[Feature Request] Expose MCP Server Toggle as a Claude-Invokable Tool for Skills

Problem

Currently, MCP server toggling via /mcp is UI-only. While #10447 requests CLI commands for hook automation (external), there's no way for Claude itself to programmatically toggle MCP servers during a conversation.

This prevents skills from dynamically managing MCP server lifecycle based on context.

Proposed Solution

Expose an internal tool that Claude can invoke:

McpServerToggle(server_name: string, enabled: bool) -> { success: bool, server: string, enabled: bool }

Or similar, mirroring what the /mcp UI does internally.

Use Case: Context-Aware Skills

A skill could intelligently enable/disable MCP servers based on the task:

# Example: database-admin.md skill

When activated, this skill:
1. Enables the `postgres-mcp` server
2. Performs database operations
3. Disables `postgres-mcp` when done to free context

This allows skills to be self-contained - they bring their own MCP dependencies online only when needed.

Why This Differs from #10447

| #10447 | This Request |
|--------|--------------|
| CLI commands for hooks/scripts | Tool Claude invokes directly |
| External automation | Internal tool call |
| Triggered by hook events | Triggered by Claude/skill logic |

Both are valuable and complementary. #10447 enables automation at session boundaries; this enables dynamic toggling mid-conversation based on Claude's reasoning.

Benefits

• Reduced context usage - MCP servers enabled only when needed
• Self-contained skills - Skills manage their own dependencies
• Dynamic workflows - Claude can adapt available tools to the task
• Composability - Chain skills that require different MCP servers

Implementation Notes

• Should respect existing approval workflows and enterprise policies
• Session-scoped only (no config file changes)
• Could log toggles for auditability