Feature request: Lazy-load MCP servers on demand to reduce context usage

Problem

MCP tool definitions consume significant context tokens (~50K+ in my case) even when the tools aren't being used in the current conversation. This eats into the available context window before the conversation even starts.

From /doctor diagnostics:

⚠ Large MCP tools context (~56,813 tokens > 25,000)
  MCP servers:
    claude_ai_Notion: 14 tools (~12,095 tokens)
    claude_ai_Google_Calendar: 9 tools (~8,879 tokens)
    claude_ai_HubSpot: 8 tools (~8,438 tokens)
    claude_ai_Slack: 13 tools (~7,432 tokens)
    sales-mcp: 16 tools (~6,674 tokens)
    (7 more servers)

Proposed Solution

Allow MCP servers to be registered but not connected by default. Claude would:

1. See a lightweight list of available MCP servers and their descriptions (minimal token cost)
2. Auto-connect to a server when the conversation requires its tools
3. Optionally auto-disconnect after the tools are no longer needed

This would keep context lean while maintaining full tool access on demand.

Current Workaround

Manually toggling servers on/off via /mcp each session, which works but adds friction and requires the user to know in advance which tools they'll need.

Impact

• On 1M context models: ~5.7% context consumed by idle MCP tools
• On 200K context models: ~28% consumed — significant impact on conversation quality
• Users with many MCP integrations (sales stack, productivity tools, dev tools) are hit hardest