diff --git a/ai-service/app/main.py b/ai-service/app/main.py
index c7890af..260dc61 100644
--- a/ai-service/app/main.py
+++ b/ai-service/app/main.py
@@ -1,9 +1,30 @@
-from fastapi import FastAPI, File, HTTPException, UploadFile
+import shutil
+import subprocess
+import threading
+import time
+
+from fastapi import FastAPI, File, HTTPException, Request, UploadFile
 
 from . import config
 from .detector import MockDetector, ONNXDetector
 
 app = FastAPI(title="Silk AI Service", version="0.1.0")
+START_TIME = time.time()
+_lock = threading.Lock()
+_requests = 0
+_latency_total = 0.0
+
+
+@app.middleware("http")
+async def count_requests(request: Request, call_next):
+    global _requests, _latency_total
+    start = time.time()
+    response = await call_next(request)
+    latency = (time.time() - start) * 1000
+    with _lock:
+        _requests += 1
+        _latency_total += latency
+    return response
 
 if config.MODEL_MODE == "onnx":
     detector = ONNXDetector(config.MODEL_PATH, tuple(config.MODEL_LABELS))
@@ -36,3 +57,66 @@ async def detect(file: UploadFile = File(...)):
             for d in detections
         ],
     }
+
+
+try:
+    import cv2  # noqa: E402
+except ImportError:
+    cv2 = None
+
+
+@app.post("/stream-detect")
+async def stream_detect(body: dict):
+    """摄像头流 AI 巡检骨架（#23）：拉流抽帧 → 检测。
+
+    二期功能：正式接入前需补摄像头视角数据与巡检任务编排。
+    """
+    url = (body or {}).get("url") or ""
+    if not url:
+        raise HTTPException(status_code=400, detail="缺少 url（rtsp/http-flv 流地址）")
+    if cv2 is None:
+        raise HTTPException(status_code=503, detail="OpenCV 未安装，无法拉流")
+    capture = cv2.VideoCapture(url)
+    if not capture.isOpened():
+        raise HTTPException(status_code=502, detail="无法连接视频流")
+    try:
+        frames = 0
+        detections = []
+        while frames < 3:
+            ok, frame = capture.read()
+            if not ok:
+                break
+            ok_encode, buf = cv2.imencode(".jpg", frame)
+            if ok_encode:
+                detections.extend(detector.detect(buf.tobytes()))
+            frames += 1
+    finally:
+        capture.release()
+    return {"model": config.MODEL_MODE, "frames": frames, "detections": detections[:10]}
+
+
+@app.get("/metrics")
+def metrics():
+    """AI 服务监控（#26 骨架）：请求量/延迟/GPU 信息（nvidia-smi 可选）。"""
+    with _lock:
+        reqs = _requests
+        avg = _latency_total / _requests if _requests else 0.0
+    gpu = None
+    if shutil.which("nvidia-smi"):
+        try:
+            out = subprocess.run(
+                ["nvidia-smi", "--query-gpu=name,temperature.gpu,memory.used,memory.total", "--format=csv,noheader,nounits"],
+                capture_output=True, text=True, timeout=10,
+            )
+            if out.returncode == 0 and out.stdout.strip():
+                parts = [p.strip() for p in out.stdout.strip().split(",")]
+                gpu = {"name": parts[0], "tempC": parts[1], "memUsedMB": parts[2], "memTotalMB": parts[3]}
+        except Exception:
+            gpu = None
+    return {
+        "model": config.MODEL_MODE,
+        "uptimeSeconds": int(time.time() - START_TIME),
+        "requests": reqs,
+        "avgLatencyMs": round(avg, 2),
+        "gpu": gpu,
+    }
diff --git a/ai-service/requirements.txt b/ai-service/requirements.txt
index c396cdf..3994f12 100644
--- a/ai-service/requirements.txt
+++ b/ai-service/requirements.txt
@@ -7,3 +7,4 @@ pydantic>=2.5
 python-multipart>=0.0.9
 pytest>=8.0
 httpx>=0.27
+opencv-python-headless>=4.8
diff --git a/ai-service/tests/test_api.py b/ai-service/tests/test_api.py
index f5b0b35..9b8e7c7 100644
--- a/ai-service/tests/test_api.py
+++ b/ai-service/tests/test_api.py
@@ -40,3 +40,16 @@ def test_detect_empty_file_rejected():
 def test_detect_invalid_image_rejected():
     r = client.post("/detect", files={"file": ("a.png", b"junk", "image/png")})
     assert r.status_code == 400
+
+
+def test_stream_detect_missing_url():
+    r = client.post("/stream-detect", json={})
+    assert r.status_code == 400
+
+
+def test_metrics_shape():
+    r = client.get("/metrics")
+    assert r.status_code == 200
+    body = r.json()
+    for key in ("model", "uptimeSeconds", "requests", "avgLatencyMs", "gpu"):
+        assert key in body