now produces a trace.

Python
TypeScript

from openai import OpenAI

client = OpenAI()
client.chat.completions.create(
    model="gpt-5-mini",
    messages=[{"role": "user", "content": "Why did my invoice change?"}],
)

import "./instrumentation";
import OpenAI from "openai";
import { provider } from "./instrumentation";

const client = new OpenAI();
await client.chat.completions.create({
  model: "gpt-5-mini",
  messages: [{ role: "user", content: "Why did my invoice change?" }],
});
await provider.shutdown();

Look at the trace

Open the tracing-quickstart project in Phoenix and click the trace: the prompt, response, tokens, and latency are on the span. To trace your own code around the LLM call, wrap it with a decorator or traceChain.

Learn more about traces

Other ways to read your traces:

In code, with Extract data from spans.
From a coding agent, with the CLI or MCP.
By asking PXI, for example “why did this trace fail?”
Across multi-turn conversations, with Sessions.

To go deeper on tracing, the Tracing Tutorial covers how to read spans, understand timing, and use trace data to debug and analyze your agent. The code on this page lives in examples/quickstarts, ready to run.