Cloudflare routes agent web search through AI Gateway with three crawler-vetted providers

Anyone who has watched an agent try three spellings of a docs path will recognise the pattern Cloudflare describes ^[9]. Each wrong guess costs a tool call and returns a 404 with nothing usable in it.

The replacement is a search query. The provider returns structured snippets from the live web, and Cloudflare says those go straight into the model's context ^[11]. The rule I like best is the attribution requirement. Every search response must include a link to the location of the crawled content ^[2]. The agent gets a real URL for its next fetch, and a site owner gets a trail from the answer back to the page.

The billing works like the rest of the gateway. Search queries draw down the AI Gateway credit balance ^[6]. They appear in the same observability logs as other gateway traffic, and admins decide who can reach which provider ^[5]. Cloudflare charges the partners' list API prices with no markup ^[8], so a query through the gateway costs the same as buying direct at list ^[14]. Teams with existing provider contracts can bring their own keys, as they already can for model inference providers ^[10].

The claim that search grounding becomes a gateway setting holds for the operational side and stops at the client. The direct path in the post is a standard REST endpoint, authenticated with the AI Gateway token ^[12]. The agent still needs a tool definition or a code path that sends the query and handles the results. What the gateway takes over is billing, logging, key handling and access policy ^[4][5][10]. On the credits path, a team already on AI Gateway does without a separate provider account and a second place to read request logs ^[5][6]. Cloudflare points to its web search developer docs for the rest ^[13].

The partner terms matter beyond this product. Each provider's crawler must meet the "Verified bots" requirements in Cloudflare's developer documentation ^[2]. The post describes this as identifying the crawler, respecting robots.txt and giving the source of search results ^[3]. Ceramic.ai, Exa and Linkup accepted those terms to get distribution through the gateway ^[1][2].

Data retention is still open. Cloudflare says it will identify partners that support Zero Data Retention ^[7], and the post does not yet name them. In my view, queries that carry customer data should stay off all three providers until Cloudflare publishes that list.