diff --git a/internal/upstream/managed/client.go b/internal/upstream/managed/client.go
index ea1e0597e..21eb9ea58 100644
--- a/internal/upstream/managed/client.go
+++ b/internal/upstream/managed/client.go
@@ -787,8 +787,19 @@ func (mc *Client) onStateChange(oldState, newState types.ConnectionState, info *
 	}
 }
 
-// startBackgroundMonitoring starts monitoring the connection health
+// startBackgroundMonitoring starts monitoring the connection health.
+//
+// Idempotent: tryReconnect() re-enters Connect() after tearing down only the
+// core client, so this is reached again on every reconnect while the existing
+// monitor is still running (stopBackgroundMonitoring is only called from the
+// managed client's Disconnect()). Without this guard each reconnect leaked
+// another backgroundHealthCheck goroutine, multiplying the ListTools liveness
+// probes sent to the upstream server for the rest of the process's life.
 func (mc *Client) startBackgroundMonitoring() {
+	if mc.monitoringStarted {
+		return
+	}
+
 	// Mark that monitoring has been started
 	mc.monitoringStarted = true
 	mc.monitoringWG.Add(1)
diff --git a/internal/upstream/managed/monitoring_leak_test.go b/internal/upstream/managed/monitoring_leak_test.go
new file mode 100644
index 000000000..901e0b4c0
--- /dev/null
+++ b/internal/upstream/managed/monitoring_leak_test.go
@@ -0,0 +1,70 @@
+package managed
+
+import (
+	"runtime"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+	"go.uber.org/zap"
+
+	"github.com/smart-mcp-proxy/mcpproxy-go/internal/config"
+)
+
+func newTestClientForMonitoring(t *testing.T) *Client {
+	t.Helper()
+	mc := &Client{
+		logger:         zap.NewNop(),
+		stopMonitoring: make(chan struct{}),
+	}
+	mc.SetConfig(&config.ServerConfig{Name: "monitor-server"})
+	return mc
+}
+
+// settledGoroutines polls runtime.NumGoroutine until it stops changing, so the
+// assertions below compare stable counts rather than racing the scheduler.
+func settledGoroutines() int {
+	prev := -1
+	for i := 0; i < 100; i++ {
+		runtime.Gosched()
+		time.Sleep(2 * time.Millisecond)
+		n := runtime.NumGoroutine()
+		if n == prev {
+			return n
+		}
+		prev = n
+	}
+	return runtime.NumGoroutine()
+}
+
+// TestStartBackgroundMonitoring_IsIdempotent pins the fix for a goroutine leak
+// that made an upstream server's health check fire N times per interval.
+//
+// startBackgroundMonitoring is reached from Connect(), and tryReconnect()
+// re-enters Connect() after tearing down only the *core* client — it never
+// calls the managed Client's Disconnect(), which is the sole caller of
+// stopBackgroundMonitoring(). So every reconnect used to add another
+// backgroundHealthCheck goroutine that lived until the process exited.
+//
+// Each of those goroutines probes liveness with ListTools(), so a server whose
+// tool catalog is large turns the leak into sustained traffic: an installation
+// observed here had ~16 concurrent tickers for a single configured upstream,
+// pushing ~80 kB per probe and accumulating 72 GB of audit rows downstream.
+func TestStartBackgroundMonitoring_IsIdempotent(t *testing.T) {
+	mc := newTestClientForMonitoring(t)
+
+	base := settledGoroutines()
+
+	// Simulates Connect() → reconnect → reconnect on one managed client.
+	mc.startBackgroundMonitoring()
+	mc.startBackgroundMonitoring()
+	mc.startBackgroundMonitoring()
+
+	assert.Equal(t, 1, settledGoroutines()-base,
+		"repeated startBackgroundMonitoring must run exactly one health-check goroutine")
+
+	mc.stopBackgroundMonitoring()
+
+	assert.Equal(t, base, settledGoroutines(),
+		"stopBackgroundMonitoring must leave no health-check goroutine behind")
+}