MCP server optimisation cuts AI site build time from 30 minutes to 4

A development team building websites via AI prompts discovered their 30-minute landing page build time was dominated not by component generation but by image uploads, which alone consumed 10 of 14 minutes after an initial round of optimisations. The root cause was that AI agents encode file uploads as base64 text, forcing the language model to generate thousands of tokens character-by-character for even small images. A secondary issue was that every write operation returned the full saved record, causing roughly 55 KB of redundant JSON to echo back per page build and inflating token usage significantly. The team addressed both problems at the server level by enabling inline image ingestion via URL — so bytes never pass through the model — and adding a verbose flag to suppress unnecessary response payloads. The combined changes reduced a full site build from 30 minutes to 4 minutes and cut token consumption from approximately 225,000 to 40,000.