Every query you send here is routed by OpenAI's Responses API to our live MCP server at
matscout.prfo.design/mcp/http/. The LLM (gpt-4o) doesn't run the 10 tools
itself — it discovers them via MCP tools/list, then calls them back over HTTP.
Open DevTools → Network and you'll see response.output_item.added events
whose type=mcp_call are real round-trips to this MCP endpoint.
Want to use the same MCP server from your own Claude (Desktop or Code)? It's a remote MCP — just point at the URL:
{
"mcpServers": {
"matscout": {
"url": "https://matscout.prfo.design/mcp/http/"
}
}
}
Drop into ~/Library/Application Support/Claude/claude_desktop_config.json
and restart Claude. No API keys, no subprocess — it's a live network endpoint. For the
stdio (local subprocess) flavour, see github.com/shorokhlev-sketch/matscout.
Type a natural-language description of a material you need. The agent will translate it to property filters, query the database, self-correct if it gets zero or too many hits, and synthesize a ranked shortlist with trade-offs.