coding

Claude Code's Lazy-Loading MCP Cuts Context 85%

Claude Code's new lazy-loading Model Context Protocol reduces token usage by 85% through on-demand resource fetching, enabling developers to work with larger

Someone found that Claude Code now has lazy-loading MCP tools that cut context usage by 85% (77K down to 8.7K tokens).

Instead of loading every tool upfront, it searches for what it needs on-demand. Pretty huge for keeping sessions fast when working with tons of integrations.

New commands worth knowing:

  • /config search <term> - finds specific settings instantly
  • /stats --filter <agent> - shows performance per agent
  • Custom keybindings in .claude/keybindings.json

The custom agents feature is wild too - create specialists that Claude calls automatically. Each gets its own isolated context window, so a “database guy” agent won’t pollute your frontend work.

Session teleportation lets you move active sessions between terminal and claude.ai/code without losing state. Background tasks run agents in parallel instead of sequentially.

Full guide at thedecipherist.com/articles/claude-code-guide-v4/ with all the configs and examples.