Feature request: lightweight persistent Q&A mode with MCP access

Problem

Each Claude Code session loads ~30k tokens of overhead (system tools, skills, memory files, MCP tool schemas). This is fine for development work, but makes it expensive for quick conversational Q&A.

The current workflow for "I have a quick question" is: open a session, wait for it to load, ask, close. There's no continuity between sessions, and interrupting a session that's doing work doesn't go well.

The browser (claude.ai) solves the persistence/conversation problem but has no access to local tools or MCP servers.

Proposed solution

A lightweight "assistant" or "chat" mode — something like claude --chat or claude --assistant — that:

• Keeps MCP server access (so tools like custom databases remain available)
• Skips loading the full tool suite (Bash, Edit, Write, Glob, Grep, LSP, etc.)
• Stays persistent / warm for quick back-and-forth questions
• Doesn't compete with or derail active development sessions

Essentially a middle ground between claude.ai (no local access) and full Claude Code (heavy, task-oriented).

Use case

Developer has a long-running Claude Code session doing implementation work, but also wants to ask quick questions throughout the day without cold-starting a new session each time or interrupting the working one.