OpenAI Realtime API

Weave’s integration with OpenAI’s Realtime API lets you automatically trace your application’s speech-to-speech interactions as they occur. You can use this to capture conversations between your agents and users to review and evaluate agent performance. This guide shows you how to enable Realtime API tracing in your application and walks through two end-to-end voice assistant examples that you can run locally.

Integrate realtime traces

This section shows the minimum code needed to enable Weave tracing for the OpenAI Realtime API in any application. Weave automatically patches into the OpenAI Realtime API and only requires a few additional lines of code to start capturing audio interactions from your application. The following code imports Weave and the Realtime API integration:

import weave
from weave.integrations import patch_openai_realtime

weave.init("your-team-name/your-project-name")
patch_openai_realtime()

# Your application's logic

Once imported into your code and executed, Weave automatically traces the interactions between the user and the OpenAI Realtime API.

Run a realtime voice assistant with the OpenAI Agents SDK

This example runs a realtime voice assistant that streams microphone audio to OpenAI’s Realtime API and plays back the AI’s spoken responses through your local machine’s speaker. The application uses the OpenAI Agents SDK with RealtimeAgent and RealtimeRunner, and enables Weave tracing by patching with patch_openai_realtime(). Use this approach when you want the higher-level Agents SDK to manage the Realtime session for you. To run the example, complete the following steps:

Start your Python environment and install the following libraries:

uv add weave openai-agents websockets pyaudio numpy

pip install weave openai-agents websockets pyaudio numpy

Create a file titled weave_voice_assistant.py and add the following code to it. The highlighted lines indicate Weave’s integration in the application. The rest of the code creates the basic voice assistant app.

weave_voice_assistant.py

import argparse
import asyncio
import queue
import sys
import termios
import threading
import tty
import weave
import pyaudio
import numpy as np
from weave.integrations import patch_openai_realtime
from agents.realtime import RealtimeAgent, RealtimeRunner

DEFAULT_WEAVE_PROJECT = "<your-team-name/your-project-name>"

FORMAT = pyaudio.paInt16
RATE = 24000  # Required by the OpenAI Realtime API.
CHUNK = 1024
MAX_INPUT_CHANNELS = 1
MAX_OUTPUT_CHANNELS = 1

INP_DEV_IDX = None
OUT_DEV_IDX = None

# Parse CLI args for Weave project name and audio device selection.
def parse_args():
    parser = argparse.ArgumentParser(description="Realtime agent with Weave logging")
    parser.add_argument(
        "--weave-project",
        default=DEFAULT_WEAVE_PROJECT,
        help=f"Weave project name (default: {DEFAULT_WEAVE_PROJECT})",
        dest="weave_project"
    )
    parser.add_argument(
        "--input-device",
        type=int,
        default=None,
        help="PyAudio input (mic) device index. Defaults to system default. Run mic_detect.py to list devices.",
        dest="input_device"
    )
    parser.add_argument(
        "--output-device",
        type=int,
        default=None,
        help="PyAudio output (speaker) device index. Defaults to system default. Run mic_detect.py to list devices.",
        dest="output_device"
    )
    return parser.parse_args()


# Initialize Weave and patch the OpenAI Realtime API for tracing.
def init_weave(project_name: str | None = None) -> None:
    name = project_name or DEFAULT_WEAVE_PROJECT
    weave.init(name)
    patch_openai_realtime()  # Enables automatic tracing of Realtime API sessions.


mic_enabled = True

# Listen for 't' key to toggle mic on/off. Runs in a daemon thread.
def start_keylistener():
    global mic_enabled
    fd = sys.stdin.fileno()
    old_settings = termios.tcgetattr(fd)
    try:
        tty.setcbreak(fd)
        while True:
            ch = sys.stdin.read(1)
            if ch.lower() == 't':
                mic_enabled = not mic_enabled
                state = "ON" if mic_enabled else "OFF"
                print(f"\n🎙  Mic {state} (press t to toggle)")
            elif ch == '\x03':  # Ctrl-C
                break
    finally:
        termios.tcsetattr(fd, termios.TCSADRAIN, old_settings)


