Problem

MCP servers disconnect intermittently during sessions (every few runs/commands). Currently, the only way to reconnect is manually running /mcp, which interrupts workflow.

Current Behavior

• MCP servers disconnect without warning during active sessions
• No automatic reconnection attempt is made
• User must manually run /mcp to reconnect
• This is disruptive when relying on MCP tools (postgres, github, etc.)

Desired Behavior

1. Auto-reconnect: When an MCP server disconnects, Claude Code should automatically attempt to reconnect
2. Configurable retry: Allow configuration for:
   • Number of retry attempts
   • Retry interval/backoff
   • Whether to notify user on reconnect
3. Graceful degradation: Only stop attempting reconnection if:
   • User explicitly disables the server
   • Max retries exceeded (with user notification)
   • Server is explicitly removed from config

Proposed Configuration

{
  "mcpServers": {
    "postgres": {
      "command": "...",
      "autoReconnect": true,
      "maxRetries": 5,
      "retryIntervalMs": 3000
    }
  }
}

Impact

This would significantly improve the experience for users who rely on MCP servers for database access, GitHub integration, and other tools during development sessions.

Environment

• OS: Windows (MINGW64)
• Claude Code: Latest version