Reclaim 30–90K context tokens with Claude Code's MCP Tool Search (`ENABLE_TOOL_SEARCH=true`)

Claude Code 2.1.7+ shipped a native flag: when registered MCP tool descriptions exceed roughly 10% of the context window, they auto-defer and Claude discovers them on demand through an MCPSearch tool. Two-line edit to settings; restart; the boot banner shows "mcp tools (loaded on-demand)" when active. One operator reported context dropping from 93.1K MCP-token usage to 0.429K after switching to the equivalent third-party proxy — the native flag now delivers the same benefit.

Heavy MCP users (Figma, Linear, GitHub, Stripe, dataforseo, Chrome) commonly burn 30–50K tokens on tool descriptions before the first user message. That's a measurable hit to long-session quality. Reclaiming this is a single-line settings change with no functional regression for most setups.

1. Upgrade Claude Code to 2.1.7 or later.
2. Edit ~/.claude/settings.json and add the env block below.
3. Restart Claude Code.
4. Verify on the boot banner: look for mcp tools (loaded on-demand).
5. Run /context and confirm MCP token cost is near zero until a tool is invoked.

{
 "env": {
 "ENABLE_TOOL_SEARCH": "true"
 }
}

On-demand loading doesn't always handle MCP servers that need auth — one operator's Figma MCP failed in this mode and had to be left in the eager-load set. Also: MCP servers that expose unusually long tool descriptions can break the search-driven import path. If a specific MCP misbehaves, run with the flag off for that server and on for the rest.