Cloudflare describes AI Search as a managed retrieval layer for structured and unstructured data. The preview accepts files, sites and database sources, then exposes indexed content through a native /search API and a Model Context Protocol endpoint at /mcp. Cloudflare also describes a custom-domain MCP path and private access through Cloudflare Access.
Managed retrieval, two agent paths
The product is designed to remove much of the indexing and hosting work from a team. The material names discovery and parsing for web content, namespaces for separating tenants or corpora, and a development stack that can surface up to ten relevant surfaces. For an application that already speaks MCP, /mcp is the obvious integration path. For a conventional application, /search provides a narrower API boundary.
That is a workflow choice, not a guarantee that retrieval quality will fit every corpus. Index freshness, parsing errors, metadata, citations, namespace design and deleted content still need to be observed on the actual data.
Preview economics
| Preview item | Cloudflare-listed rate |
|---|---|
| Ingest | $0.75 per million tokens |
| Storage | $2 per GB-month |
| Semantic search | $0.75 per 1,000 queries |
| Full-text search | $0.10 per 1,000 queries |
Cloudflare says the preview includes free allotments and that prices may change. The table is a planning baseline, not a quote. A team should add crawl, embedding, storage, egress, application and model costs before comparing it with a self-hosted index.
Private data needs a private route
Cloudflare Access can protect the custom-domain MCP endpoint, but an access policy does not replace corpus authorization. Test tenant isolation, namespace mistakes, document deletion, prompt-injected content and citation behavior. Keep credentials and production data out of a first pilot until the route and logs are understood.
Test the corpus, not the brochure
- Load a small owned corpus with known answers, stale documents and deliberate duplicates.
- Compare
/searchand/mcpfor recall, citations, latency and failure behavior. - Verify Access policy, namespace isolation and document deletion.
- Measure query, storage and application cost for a fixed workload.
- Set an exit rule if citations are missing, freshness is unclear or data crosses a boundary.
Watch for general availability, changed pricing, independent retrieval evaluations and evidence about private-data controls. AI Search is promising as a managed retrieval building block; it is not self-validating simply because the endpoint is managed.
Sources: Cloudflare Blog — AI Search, AI Search documentation, Cloudflare Agents documentation and Cloudflare Access documentation.