[Feature Request] VS Code extension: share MCP servers across conversation tabs instead of spawning per-tab

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single feature request

Problem

In the VS Code extension, each open conversation tab spawns its own claude.exe process, and each claude.exe spawns its own full set of MCP server processes.

With N conversation tabs and M configured MCP servers, this creates N × M processes.

Concrete example: 3 open conversations × 31 MCP servers = 93 child processes (node.exe, python.exe, native binaries). On a 16 GB RAM machine this causes:

• ~2-3 GB RAM consumed by duplicate MCP processes alone
• Windows Memory Compression kicks in (~700 MB)
• VS Code input field lag (keystrokes appear with delay)
• New conversation startup is slow (must spawn all MCP servers)

Verified experimentally:

• Each claude.exe has a different PID and different --resume session ID
• Each spawns its own MCP server processes (confirmed via ParentProcessId)
• When a conversation tab is closed, its claude.exe + MCP processes are correctly cleaned up
• The issue is purely about simultaneous tabs, not zombie processes

Proposed Solution

Share MCP server processes across conversation tabs within the same VS Code window. Since all tabs read the same .mcp.json config, the MCP servers could be managed by the extension host and multiplexed to individual claude.exe processes.

Alternatives Considered

• Reducing the number of MCP servers — works but limits functionality
• Using fewer simultaneous conversation tabs — works but limits workflow
• Lazy MCP loading (already implemented) — helps but once a server is called it stays alive per-conversation

Environment

• Claude Code: v2.1.81 (VS Code extension)
• OS: Windows 11, 16 GB RAM, Ryzen 5 5600H
• MCP servers: 31 configured

Related

• #27549 — High baseline CPU/RAM per claude.exe instance on Windows (compounds this issue)