Expose MCP tools as a filesystem code tree so agents load only the definitions they touch (98.7% fewer tokens)

Anthropic Engineering·high signal

Instead of loading every MCP tool definition upfront, Anthropic's code-execution pattern lays tools out as files (e.g. ./servers/google-drive/getDocument.ts) that the agent discovers by browsing and reads on demand — or via a search_tools meta-tool that returns name-only, name+description, or full schema. In the Google Drive→Salesforce example this cut a task from 150K to 2K tokens (98.7%), and intermediate data stays in the execution environment rather than passing through the model's context. Pair it with a ./skills/ directory plus SKILL.md files to persist working implementations as a reusable toolbox.