# Drain the audio queue and write to the speaker in a background thread.
def play_audio(output_stream: pyaudio.Stream, audio_output_queue: queue.Queue):
    while True:
        data = audio_output_queue.get()
        if data is None:
            break
        output_stream.write(data)


# Open audio streams, start a Realtime session, and run the send/receive loop.
async def main(*, input_device_index: int | None = None, output_device_index: int | None = None):
    p = pyaudio.PyAudio()

    if input_device_index is None:
        input_device_index = int(p.get_default_input_device_info()['index'])
    if output_device_index is None:
        output_device_index = int(p.get_default_output_device_info()['index'])

    # Clamp channel count to device capabilities to avoid pyaudio errors.
    input_info = p.get_device_info_by_index(input_device_index)
    output_info = p.get_device_info_by_index(output_device_index)

    input_channels = min(int(input_info['maxInputChannels']), MAX_INPUT_CHANNELS)
    output_channels = min(int(output_info['maxOutputChannels']), MAX_OUTPUT_CHANNELS)

    mic = p.open(
        format=FORMAT,
        channels=input_channels,
        rate=RATE,
        input=True,
        output=False,
        frames_per_buffer=CHUNK,
        input_device_index=input_device_index,
        start=False,
    )
    speaker = p.open(
        format=FORMAT,
        channels=output_channels,
        rate=RATE,
        input=False,
        output=True,
        frames_per_buffer=CHUNK,
        output_device_index=output_device_index,
        start=False,
    )
    mic.start_stream()
    speaker.start_stream()

    # Buffer audio through a queue so in-flight playback can be flushed on interrupt.
    audio_output_queue = queue.Queue()
    threading.Thread(
        target=play_audio, args=(speaker, audio_output_queue), daemon=True
    ).start()

    s_agent = RealtimeAgent(
        name="Speech Assistant",
        instructions="You are a tool using AI. Use tools to accomplish a task whenever possible"
    )

    s_runner = RealtimeRunner(s_agent, config={
        "model_settings": {
            "model_name": "gpt-realtime",
            "modalities": ["audio"],
            "output_modalities": ["audio"],
            "input_audio_format": "pcm16",
            "output_audio_format": "pcm16",
            "speed": 1.2,
            "turn_detection": {
                "prefix_padding_ms": 100,
                "silence_duration_ms": 100,
                "type": "server_vad",
                "interrupt_response": True,
                "create_response": True,
            },
        }
    })
    print("--- Session Active (Speak into mic) ---")
    print("🎙  Mic ON (press t to toggle)")

    threading.Thread(target=start_keylistener, daemon=True).start()

    async with await s_runner.run() as session:
        # Stream mic input to the Realtime API, sending silence when muted.
        async def send_mic_audio():
            silence = b'\x00' * CHUNK * 2  # 2 bytes per sample (16-bit PCM).
            try:
                while True:
                    raw_data = mic.read(CHUNK, exception_on_overflow=False)

                    if mic_enabled:
                        audio_data = np.frombuffer(raw_data, dtype=np.int16).astype(np.float64)
                        rms = np.sqrt(np.mean(audio_data**2))
                        meter = int(min(rms / 50, 50))
                        print(f"Mic Level: {'█' * meter}{' ' * (50-meter)} | 🎙 ON ", end="\r")
                        await session.send_audio(raw_data)
                    else:
                        print(f"Mic Level: {' ' * 50} | 🎙 OFF", end="\r")
                        await session.send_audio(silence)

                    await asyncio.sleep(0)  # Yield to the event loop between reads.
            except Exception:
                pass

