Infra notwithstanding - I'd be interested in hearing how much success they actually had using a locally hosted MCP-capable LLM (and which ones in particular) because the E2E tests in the article seem to be against remote models like Claude.