Feat: Implement MCP Adapter

[goh-chunlin]

Objective

Enable SilverAI to be used as an MCP Server, allowing LLM clients (Claude Desktop, Cursor, etc.) to discover and use protected tools directly.

Tasks

• Add mcp dependency to pyproject.toml;
• Create src/silver_ai/mcp.py;
• Implement an MCPAdapter class that:
  • Wraps @guard functions;
  • Exposes them as MCP Tools (JSON Schema);
  • Handles the "Zero-Crash" return types for the LLM.
• Create a server.py entry point to run the stdio server;
• Update README with MCP configuration instructions.

Acceptance Criteria

• A user can configure claude_desktop_config.json to point to SilverAi;
• Claude can "see" the IndustrialRobot tools;
• When Claude tries to execute an unsafe action, it receives the SilverAi JSON error payload.

graph TD
    subgraph Agentic Loop [Strands Agent Loop]
        direction TB
        Agent[🤖 Agent Brain]
        Model[🧠 LLM Model]
        
        %% The Thinking Loop
        Agent -->|1. Prompt| Model
        Model -->|2. Reasoning & Tool Selection| Agent
        
        %% The Action Loop (Where SilverAi Lives)
        Agent -->|3. Attempt Tool Call| Firewall{🛡️ SilverAi Guard}
        
        %% SilverAi Logic
        Firewall -->|❌ BLOCK: Unsafe State| Error[⚠️ Structured Error]
        Firewall -->|✅ PASS: Safe State| Tool[🔧 Real Tool / Hardware]
        
        %% The Feedback Loop
        Tool -->|4. Success Result| Agent
        Error -->|4. Failure Feedback| Agent
    end

    style Firewall fill:#ff4d4d,stroke:#333,stroke-width:2px,color:white
    style Agent fill:#00aaff,stroke:#333,stroke-width:2px,color:white
    style Tool fill:#00aaff,stroke:#333,stroke-width:2px,color:white
    style Model fill:#00aaff,stroke:#333,stroke-width:2px,color:white

Loading

[goh-chunlin]

Food for thoughts:

An MCP server is just a list of Python functions.

@mcp.tool()
def clean_floor(zone: str):
    # Logic
    return "Cleaning"

The SilverAi wrapper:

@mcp.tool()
@guard(rules.BatteryMin(15))
def clean_floor(zone: str):
    # Logic
    return "Cleaning"

Things we need to think:

• Verify Compatibility: Does @mcp.tool play nice with @guard because decorators order matters?
• Schema Injection: MCP reads the function signature to tell the LLM what args it needs. Does @guard hide the arguments? We may need to use funtools.wraps correctly;
• Error Handling: If SilverAi blocks it, does MCP crash? Or does it return the error string to the LLM?

We may need to build an mcp_demo.py that proves these two decorators (mcp and guard) work together.