diff --git a/scripts/collect-pool-metrics.sh b/scripts/collect-pool-metrics.sh
new file mode 100755
index 0000000..4fd10e3
--- /dev/null
+++ b/scripts/collect-pool-metrics.sh
@@ -0,0 +1,233 @@
+#!/bin/bash
+# Collect pool metrics during test execution
+# Runs in background and samples CPU, memory, latency from /cluster and /pools
+#
+# Usage: scripts/collect-pool-metrics.sh start|stop|report [RESULTS_DIR]
+#
+# Metrics collected:
+#   - Cluster load average (1m, 5m, 15m)
+#   - Memory usage (used_gb, available_gb, used_percent)
+#   - Pool availability (available, allocated, spawning)
+#   - API latency (ms) - time to hit /languages endpoint
+#   - Per-pool breakdown
+
+set -e
+
+ACTION="${1:-start}"
+RESULTS_DIR="${2:-pool-metrics}"
+METRICS_FILE="$RESULTS_DIR/pool-metrics.json"
+PID_FILE="$RESULTS_DIR/.collector.pid"
+RAW_FILE="$RESULTS_DIR/raw-samples.csv"
+
+API_BASE="https://api.unsandbox.com"
+
+mkdir -p "$RESULTS_DIR"
+
+get_auth_header() {
+    if [ -n "$UNSANDBOX_PUBLIC_KEY" ]; then
+        echo "Authorization: Bearer $UNSANDBOX_PUBLIC_KEY"
+    else
+        echo "Authorization: Bearer none"
+    fi
+}
+
+# Background collector loop
+collector_loop() {
+    echo "timestamp,load1,load5,load15,mem_used_gb,mem_avail_gb,mem_pct,available,allocated,spawning,latency_ms,http_code" > "$RAW_FILE"
+
+    while true; do
+        ts=$(date +%s)
+
+        # Get cluster metrics
+        start_ns=$(date +%s%N)
+        cluster=$(curl -s "$API_BASE/cluster" -H "$(get_auth_header)" 2>/dev/null || echo "{}")
+        end_ns=$(date +%s%N)
+        latency_ms=$(( (end_ns - start_ns) / 1000000 ))
+
+        # Parse cluster data
+        load1=$(echo "$cluster" | jq -r '.load_avg.load1 // 0')
+        load5=$(echo "$cluster" | jq -r '.load_avg.load5 // 0')
+        load15=$(echo "$cluster" | jq -r '.load_avg.load15 // 0')
+        mem_used=$(echo "$cluster" | jq -r '.memory.used_gb // 0')
+        mem_avail=$(echo "$cluster" | jq -r '.memory.available_gb // 0')
+        mem_pct=$(echo "$cluster" | jq -r '.memory.used_percent // 0')
+        available=$(echo "$cluster" | jq -r '.available // 0')
+        allocated=$(echo "$cluster" | jq -r '.allocated // 0')
+        spawning=$(echo "$cluster" | jq -r '.spawning // 0')
+
+        # Determine HTTP status
+        if [ -n "$cluster" ] && [ "$cluster" != "{}" ]; then
+            http_code="200"
+        else
+            http_code="000"
+        fi
+
+        echo "$ts,$load1,$load5,$load15,$mem_used,$mem_avail,$mem_pct,$available,$allocated,$spawning,$latency_ms,$http_code" >> "$RAW_FILE"
+
+        sleep 5  # Sample every 5 seconds
+    done
+}
+
+generate_report() {
+    if [ ! -f "$RAW_FILE" ]; then
+        echo "No metrics collected yet"
+        exit 1
+    fi
+
+    # Get pools data for the report
+    pools=$(curl -s "$API_BASE/pools" -H "$(get_auth_header)" 2>/dev/null || echo "{}")
+
+    # Parse raw samples and generate stats
+    python3 - "$RAW_FILE" "$METRICS_FILE" "$pools" << 'PYTHON'
+import sys
+import json
+import statistics
+
+raw_file = sys.argv[1]
+output_file = sys.argv[2]
+pools_json = sys.argv[3] if len(sys.argv) > 3 else "{}"
+
+try:
+    pools_data = json.loads(pools_json)
+except:
+    pools_data = {}
+
+samples = {
+    "load1": [], "load5": [], "load15": [],
+    "mem_used_gb": [], "mem_avail_gb": [], "mem_pct": [],
+    "available": [], "allocated": [], "spawning": [],
+    "latency_ms": []
+}
+timestamps = []
+http_codes = {}
+
+with open(raw_file) as f:
+    next(f)  # skip header
+    for line in f:
+        parts = line.strip().split(',')
+        if len(parts) >= 12:
+            timestamps.append(int(parts[0]))
+            samples["load1"].append(float(parts[1]))
+            samples["load5"].append(float(parts[2]))
+            samples["load15"].append(float(parts[3]))
+            samples["mem_used_gb"].append(float(parts[4]))
+            samples["mem_avail_gb"].append(float(parts[5]))
