[FEATURE] Add UI toggles for MCP servers and built-in tools to reduce baseline token usage

Resolved 💬 3 comments Opened Apr 17, 2026 by leanddouglas Closed Apr 20, 2026

Preflight Checklist

  • [x] I have searched existing requests and this feature hasn't been requested yet
  • [x] This is a single feature request (not multiple features)

Problem Statement

Every Claude Code session loads tool schemas into context before the user sends a single message. On my setup: i know i can disable in settings , but sometimes I am working with different models simultaneously and its time consuming to go back to settings every time i need to turn on and off a tool to optimize for token wasted.

System tools: ~11.8k tokens
MCP tools (GitHub server): ~14.5k tokens
System prompt: ~5.8k tokens
Memory files: ~2.8k tokens
That's ~35k tokens consumed before I type anything. If I'm only asking a quick question or doing work that doesn't touch GitHub, those MCP schemas are pure waste — and they count against my usage.

Proposed Solution

  1. In-session UI toggle — a simple panel (e.g. in /mcp or a new /tools command) to enable/disable MCP servers and tool groups on the fly, without editing settings.json and restarting.
  1. Smarter auto-loading — Claude Code should be able to read the user's prompt and decide which MCP servers / tool groups are actually relevant, loading schemas on demand. Deferred tools via ToolSearch are a step in this direction, but MCP servers are still fully loaded upfront.
  1. Per-session overrides — let me say "this session doesn't need GitHub" without permanently disabling it.

Why it matters

For consumer/Pro-plan users on token-limited sessions, paying 14k+ tokens per session for tools I'm not using feels like a tax. The infrastructure (deferred tools) already exists — extend it to MCP servers and expose a UI.

Alternative Solutions

_No response_

Priority

Critical - Blocking my work

Feature Category

MCP server integration

Use Case Example

_No response_

Additional Context

_No response_

View original on GitHub ↗

This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