        # Receive events from the session and route audio to the speaker.
        async def handle_events():
            async for event in session:
                if event.type == "audio":
                    audio_output_queue.put(event.audio.data)
                elif event.type == "audio_interrupted":
                    # Flush queued AI audio so it doesn't talk over the user.
                    while not audio_output_queue.empty():
                        try:
                            audio_output_queue.get_nowait()
                        except queue.Empty:
                            break

        mic_task = asyncio.create_task(send_mic_audio())
        try:
            await handle_events()
        finally:
            mic_task.cancel()

    # Cleanup
    audio_output_queue.put(None)  # Signal playback thread to exit.
    mic.close()
    speaker.close()
    p.terminate()

if __name__ == "__main__":
    args = parse_args()
    init_weave(args.weave_project)
    fd = sys.stdin.fileno()
    old_settings = termios.tcgetattr(fd)
    try:
        asyncio.run(main(input_device_index=args.input_device, output_device_index=args.output_device))
    finally:
        termios.tcsetattr(fd, termios.TCSADRAIN, old_settings)

Update the DEFAULT_WEAVE_PROJECT value with your team and project names.
Set your OPENAI_API_KEY environment variable.
Start the voice assistant:
python weave_voice_assistant.py

Once the assistant is running, press T on your keyboard to mute or unmute the mic. The assistant uses server-side voice activity detection to handle turn-taking and interruptions. As you speak to the assistant, Weave captures traces that you can explore in the Weave UI, including the audio from the session.

Run a realtime voice assistant with WebSockets

The following example connects directly to the OpenAI Realtime API over WebSockets. It streams microphone audio to the API, plays back spoken responses, and supports tool calling (weather lookups, math evaluation, code execution, and file writing). Weave traces the session using weave.init() and patch_openai_realtime(). Use this approach when you want full control over the Realtime session without depending on the Agents SDK. To run the example, complete the following steps:

Start your Python environment and install the following libraries:

uv add weave websockets pyaudio numpy

pip install weave websockets pyaudio numpy

Create a file titled tool_definitions.py and add the following tool definitions to it. The main application imports from this module.

tool_definitions.py

import json
import subprocess
import tempfile
from pathlib import Path
import weave


# @function_tool
@weave.op
def get_weather(city: str) -> str:
    """Get the current weather for a city.

    Args:
        city: The city name to get weather for.
    """
    return json.dumps({"city": city, "temperature": "72°F", "condition": "sunny"})


@weave.op
def calculate(expression: str) -> str:
    """Evaluate a math expression and return the result.

    Args:
        expression: A math expression to evaluate, e.g. '2 + 2'.
    """
    try:
        result = eval(expression)
        return str(result)
    except Exception as e:
        return f"Error: {e}"


@weave.op
def run_python_code(code: str) -> str:
    """Write and execute a Python script, returning its stdout/stderr.

    Args:
        code: The Python source code to execute.
    """
    with tempfile.NamedTemporaryFile(
        mode="w", suffix=".py", dir=tempfile.gettempdir(), delete=False
    ) as f:
        f.write(code)
        script_path = Path(f.name)

    try:
        result = subprocess.run(
            ["python", str(script_path)],
            capture_output=True,
            text=True,
            timeout=30,
        )
        output = result.stdout
        if result.stderr:
            output += f"\nSTDERR:\n{result.stderr}"
        if result.returncode != 0:
            output += f"\n(exit code {result.returncode})"
        return output or "(no output)"
    except subprocess.TimeoutExpired:
        return "Error: script timed out after 30 seconds."
    finally:
        script_path.unlink(missing_ok=True)


@weave.op
async def write_file(file_path: str, content: str) -> str:
    """Write content to a file on disk.

