diff --git a/sendspin/serve/coordinator.py b/sendspin/serve/coordinator.py
index acdea45..a8bbe7c 100644
--- a/sendspin/serve/coordinator.py
+++ b/sendspin/serve/coordinator.py
@@ -15,6 +15,7 @@
 import time
 from contextlib import suppress
 
+from aiosendspin.clock import RawMonotonicClock
 from aiosendspin.server.push_stream import DEFAULT_INITIAL_DELAY_US
 
 from sendspin.serve import get_local_ip
@@ -35,24 +36,6 @@
 MAX_BUFFER_AHEAD_US = 5_000_000
 
 
-def server_now_us() -> int:
-    """Current time in microseconds on the aiosendspin server clock domain.
-
-    The aiosendspin server timestamps audio with ``RawMonotonicClock``
-    (``CLOCK_MONOTONIC_RAW`` on Linux), which — unlike
-    ``time.monotonic()`` — is not slewed by NTP/adjtime
-    (``aiosendspin/clock.py``). ``play_start_us`` is an absolute timestamp
-    consumed verbatim by workers/clients, so it must be sampled from the
-    same clock domain; ``time.monotonic()`` drifts away from it by every
-    NTP adjustment since boot (measured 102 s apart on one Linux host).
-    Interval-only measurements (stats/health timers) stay on
-    ``time.monotonic()``, which is the correct clock for durations.
-    """
-    if hasattr(time, "CLOCK_MONOTONIC_RAW"):
-        return time.clock_gettime_ns(time.CLOCK_MONOTONIC_RAW) // 1_000
-    return time.monotonic_ns() // 1_000
-
-
 class ServeCoordinator:
     """Orchestrates multi-worker serve mode."""
 
@@ -89,6 +72,8 @@ def __init__(
         self._run_task: asyncio.Task[int] | None = None
         self._failed_workers: set[int] = set()
         self._reported_crashed: set[int] = set()
+        # Workers schedule audio on this clock, so play_start_us must come from it too.
+        self._clock = RawMonotonicClock()
 
     async def run(self) -> int:
         """Main coordinator loop."""
@@ -264,7 +249,7 @@ async def _stream_audio_loop(self) -> None:
                 audio_source = await decode_audio(self.source, source_format=self.source_format)
                 fmt = audio_source.format
 
-                stream_start_us = server_now_us() + DEFAULT_INITIAL_DELAY_US
+                stream_start_us = self._clock.now_us() + DEFAULT_INITIAL_DELAY_US
                 play_start_us = stream_start_us
                 samples_sent = 0
 
@@ -301,7 +286,7 @@ async def _stream_audio_loop(self) -> None:
                     samples_sent += len(pcm_chunk) // frame_stride
                     play_start_us = stream_start_us + samples_sent * 1_000_000 // fmt.sample_rate
 
-                    now_us = server_now_us()
+                    now_us = self._clock.now_us()
                     ahead_us = play_start_us - DEFAULT_INITIAL_DELAY_US - now_us
                     if ahead_us > MAX_BUFFER_AHEAD_US:
                         await asyncio.sleep((ahead_us - MAX_BUFFER_AHEAD_US) / 1_000_000)
diff --git a/tests/serve/test_coordinator.py b/tests/serve/test_coordinator.py
index 47639a2..54447d9 100644
--- a/tests/serve/test_coordinator.py
+++ b/tests/serve/test_coordinator.py
@@ -6,7 +6,9 @@
 from unittest.mock import MagicMock, patch
 
 import pytest
+from aiosendspin.clock import RawMonotonicClock
 from aiosendspin.server import AudioFormat
+from aiosendspin.server.push_stream import DEFAULT_INITIAL_DELAY_US
 
 from sendspin.serve.coordinator import ServeCoordinator
 from sendspin.serve.ipc import AudioChunk, WorkerClientCount, WorkerError, WorkerListening
@@ -124,3 +126,24 @@ async def chunks() -> AsyncGenerator[bytes, None]:
         await coordinator._stream_audio_loop()
 
     assert sent[3].play_start_us - sent[0].play_start_us == 256_000
+
+
+@pytest.mark.asyncio
+async def test_chunk_timestamps_use_worker_clock(coordinator: ServeCoordinator) -> None:
+    """Chunks are stamped on the clock the workers schedule audio with."""
+    sent: list[AudioChunk] = []
+    coordinator._audio_queues = [MagicMock(put=sent.append)]
+    coordinator._total_listeners.value = 1
+
+    async def chunks() -> AsyncGenerator[bytes, None]:
+        yield bytes(4096 * 4)
+        coordinator._shutdown_requested = True
+
+    source = MagicMock(generator=chunks(), format=AudioFormat(48000, 16, 2))
+    with (
+        patch("sendspin.serve.coordinator.decode_audio", return_value=source),
+        patch.object(RawMonotonicClock, "now_us", return_value=1_000_000),
+    ):
+        await coordinator._stream_audio_loop()
+
+    assert sent[0].play_start_us == 1_000_000 + DEFAULT_INITIAL_DELAY_US