Built an MCP server that turns awwwards.com into an agent-readable design library. 7 tools over stdio (official SDK, TypeScript, zero native deps):
search_sites — filter by color hex / tag slugs / technology / award; results carry inline JPEG screenshots (agents judge visually)
get_site_details / get_site_elements — palette, stack, jury score; element-level posters + mp4 URLs for motion study
capture_live_site, analyze_page_structure, record_site_motion — optional playwright tools: full-page capture, DOM band-map JSON, and a ~27s scripted motion recording returned as filmstrip + .webm
list_categories — the full filter taxonomy (198 tag slugs, 27 colors)
source & further reading
dev.to — original article
Token-Efficient Agentic Development — Part 1: What Are You Actually Paying For?
The model is the easy part. What building Dantiva taught me
Claude Code Session Compaction in 2026: How Context Summarization Works and What Your Agent Forgets