Problem

Plugins (especially knowledge-work-plugins) bundle multiple MCP servers, but there's no way to selectively disable individual MCP servers within an enabled plugin. It's all-or-nothing at the plugin level.

This creates two problems:

1. Massive MCP server duplication across plugins

When multiple knowledge-work plugins are enabled, the same MCP servers load multiple times. For example, with just engineering, design, and data enabled:

The slack MCP server alone is declared in 12+ knowledge-work plugins.

2. Unnecessary context window consumption

Each MCP server injects tool definitions into the deferred tools list, consuming context space on every turn — even if the user never calls those tools. Users who want a plugin's skills/agents but not all of its MCP servers have no way to reduce this overhead.

3. Disabling MCP servers doesn't persist

Users report disabling MCP servers during a session, only to have them reactivate on the next session because the plugin re-injects them from its .mcp.json manifest.

Proposed Solution

Add per-MCP-server toggles within plugins. For example, in settings.json:

{
  "enabledPlugins": {
    "engineering@knowledge-work-plugins": {
      "enabled": true,
      "disabledMcpServers": ["pagerduty", "datadog", "slack"]
    }
  }
}

Or alternatively, a global MCP server blocklist:

{
  "disabledMcpServers": ["pagerduty", "datadog", "intercom", "hex"]
}

Additional considerations

• Deduplication: When multiple plugins declare the same MCP server (e.g., slack), only one instance should load. Currently each plugin loads its own copy.
• Hooks are insufficient: A PreToolUse hook can block MCP tool execution, but the tool definitions still load into context, so the context overhead remains.

Environment

• Claude Code CLI (macOS)
• Multiple knowledge-work-plugins + standalone plugins enabled simultaneously

🤖 Generated with Claude Code