Cloudflare AI Search combines the parts of a retrieval system—crawling, parsing, embeddings, vector storage and a search API—into one managed service for AI agents and applications. Organizations can index structured or unstructured sources and query several websites or instances as one corpus.
A new discovery mode can find pages on sites that do not publish a sitemap. Developers create an index with a Wrangler command that specifies the source and parser, then connect it to an application or Model Context Protocol server through a Worker. Cloudflare also offers public search and MCP endpoints that can be shared without a separate deployment or authentication.
The service builds on existing Cloudflare components including Workers AI, AI Gateway, Vectorize, R2 and Browser Rendering. It can support multimodal retrieval and already indexes documentation for services and projects such as Cloudflare’s API, Astro, Vite, Hono and Replicate.
AI Search remains in beta and is free during that period. At general availability, Cloudflare plans to make embedding and reranking free when customers use default or selected Workers AI models; answer generation and query rewriting will be billed according to model usage. Public unauthenticated endpoints are convenient for open documentation, but teams indexing private or sensitive material will need to put access controls around their own integration.