diff --git a/db_async.py b/db_async.py
index f2c28ab7..c7adf1dc 100644
--- a/db_async.py
+++ b/db_async.py
@@ -49,6 +49,11 @@ telemetry_events = Table(
     Column("cpu_pct", Float, nullable=True),
     Column("mem_handles", Integer, nullable=True),
     Column("latency_ms", Float, nullable=True),
+    # Server-side receive time. The `timestamp` column above is the CLIENT's
+    # self-reported wall clock and drifts up to ~90s across machines, so the
+    # "online" window must use this server-stamped value instead (see /live
+    # cache query). Nullable so pre-migration rows fall back to `timestamp`.
+    Column("received_at", DateTime(timezone=True), nullable=True),
 )
 # Composite index to accelerate Grafana queries filtering by character_name then ordering by timestamp
 Index(
@@ -256,6 +261,18 @@ async def init_db_async():
         print(
             f"Warning: failed to create composite index ix_telemetry_events_char_ts: {e}"
         )
+    # Add the server-receive-time column to existing deployments (idempotent).
+    # Used as the clock-skew-proof basis for the "online" window in /live.
+    try:
+        with engine.connect() as conn:
+            conn.execute(
+                text(
+                    "ALTER TABLE telemetry_events "
+                    "ADD COLUMN IF NOT EXISTS received_at TIMESTAMPTZ"
+                )
+            )
+    except Exception as e:
+        print(f"Warning: failed to add telemetry_events.received_at column: {e}")
     # Add retention and compression policies on the hypertable
     try:
         with engine.connect().execution_options(isolation_level="AUTOCOMMIT") as conn:
diff --git a/main.py b/main.py
index a3dd906e..2eaa2433 100644
--- a/main.py
+++ b/main.py
@@ -824,8 +824,16 @@ async def _refresh_cache_loop() -> None:
 
     while True:
         try:
-            # Recompute live players (last 30s)
+            # Recompute live players (last 30s) using the SERVER receive-time,
+            # not the client's self-reported timestamp. Game machines' clocks
+            # drift up to ~90s apart, which put characters on the 30s boundary
+            # and flapped the count. COALESCE falls back to the client timestamp
+            # for pre-migration rows (and the brief window after deploy before
+            # new rows carry received_at). The coarse `timestamp` bound is kept
+            # ONLY so TimescaleDB can prune chunks; it's wide enough (10 min) to
+            # cover any plausible clock skew.
             cutoff = datetime.now(timezone.utc) - ACTIVE_WINDOW
+            chunk_cutoff = datetime.now(timezone.utc) - timedelta(minutes=10)
             sql_live = """
                 SELECT sub.*,
                        COALESCE(rs.total_rares, 0)      AS total_rares,
@@ -834,7 +842,8 @@ async def _refresh_cache_loop() -> None:
                 FROM (
                     SELECT DISTINCT ON (character_name) *
                     FROM telemetry_events
-                    WHERE timestamp > :cutoff
+                    WHERE timestamp > :chunk_cutoff
+                      AND COALESCE(received_at, timestamp) > :cutoff
                     ORDER BY character_name, timestamp DESC
                 ) sub
                 LEFT JOIN rare_stats rs
@@ -848,7 +857,9 @@ async def _refresh_cache_loop() -> None:
 
             # Use a single connection for both queries to reduce connection churn
             async with database.connection() as conn:
-                rows = await conn.fetch_all(sql_live, {"cutoff": cutoff})
+                rows = await conn.fetch_all(
+                    sql_live, {"cutoff": cutoff, "chunk_cutoff": chunk_cutoff}
+                )
                 new_players = [dict(r) for r in rows]
 
                 # Track player changes for debugging
@@ -3125,6 +3136,11 @@ async def ws_receive_snapshots(
                     # Prepare data and compute kill delta
                     db_data = snap.dict()
                     db_data["rares_found"] = 0
+                    # Stamp the SERVER receive time. snap.timestamp is the
+                    # client's wall clock (drifts across machines); the /live
+                    # "online" window uses received_at so clock skew can't flap
+                    # the player count.
+                    db_data["received_at"] = datetime.now(timezone.utc)
                     key = (snap.session_id, snap.character_name)
 
                     # Get last recorded kill count for this session