Cloudflare services for private origins is now generally available
Source
blog.cloudflare.com
Author
Enrique Somoza
Date
Why it matters
Engineers can now front private MCPThe Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.Full definition → servers and AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → backends with Cloudflare's WAF, bot management and rate limiting without a public IP, open firewall port, or cloudflared on the origin. Argo Smart Routing is among the remaining limitations.
Key takeaways · AI-distilled
Cloudflare's private-origin support is now generally available: a proxied A or AAAA record pointed at a private IP puts WAF, bot management, rate limiting, caching, rewrites and Workers in front of it, reached over an existing Cloudflare One connection.
The origin keeps its private IP with no inbound firewall exception. With Cloudflare WAN no connector software runs on the origin; with Cloudflare Mesh one connector inside the private network can serve multiple origins.
Private routing turns on automatically, and cannot be disabled, for RFC 1918 IPv4, RFC 6598 CGNAT (100.64.0.0/10) and RFC 4193 IPv6 (fc00::/7) addresses. In the API it is a private_routing attribute on a standard DNS record.
Limits remain: Argo Smart Routing and Dedicated CDN Egress IPs are unsupported, origin health checks need Load Balancing, and access needs a Private Origins entitlement from the account team plus Unified Routing unless the connection is Cloudflare Tunnel.
Cloudflare says the most common beta setup failure was a missing return route for its source ranges on WAN or Mesh. Private-to-private traffic, such as agents on one private network reaching apps on another, is targeted for closed beta in early 2027.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.
