"Run a new experiment with lr=0.001, compare it to my last 5 runs, 
register the best one, and deploy it to staging."

╔══════════════════════════════════════════════════════════╗
║        mlctl — ML Platform Agent                         ║
╚══════════════════════════════════════════════════════════╝

USER: Compare all experiments, find the best one, register 
      it as netflix_recommender and deploy to staging.

🔧 Tool Call: list_experiments
   Result: [run_001 (acc: 0.882), run_002 (acc: 0.910), run_003 (acc: 0.925)]

🔧 Tool Call: compare_runs
   Result: Best run is run_003 with accuracy 0.9247

🔧 Tool Call: register_model
   Args: { run_id: run_003, model_name: netflix_recommender, stage: staging }
   Result: { registered: true, version: 3.0 }

🔧 Tool Call: deploy_model
   Args: { model_name: netflix_recommender, version: 3.0, environment: staging }
   Result: { deployed: true, endpoint: https://ml-platform.netflix.internal/... }

🤖 mlctl: Done. I compared all 3 experiments and identified run_003 
  (attention_recommender, accuracy: 92.47%) as the best performer. 
  It has been registered as netflix_recommender v3.0 and is now 
  live in staging at the endpoint above. Ready to promote to production 
  when you give the go-ahead.

┌─────────────────────────────────────────────────────┐
│                  MCP Interface                       │
│   Claude Desktop · VS Code · Slack · CLI             │
└──────────────────────┬──────────────────────────────┘
                       │
┌──────────────────────▼──────────────────────────────┐
│              MLOrchestrator (agent)                  │
│   Multi-step reasoning loop · Tool dispatch          │
│   Conversation history · Anomaly flagging            │
└──────┬──────────────────────────────────────────────┘
       │
       ├── ExperimentTools   → run, list, compare runs
       ├── ModelTools        → register, deploy, rollback
       └── PipelineTools     → trigger, monitor pipelines
                       │
┌──────────────────────▼──────────────────────────────┐
│              Model Adapter (swappable)               │
│   OpenAI (demo) · Netflix Internal Model (prod)      │
└──────────────────────┬──────────────────────────────┘
                       │
┌──────────────────────▼──────────────────────────────┐
│              Platform Backend                        │
│   Mock (demo) → Netflix AI Platform APIs (prod)      │
└─────────────────────────────────────────────────────┘

# 1. Clone
git clone https://github.com/ssengupta93/Agents
cd Agents/mlctl

# 2. Install dependencies
pip install -e ".[dev]"

# 3. Set your API key
export OPENAI_API_KEY=your-key-here

# 4. Run the demo
python examples/demo.py

{
  "mcpServers": {
    "mlctl": {
      "command": "python",
      "args": ["/path/to/mlctl/server.py"],
      "env": {
        "OPENAI_API_KEY": "your-key-here",
        "MODEL_PROVIDER": "openai"
      }
    }
  }
}

# mlctl/adapters/model_adapter.py

class NetflixModelAdapter(BaseModelAdapter):
    def __init__(self, endpoint: str, api_key: str):
        self.endpoint = endpoint
        self.api_key = api_key

    def chat(self, messages: list[dict], tools: list[dict] = None) -> dict:
        # Replace with Netflix's internal LLM SDK call
        raise NotImplementedError("Inject Netflix's internal model client here.")

mlctl/
├── server.py                  # MCP server entry point
├── mlctl/
│   ├── agent/
│   │   └── orchestrator.py    # Multi-step reasoning loop
│   ├── tools/
│   │   ├── experiments.py     # Experiment management tools
│   │   ├── models.py          # Model registry + deployment tools
│   │   └── pipelines.py       # Pipeline orchestration tools
│   └── adapters/
│       └── model_adapter.py   # Swappable LLM interface
├── mock/
│   └── platform_mock.py       # Simulated Netflix platform APIs
└── examples/
    └── demo.py                # Full lifecycle demo