    Args:
        file_path: The path to write the file to.
        content: The content to write into the file.
    """
    try:
        path = Path(file_path)
        path.parent.mkdir(parents=True, exist_ok=True)
        path.write_text(content)
        return f"Wrote {len(content)} bytes to {file_path}"
    except Exception as e:
        return f"Error writing file: {e}"

Create a file titled weave_ws_voice_assistant.py in the same directory and add the following code to it.

weave_ws_voice_assistant.py

import asyncio
import base64
import json
import os
import queue
import threading
from typing import Any, Callable

import numpy as np
import pyaudio
import websockets

import weave
weave.init("<your-team-name/your-project-name>")
from weave.integrations import patch_openai_realtime
patch_openai_realtime()

from tool_definitions import (
    calculate,
    get_weather,
    run_python_code,
    write_file,
)

# Audio format (must be PCM16 for the Realtime API).
FORMAT = pyaudio.paInt16
RATE = 24000
CHUNK = 1024
MAX_INPUT_CHANNELS = 2
MAX_OUTPUT_CHANNELS = 2

OPENAI_API_KEY = os.environ.get("OPENAI_API_KEY", "")
REALTIME_URL = "wss://api.openai.com/v1/realtime?model=gpt-realtime"

DEBUG_WRITE_LOG = False

# Map tool name -> callable for function call dispatch.
TOOL_REGISTRY: dict[str, Callable[..., Any]] = {
    "get_weather": get_weather,
    "calculate": calculate,
    "run_python_code": run_python_code,
    "write_file": write_file,
}

# Raw tool definitions for the Realtime API session config.
TOOL_DEFINITIONS = [
    {
        "type": "function",
        "name": "get_weather",
        "description": "Get the current weather for a city.",
        "parameters": {
            "type": "object",
            "properties": {
                "city": {
                    "type": "string",
                    "description": "The city name to get weather for.",
                }
            },
            "required": ["city"],
        },
    },
    {
        "type": "function",
        "name": "calculate",
        "description": "Evaluate a math expression and return the result.",
        "parameters": {
            "type": "object",
            "properties": {
                "expression": {
                    "type": "string",
                    "description": "A math expression to evaluate, e.g. '2 + 2'.",
                }
            },
            "required": ["expression"],
        },
    },
    {
        "type": "function",
        "name": "run_python_code",
        "description": "Write and execute a Python script, returning its stdout/stderr.",
        "parameters": {
            "type": "object",
            "properties": {
                "code": {
                    "type": "string",
                    "description": "The Python source code to execute.",
                }
            },
            "required": ["code"],
        },
    },
    {
        "type": "function",
        "name": "write_file",
        "description": "Write content to a file on disk.",
        "parameters": {
            "type": "object",
            "properties": {
                "file_path": {
                    "type": "string",
                    "description": "The path to write the file to.",
                },
                "content": {
                    "type": "string",
                    "description": "The content to write into the file.",
                },
            },
            "required": ["file_path", "content"],
        },
    },
]


async def send_event(ws, event: dict) -> None:
    await ws.send(json.dumps(event))


async def configure_session(ws) -> None:
    event = {
        "type": "session.update",
        "session": {
            "type": "realtime",
            "model": "gpt-realtime",
            "output_modalities": ["audio"],
            "instructions": (
                "You are a helpful AI assistant with access to tools. "
                "Use tools to accomplish tasks whenever possible. "
                "Speak clearly and briefly."
            ),
            "tools": TOOL_DEFINITIONS,
            "tool_choice": "auto",
            "audio": {
                "input": {
                    "format": {"type": "audio/pcm", "rate": 24000},
                    "transcription": {"model": "gpt-4o-transcribe"},
                    "turn_detection": {
                        "type": "server_vad",
                        "threshold": 0.5,
                        "prefix_padding_ms": 300,
                        "silence_duration_ms": 500,
                    },
                },
                "output": {
                    "format": {"type": "audio/pcm", "rate": 24000},
                },
            },
        },
    }
    await send_event(ws, event)
    print("Session configured.")