+            samples["mem_pct"].append(float(parts[6]))
+            samples["available"].append(int(parts[7]))
+            samples["allocated"].append(int(parts[8]))
+            samples["spawning"].append(int(parts[9]))
+            samples["latency_ms"].append(int(parts[10]))
+            code = parts[11]
+            http_codes[code] = http_codes.get(code, 0) + 1
+
+def calc_stats(values):
+    if not values:
+        return {"min": 0, "max": 0, "avg": 0, "median": 0, "stdev": 0}
+    return {
+        "min": round(min(values), 2),
+        "max": round(max(values), 2),
+        "avg": round(statistics.mean(values), 2),
+        "median": round(statistics.median(values), 2),
+        "stdev": round(statistics.stdev(values), 2) if len(values) > 1 else 0
+    }
+
+duration = (max(timestamps) - min(timestamps)) if timestamps else 0
+
+report = {
+    "summary": {
+        "total_samples": len(timestamps),
+        "collection_duration_seconds": duration,
+        "success_rate": round(http_codes.get("200", 0) / len(timestamps) * 100, 1) if timestamps else 0
+    },
+    "cpu": {
+        "load1": calc_stats(samples["load1"]),
+        "load5": calc_stats(samples["load5"]),
+        "load15": calc_stats(samples["load15"])
+    },
+    "memory": {
+        "used_gb": calc_stats(samples["mem_used_gb"]),
+        "available_gb": calc_stats(samples["mem_avail_gb"]),
+        "used_percent": calc_stats(samples["mem_pct"])
+    },
+    "pool": {
+        "available": calc_stats(samples["available"]),
+        "allocated": calc_stats(samples["allocated"]),
+        "spawning": calc_stats(samples["spawning"])
+    },
+    "latency_ms": calc_stats(samples["latency_ms"]),
+    "http_codes": http_codes
+}
+
+# Add pool breakdown if available
+if pools_data and "pools" in pools_data:
+    report["pools"] = []
+    for pool in pools_data["pools"]:
+        report["pools"].append({
+            "id": pool.get("id"),
+            "host": pool.get("host"),
+            "status": pool.get("status"),
+            "vcpu_count": pool.get("vcpu_count"),
+            "load_avg": pool.get("load_avg"),
+            "available": pool.get("available"),
+            "total": pool.get("total")
+        })
+
+with open(output_file, 'w') as f:
+    json.dump(report, f, indent=2)
+
+print(json.dumps(report, indent=2))
+PYTHON
+}
+
+case "$ACTION" in
+    start)
+        if [ -f "$PID_FILE" ] && kill -0 $(cat "$PID_FILE") 2>/dev/null; then
+            echo "Collector already running (PID: $(cat $PID_FILE))"
+            exit 0
+        fi
+        echo "Starting pool metrics collector..."
+        collector_loop &
+        echo $! > "$PID_FILE"
+        echo "Collector started (PID: $(cat $PID_FILE))"
+        echo "Samples: $RAW_FILE"
+        ;;
+
+    stop)
+        if [ -f "$PID_FILE" ]; then
+            pid=$(cat "$PID_FILE")
+            kill "$pid" 2>/dev/null || true
+            rm -f "$PID_FILE"
+            echo "Collector stopped"
+        else
+            echo "Collector not running"
+        fi
+        ;;
+
+    report)
+        generate_report
+        ;;
+
+    status)
+        if [ -f "$PID_FILE" ] && kill -0 $(cat "$PID_FILE") 2>/dev/null; then
+            echo "Collector running (PID: $(cat $PID_FILE))"
+            wc -l "$RAW_FILE" 2>/dev/null || echo "No samples yet"
+        else
+            echo "Collector not running"
+        fi
+        ;;
+
+    snapshot)
+        # One-shot snapshot of current state
+        curl -s "$API_BASE/cluster" -H "$(get_auth_header)" | jq '{
+            load_avg,
+            memory: {used_gb: .memory.used_gb, available_gb: .memory.available_gb, used_percent: .memory.used_percent},
+            containers: {available, allocated, spawning},
+            version
+        }'
+        ;;
+
+    *)
+        echo "Usage: $0 start|stop|report|status|snapshot [RESULTS_DIR]"
+        exit 1
+        ;;
+esac