diff --git a/deploy/runpod/README.md b/deploy/runpod/README.md
new file mode 100644
index 0000000..7b57553
--- /dev/null
+++ b/deploy/runpod/README.md
@@ -0,0 +1,97 @@
+# Deploy OpenClaw Voice on RunPod
+
+Deploy OpenClaw Voice on RunPod for GPU-accelerated voice inference.
+
+## Quick Deploy
+
+### 1. Create a RunPod Account
+Sign up at [runpod.io](https://runpod.io)
+
+### 2. Deploy from Template
+
+Use our pre-built template:
+
+```
+Template ID: openclaw-voice
+Docker Image: ghcr.io/purple-horizons/openclaw-voice:latest
+```
+
+Or deploy manually:
+
+### 3. Manual Deployment
+
+**Create a new Pod:**
+- GPU: RTX 4090 recommended ($0.44/hr)
+- Docker Image: `ghcr.io/purple-horizons/openclaw-voice:latest`
+- Expose Port: 8765
+- Volume: 20GB (for model cache)
+
+**Environment Variables:**
+```
+OPENCLAW_STT_MODEL=large-v3-turbo
+OPENCLAW_STT_DEVICE=cuda
+OPENCLAW_REQUIRE_AUTH=true
+OPENCLAW_MASTER_KEY=<generate-a-secure-key>
+OPENAI_API_KEY=<your-openai-key>
+```
+
+### 4. Connect
+
+Once deployed, get your Pod's public URL:
+```
+wss://<pod-id>-8765.proxy.runpod.net/ws?api_key=<your-key>
+```
+
+## Cost Optimization
+
+### On-Demand (Pay per hour)
+- RTX 4090: ~$0.44/hr
+- RTX 3090: ~$0.30/hr
+- A10: ~$0.50/hr
+
+### Spot Instances (Up to 80% cheaper)
+- Use spot instances for non-critical workloads
+- May be interrupted with 30s notice
+
+### Reserved (Best for production)
+- Reserve a GPU for consistent pricing
+- No interruptions
+
+## Scaling
+
+### Single Pod
+- Handles ~10-20 concurrent voice sessions
+- Good for testing and small deployments
+
+### Multiple Pods with Load Balancer
+1. Deploy 2+ pods
+2. Use RunPod's serverless endpoints
+3. Or put behind Cloudflare Load Balancer
+
+## Monitoring
+
+### Health Check
+```bash
+curl https://<pod-url>/
+```
+
+### Check Logs
+```bash
+runpodctl logs <pod-id>
+```
+
+## Troubleshooting
+
+### GPU Not Detected
+- Ensure NVIDIA drivers are loaded
+- Check `nvidia-smi` in pod terminal
+
+### Out of Memory
+- Reduce batch size
+- Use smaller Whisper model (base vs large)
+- Upgrade to GPU with more VRAM
+
+### High Latency
+- Use larger GPU
+- Enable streaming responses
+- Check network latency to RunPod region
diff --git a/scripts/download_models.py b/scripts/download_models.py
new file mode 100644
index 0000000..7b4c9b1
--- /dev/null
+++ b/scripts/download_models.py
@@ -0,0 +1,84 @@
+#!/usr/bin/env python3
+"""
+Download Whisper models for offline use.
+
+Usage:
+    python scripts/download_models.py [model_name]
+    
+Models:
+    tiny    - 39M params, ~1GB VRAM (fastest)
+    base    - 74M params, ~1GB VRAM
+    small   - 244M params, ~2GB VRAM
+    medium  - 769M params, ~5GB VRAM
+    large-v3-turbo - 809M params, ~6GB VRAM (best quality/speed)
+"""
+
+import sys
+import os
+
+
+def download_model(model_name: str = "base"):
+    """Download a Whisper model."""
+    print(f"Downloading Whisper model: {model_name}")
+    print("This may take a few minutes...")
+    
+    try:
+        from faster_whisper import WhisperModel
+        
+        # This will download the model if not cached
+        model = WhisperModel(model_name, device="cpu", compute_type="int8")
+        
+        print(f"✅ Model '{model_name}' downloaded successfully!")
+        print(f"   Cached at: ~/.cache/huggingface/")
+        
+        # Test the model
+        import numpy as np
+        audio = np.zeros(16000, dtype=np.float32)
+        segments, info = model.transcribe(audio)
+        list(segments)  # Consume generator
+        
+        print(f"✅ Model tested successfully!")
+        
+    except ImportError:
+        print("❌ faster-whisper not installed. Run:")
+        print("   pip install faster-whisper")
+        sys.exit(1)
+    except Exception as e:
+        print(f"❌ Error: {e}")
+        sys.exit(1)
+
+
+def list_models():
+    """List available models."""
+    models = {
+        "tiny": "39M params, ~1GB VRAM, fastest",
+        "base": "74M params, ~1GB VRAM, good balance",
+        "small": "244M params, ~2GB VRAM",
+        "medium": "769M params, ~5GB VRAM",
+        "large-v3": "1.5B params, ~10GB VRAM, best quality",
+        "large-v3-turbo": "809M params, ~6GB VRAM, best quality/speed ratio",
+    }
+    
+    print("Available Whisper models:")
+    print()
+    for name, desc in models.items():
+        print(f"  {name:20} - {desc}")
+    print()
+    print("Recommended: large-v3-turbo (GPU) or base (CPU)")
+
+
+if __name__ == "__main__":
+    if len(sys.argv) < 2:
+        list_models()
+        print()
+        model = input("Enter model name to download (or 'q' to quit): ").strip()
+        if model.lower() == 'q':
+            sys.exit(0)
+    else:
+        model = sys.argv[1]
+        
+        if model in ["-h", "--help", "help"]:
+            list_models()
+            sys.exit(0)
+    
+    download_model(model)
diff --git a/scripts/generate_master_key.py b/scripts/generate_master_key.py
new file mode 100644
index 0000000..74601ea
--- /dev/null
+++ b/scripts/generate_master_key.py
@@ -0,0 +1,30 @@
+#!/usr/bin/env python3
+"""
+Generate a secure master API key for OpenClaw Voice.
+
+Usage:
+    python scripts/generate_master_key.py
+    
+Add the output to your .env file or deployment environment.
+"""
+
+import secrets
+
+
+def generate_master_key() -> str:
+    """Generate a secure master key."""
+    return f"ocv_master_{secrets.token_urlsafe(32)}"
+
+
+if __name__ == "__main__":
+    key = generate_master_key()
+    
+    print("=" * 60)
+    print("OpenClaw Voice Master Key")
+    print("=" * 60)
+    print()
+    print(f"OPENCLAW_MASTER_KEY={key}")
+    print()
+    print("Add this to your .env file or deployment environment.")
+    print("Keep this key SECRET - it has full admin access.")
+    print("=" * 60)