[Feature Request] Per-connector toggle for claude.ai-synced MCP servers in Claude Code

Bug Description
Feature request: per-connector toggle for claude.ai-synced MCP connectors in Claude Code

Problem: MCP connectors enabled on claude.ai web (Gmail, Google Calendar, Google Drive etc.) are auto-synced into every Claude Code CLI session. Their tool names are injected into the deferred-tools list at session start, and some (like Nia) inject full MCP instruction blocks — hundreds of tokens per session — even when the user never intends
to invoke them from CC.

Currently there's no way to scope these: disconnecting on claude.ai removes them everywhere, and settings.json has no disabledMcpServers or equivalent that matches mcp__claude_ai_*
patterns. The permissions system can block execution but doesn't reclaim the context tokens consumed by tool names and instruction blocks.

Proposed solutions (either works):

1. Settings-level exclusion list — support a disabledMcpServers array in ~/.claude/settings.json that accepts server name patterns (e.g. "mcp__claude_ai_Gmail",

"mcp__claude_ai_Google_*", "apigcp"). Applied before tool-name injection at session start.

1. Per-connector toggle on claude.ai — each connector gets a "Enable in Claude Code CLI" checkbox independent of its web-app availability.

Use case: I use Gmail/Calendar/Drive from claude.ai daily but never from Claude Code sessions on my dev machine. Right now I'm paying context overhead on every CC session for tools I'll never call there. For heavy users, this compounds across many sessions per day.

Environment Info

• Platform: linux
• Terminal: ghostty
• Version: 2.1.119
• Feedback ID: 8c2c31f9-349a-4ce7-912d-a6190251ea9a