async def handle_function_call(ws, call_id: str, name: str, arguments: str) -> None:
    if not name:
        raise Exception("Did not get a function name")

    print(f"\n[Function Call] {name}({arguments})")
    tool_fn = TOOL_REGISTRY.get(name)
    if tool_fn is None:
        result = json.dumps({"error": f"Unknown function: {name}"})
    else:
        try:
            args = json.loads(arguments)
            result = tool_fn(**args)
            if asyncio.iscoroutine(result):
                result = await result
        except Exception as e:
            result = json.dumps({"error": str(e)})

    print(f"[Function Result] {result}")

    # Send the function call output back to the model.
    await send_event(ws, {
        "type": "conversation.item.create",
        "item": {
            "type": "function_call_output",
            "call_id": call_id,
            "output": result if isinstance(result, str) else json.dumps(result),
        },
    })

    # Trigger a new response so the model incorporates the function result.
    await send_event(ws, {"type": "response.create"})


def play_audio(output_stream: pyaudio.Stream, audio_output_queue: queue.Queue):
    """Runs in a separate thread because pyaudio's write() blocks until the
    sound card consumes the samples. Decoupling playback from the async event
    loop lets us flush the queue on interrupt without waiting for in-flight
    writes to finish."""
    while True:
        data = audio_output_queue.get()
        if data is None:
            break
        output_stream.write(data)


async def send_mic_audio(ws, mic) -> None:
    try:
        while True:
            raw_data = mic.read(CHUNK, exception_on_overflow=False)

            # Visual volume meter.
            audio_data = np.frombuffer(raw_data, dtype=np.int16).astype(np.float64)
            rms = np.sqrt(np.mean(audio_data**2))
            meter = int(min(rms / 50, 50))
            print(f"Mic Level: {'█' * meter}{' ' * (50 - meter)} |", end="\r")

            # Base64-encode and send audio chunk.
            b64_audio = base64.b64encode(raw_data).decode("utf-8")
            await send_event(ws, {
                "type": "input_audio_buffer.append",
                "audio": b64_audio,
            })

            await asyncio.sleep(0)
    except asyncio.CancelledError:
        pass


async def receive_events(ws, audio_output_queue: queue.Queue) -> None:
    # Accumulate function call arguments across delta events.
    pending_calls: dict[str, dict] = {}
    async for raw_message in ws:
        if DEBUG_WRITE_LOG:
            with open("data.jsonl", "a", encoding="utf-8") as f:
                f.write(json.dumps(raw_message) + "\n")

        event = json.loads(raw_message)
        event_type = event.get("type", "")

        if event_type == "session.created":
            print(raw_message)

        elif event_type == "session.updated":
            print(raw_message)

        elif event_type == "error":
            print(f"\n[Error] {event}")

        elif event_type == "input_audio_buffer.speech_started":
            # Flush queued AI audio so it doesn't talk over the user.
            while not audio_output_queue.empty():
                try:
                    audio_output_queue.get_nowait()
                except queue.Empty:
                    break

        elif event_type == "input_audio_buffer.speech_stopped":
            pass

        elif event_type == "input_audio_buffer.committed":
            pass

        elif event_type == "response.created":
            pass

        elif event_type == "response.output_text.delta":
            pass

        elif event_type == "response.output_text.done":
            pass

        # Audio output deltas - queue for playback.
        elif event_type == "response.output_audio.delta":
            audio_bytes = base64.b64decode(event.get("delta", ""))
            audio_output_queue.put(audio_bytes)

        elif event_type == "response.output_audio_transcript.delta":
            pass

        elif event_type == "response.output_audio_transcript.done":
            pass

        # Function call started - initialize pending call.
        elif event_type == "response.output_item.added":
            item = event.get("item", {})
            if item.get("type") == "function_call" and item.get("status") == "in_progress":
                item_id = item.get("id", "")
                pending_calls[item_id] = {
                    "call_id": item.get("call_id", ""),
                    "name": item.get("name", ""),
                    "arguments": "",
                }
                print(f"\n[Function Call Started] {item.get('name', '')}")

