---
description: MCP tool implementation pattern and structure
globs: ["src/tools/**/*.py"]
alwaysApply: true
---

# MCP Tool Implementation Pattern

Every MCP tool should follow this structure:

```python
from typing import Optional, Literal
from pydantic import BaseModel, Field
from fastmcp import FastMCP

mcp = FastMCP("UniFi MCP Server")

@mcp.tool()
async def tool_name(
    # Required parameters first
    site_id: str,
    resource_id: str,

    # Optional parameters with defaults
    optional_param: Optional[str] = None,

    # Safety parameters for mutating operations
    confirm: bool = False,
    dry_run: bool = False
) -> dict:
    """
    Clear, concise description of what the tool does.

    Args:
        site_id: Always include site_id for multi-site support
        resource_id: Specific resource identifier
        optional_param: Description of optional parameter
        confirm: REQUIRED True for mutating operations (safety)
        dry_run: Preview changes without applying them

    Returns:
        Dictionary with operation results. Specify structure.

    Raises:
        MCPError: For all MCP-specific errors
        ValueError: For invalid inputs
        UniFiConnectionError: For API connection issues
    """
    # 1. Input validation
    if not site_id:
        raise ValueError("site_id is required")

    # 2. Safety checks for mutating operations
    if not confirm and not dry_run:
        raise MCPError("This operation requires confirm=True")

    # 3. Audit logging
    logger.info(
        f"tool_name: site={site_id}, id={resource_id}, dry_run={dry_run}",
        extra={"operation": "tool_name", "site_id": site_id}
    )

    # 4. Dry-run mode
    if dry_run:
        preview = await generate_preview(site_id, resource_id)
        return {"status": "preview", "changes": preview}

    # 5. Execute operation
    try:
        result = await execute_operation(site_id, resource_id)
    except UniFiAPIError as e:
        logger.error(f"Operation failed: {e}")
        raise MCPError(f"Failed to execute operation: {str(e)}")

    # 6. Cache invalidation (if applicable)
    if result.get("success"):
        await cache.invalidate(f"resource:{site_id}")

    # 7. Return structured result
    return {
        "status": "success",
        "resource_id": resource_id,
        "data": result
    }
```

## MCP-Specific Guidelines

- Use `@mcp.tool()` decorator for tool definitions
- Use `@mcp.resource()` decorator for resource endpoints
- Always include `site_id` for multi-site support
- Implement proper error handling with custom exceptions
- Return structured dictionaries for consistency