AirBridge for Windows

AirBridge sends live Windows system audio—or one application's audio—to one or more AirPlay speakers. It runs from the system tray and adds per-speaker volume, multi-speaker alignment, silence standby, browser picture-delay correction, and an optional GPT-5.6 voice assistant.

Judge quick start

Supported platform: Windows 10 22H2 (build 19045) or newer, or Windows 11. Testing requires an AirPlay/RAOP receiver on the same trusted Private network. An OpenAI API key is optional; core audio streaming works without one.

1. Download AirBridge-Setup.exe from the latest GitHub release. No build tools are required.
2. Run the installer and launch AirBridge from the Start menu. The installer is not code-signed, so Windows SmartScreen may require More info → Run anyway.
3. Open the tray flyout, select a discovered speaker, and choose Start. A receiver starts at 30% volume the first time it is used, so raise its slider if the test audio is too quiet.
4. Play audio on Windows and confirm that it moves to the selected speaker. Adjust its volume from the flyout.
5. Test push-to-talk using the walkthrough below. This requires a microphone, internet access, and a judge-provided OpenAI API key.

For a second test path, start one speaker, open Settings → Browser sync, and choose Measure delay. Load the browser extension as described below, enter the measured delay, and enable it on a video site to delay the picture while leaving its audio playing normally.

Push-to-talk judge walkthrough

1. Open the tray flyout, choose the Settings gear, and select Assistant.
2. Paste an OpenAI API key, leave Enable the AirBridge assistant when an API key is available checked, and choose Save. The key is stored for the current Windows user in Windows Credential Manager and is never displayed again.
3. In Settings → General, confirm the push-to-talk shortcut. The default global shortcut is Ctrl+Alt+Space and it works while AirBridge is in the tray.
4. Hold Ctrl+Alt+Space for at least 250 ms. When the HUD says Listening, say “What speakers are available?” and release the shortcut. The HUD advances through Transcribing and Thinking, then shows the answer. This first command is read-only and verifies the microphone, transcription, GPT-5.6, and local tool path.
5. Use the exact receiver name returned by the assistant for an action, for example: “AirPlay all system audio to Kitchen,” “Set Kitchen volume to 50 percent,” or “Stop playing system audio.” AirBridge asks for confirmation when local policy requires it.
6. Open Settings → Advanced → Open AI Activity Inspector to verify the transcription event, gpt-5.6 Responses API request, local policy decision, tool call, result, latency, token usage, and cost estimate.

Press Escape while recording to cancel a command. If the HUD reports that the microphone is unavailable or no speech was detected, enable Windows microphone access for desktop apps, confirm that the default input is not muted, and try again. If the shortcut does not register, another application may be using it; click the shortcut field under Settings → General, press a different modified key combination, and save. See troubleshooting for receiver, streaming, and log guidance.

What it does

• Captures the full Windows mix or one application's process tree through WASAPI.
• Streams live, file-free PCM to independently controlled AirPlay/RAOP receivers.
• Routes to multiple speakers and compensates for constant receiver latency with acoustic measurement and per-speaker trims.
• Releases idle speaker sessions through silence standby, then reconnects the same route when audio resumes.
• Delays browser video frames to correct AirPlay lip sync without delaying the captured audio.
• Accepts optional voice commands through GPT-5.6 with local policy enforcement and an inspectable activity log.

AirBridge uses independent RAOP sessions, not AirPlay 2 buffered-mode multi-room synchronization. Receiver alignment corrects constant latency but does not provide a shared PTP playback clock.

How it works

Windows WASAPI loopback
  → 44.1 kHz s16le stereo normalization
  → shared bounded in-memory PCM fanout
  → per-receiver trim, named pipe, and RAOP session
  → AirPlay receiver

No temporary media file is created in the streaming path. Raw audio, receiver addresses, credentials, executable paths, and window titles are never sent to OpenAI. See architecture, hardware validation, AI tools and privacy, and troubleshooting for details.

