Weekend fun experiment: What if AI agents could see icon libraries the way humans do? Iconic is an MCP service that helps agents like Anthropic Claude find icons by intent. How?
I ingested ~20,000 icons across permissively licensed icon sets, then ran every set through a local LLM (qwen2.5-coder) to extract semantic metadata (keywords, categories, use cases, visual description), then exposed the catalog via MCP for agentic contextual search.
It empowers agents to search for intents like:
"Suggest me icons for a live music event"
"Show me thin-line options for a settings page"
"Find icons similar to this one"
.. And so on.
Up next: Started playing with a RAG scaffolding to generate brand new icons in the style of existing icon libraries. We'll see where we get from here 😄