        # Function call argument deltas - accumulate.
        elif event_type == "response.function_call_arguments.delta":
            item_id = event.get("item_id", "")
            if item_id in pending_calls:
                pending_calls[item_id]["arguments"] += event.get("delta", "")

        elif event_type == "response.function_call_arguments.done":
            item_id = event.get("item_id", "")
            call_info = pending_calls.pop(item_id, None)
            if call_info is None:
                # Fallback: use data directly from the done event.
                call_info = {
                    "call_id": event.get("call_id"),
                    "name": event.get("name"),
                    "arguments": event.get("arguments"),
                }
            try:
                await handle_function_call(
                    ws,
                    call_info["call_id"],
                    call_info["name"],
                    call_info["arguments"],
                )
            except Exception as e:
                print(f"Failed to call function for message {call_info}: error - {e}")

        elif event_type == "response.done":
            pass

        elif event_type == "rate_limits.updated":
            pass

        else:
            print(f"\n[Event: {event_type}]")


async def main():
    if not OPENAI_API_KEY:
        print("Error: OPENAI_API_KEY environment variable not set")
        return

    p = pyaudio.PyAudio()

    input_device_index = int(p.get_default_input_device_info()['index'])
    output_device_index = int(p.get_default_output_device_info()['index'])

    # Channel count must match the device's capabilities or pyaudio errors upon open.
    input_info = p.get_device_info_by_index(input_device_index)
    output_info = p.get_device_info_by_index(output_device_index)
    input_channels = min(int(input_info['maxInputChannels']), 1)
    output_channels = min(int(output_info['maxOutputChannels']), 1)

    mic = p.open(
        format=FORMAT,
        channels=input_channels,
        rate=RATE,
        input=True,
        output=False,
        frames_per_buffer=CHUNK,
        input_device_index=input_device_index,
        start=False,
    )
    speaker = p.open(
        format=FORMAT,
        channels=output_channels,
        rate=RATE,
        input=False,
        output=True,
        frames_per_buffer=CHUNK,
        output_device_index=output_device_index,
        start=False,
    )
    mic.start_stream()
    speaker.start_stream()

    # Audio goes through a queue so we can flush it when the user interrupts.
    # Writing directly to the speaker makes it impossible to cancel in-flight audio.
    audio_output_queue = queue.Queue()
    threading.Thread(
        target=play_audio, args=(speaker, audio_output_queue), daemon=True
    ).start()

    headers = {
        "Authorization": f"Bearer {OPENAI_API_KEY}",
    }

    print("Connecting to OpenAI Realtime API...")

    async with websockets.connect(
        REALTIME_URL,
        additional_headers=headers,
    ) as ws:
        print("Connected! Configuring session...")
        await configure_session(ws)

        print("--- Session Active (Speak into mic) ---")

        mic_task = asyncio.create_task(send_mic_audio(ws, mic))
        try:
            await receive_events(ws, audio_output_queue)
        finally:
            mic_task.cancel()
            try:
                await mic_task
            except asyncio.CancelledError:
                pass

    # Cleanup
    audio_output_queue.put(None)  # Signal playback thread to exit.
    mic.close()
    speaker.close()
    p.terminate()
    print("\nSession ended.")


if __name__ == "__main__":
    asyncio.run(main())

Update the weave.init() call with your team and project names.
Set your OPENAI_API_KEY environment variable.
Start the voice assistant:
python weave_ws_voice_assistant.py

Get Started

Guides

Cookbooks

Reference

Details & Support

Integrate realtime traces

Run a realtime voice assistant with the OpenAI Agents SDK

Run a realtime voice assistant with WebSockets

​Integrate realtime traces

​Run a realtime voice assistant with the OpenAI Agents SDK

​Run a realtime voice assistant with WebSockets

Integrate realtime traces

Run a realtime voice assistant with the OpenAI Agents SDK

Run a realtime voice assistant with WebSockets