Build from source

Requirements: the .NET 9 SDK, Python 3.12, Windows 10/11, and an AirPlay receiver on the same trusted Private network.

py -3.12 -m venv .venv
.\.venv\Scripts\python.exe -m pip install -r src\AirBridge.RaopHost\requirements.txt
dotnet run --project src\AirBridge.App\AirBridge.App.csproj

On first launch, refresh the output list, select one or more receivers, and choose Start. Apple TV pairing prompts for the code displayed on the TV. For a Mac receiver, set Allow AirPlay for to Anyone on the Same Network.

The larger developer diagnostics dashboard is available with:

dotnet run --project src\AirBridge.App\AirBridge.App.csproj -- --preview

Browser extension

• Chrome or Edge: load src/AirBridge.BrowserExtension as an unpacked Manifest V3 extension.
• Firefox: open about:debugging#/runtime/this-firefox, choose Load Temporary Add-on, and select src/AirBridge.FirefoxExtension/manifest.json.

Start exactly one speaker, use Measure delay in AirBridge, enter the result in the extension popup, and enable the current video site. Protected DRM/EME video is restored and skipped when frame capture is unavailable. More detail is in browser sync.

Optional GPT-5.6 assistant

Follow the push-to-talk judge walkthrough for the shortest end-to-end test. OPENAI_API_KEY can be used instead of the Settings field as a managed override.

The assistant uses gpt-5.6 through the Responses API for reasoning, intent resolution, diagnosis, and function calling. gpt-4o-transcribe handles user-approved push-to-talk transcription. A local allowlist independently classifies each tool call as read-only, reversible, confirmation-required, or forbidden.

Open Settings → Advanced → Open AI Activity Inspector to view transcription events, Responses API requests, policy decisions, tool calls, results, latency, token usage, and cost estimates. The model never receives raw audio, credentials, or receiver network details.

Built with Codex and GPT-5.6

AirBridge was built during OpenAI Build Week in July 2026 with Codex as the primary implementation partner across C#, Python, and JavaScript.

Human direction supplied the product requirements, receiver environment, network-permission diagnosis, acoustic observations, and final product, safety, and scope decisions. Codex accelerated the work by:

• tracing pyatv's RAOP internals to find the source-opening seam for a permanently open, file-free AudioSource;
• designing and debugging the shared sender clock, bounded receiver queues, readiness gate, and acoustic alignment flow;
• running specification and implementation audits to identify drift between intended and implemented behavior;
• writing focused .NET, Python, and Node regressions for stalled receivers, standby restarts, and calibration feedback loops;
• iterating on the tray UI with rendered snapshots and analyzing real HomePod, Apple TV, and Mac telemetry; and
• building the MSI, portable package, and AirBridge-Setup.exe release pipeline.

GPT-5.6 contributed directly to the final product as the optional in-app assistant described above. The key engineering decisions remained local and deterministic: audio stays in bounded memory, a stalled receiver cannot block its siblings, dangerous actions are excluded or require confirmation, and assistant fixes are verified against real telemetry. The specification audit records additional implementation checks.

Tests and packaging

dotnet test AirBridge.sln -c Release
.\.venv\Scripts\python.exe -m unittest discover -s src\AirBridge.RaopHost -p "test_*.py" -v
node --test tests\browser-extension.test.js
.\scripts\package.ps1

Machine-level WASAPI tests are opt-in so ordinary CI cannot hang on a driver COM call:

$env:AIRBRIDGE_RUN_HARDWARE_TESTS = "1"
dotnet test AirBridge.sln -c Release --filter "Category=Hardware"

Packaging produces artifacts\AirBridge-Setup.exe, artifacts\AirBridge.msi, and a portable application under artifacts\publish.

Acknowledgments

AirBridge streams RAOP audio through pyatv, Pierre Ståhl's excellent open-source Apple TV and AirPlay library for Python. Its clean protocol implementation is what made a file-free live AudioSource on Windows possible.

License

AirBridge is available under the MIT License.