Unused connectors and MCP servers still cost tokens in every session

In the desktop app's Code tab, a session that had received only "hi" was already at 56.8k tokens of context. Deferred tool schemas already help, but integrations a session never uses still cost tokens every time:

• Every deferred tool name is listed when the session starts. In my session that was 163 names, and 65 of them belonged to Claude in Chrome, computer use and a Google Drive connector that the session never used. Connector tool names start with a 36-character UUID (mcp__<uuid>__copy_file), which takes extra tokens.
• The MCP server instructions for Claude in Chrome and computer use (1.7k tokens) load even when neither is used.
• The visualize connector is listed twice: once as loaded mcp__visualize__* tools and again as UUID-prefixed deferred tools.
• Turning on a claude.ai connector for one session also makes it the default for new sessions, so using a connector occasionally means remembering to turn it off afterwards.
• The context panel's Messages row showed 11.9k tokens at the first message. By my estimate, the session-start reminders I could identify account for 4–5k of that, and the panel doesn't break down the rest.

Suggestions:

1. Show each unused server as one line (name, tool count, purpose), and load its tool names and instructions only when the model asks for that server.
2. Load a server's instructions at the same time as its tool schemas.
3. Use the server name in connector tool names in place of the UUID, and remove the duplicate visualize listing.
4. Allow turning on a connector for one session without changing the default for new sessions.
5. Break down the Messages row so session-start reminders are counted separately from the conversation.

Related: #13717, #47881