diff --git a/docs/mac_audio_dump.md b/docs/mac_audio_dump.md
new file mode 100644
index 00000000..533b2534
--- /dev/null
+++ b/docs/mac_audio_dump.md
@@ -0,0 +1,24 @@
+# macOS System Audio Capture with `SystemAudioDump`
+
+Glass uses a small native helper, `SystemAudioDump`, to stream macOS system audio into the Speech‑to‑Text (STT) pipeline. The helper taps the CoreAudio output device and writes raw PCM data to `stdout`, which the STT service ingests and forwards to the "Them" transcription session.
+
+## Launch sequence
+1. **Renderer request** – Starting capture in the UI triggers the IPC method `listen:startMacosSystemAudio`.
+2. **Service preparation** – The `SttService` kills any lingering helpers via `pkill -f SystemAudioDump` and resolves the path to the bundled binary at `src/ui/assets/SystemAudioDump` (or the unpacked equivalent in production).
+3. **Spawning** – The helper is spawned with `stdio` wired so that the service can read PCM bytes from `stdout` and log diagnostics from `stderr`.
+
+## Data flow
+- The helper emits 24 kHz stereo, 16‑bit PCM. The service buffers this stream into 100 ms (`CHUNK_SIZE = 24000 * 2 bytes * 2 channels * 0.1s`) chunks.
+- Each chunk is converted to mono by discarding the right channel, then base64‑encoded.
+- The encoded audio is forwarded both to the renderer (`system-audio-data`) and to the active "Them" STT WebSocket session. Provider‑specific payloads are constructed as needed (e.g., Gemini expects `{ audio: { data, mimeType } }`, Deepgram receives raw buffers).
+
+## Lifecycle management
+- `stderr` output from `SystemAudioDump` is logged for troubleshooting.
+- When the helper exits or errors, the service clears its reference and stops sending audio.
+- Calling `stopMacOSAudioCapture` sends a `SIGTERM` to the helper, ensuring a clean shutdown.
+
+## Integration notes
+- On macOS, screen capture is obtained separately via `getDisplayMedia`; system audio never travels through the browser APIs.
+- Windows relies on `getDisplayMedia({ audio: true })` for loopback capture, while Linux currently disables system audio capture.
+
+`SystemAudioDump` enables reliable separation of microphone and system audio on macOS, letting Glass transcribe remote voices without echoing the user's own microphone input.