Cloudflare launches Web Search API beta with three search providers behind AI Gateway

Cloudflare has launched a Web Search API in beta, a single service that lets AI agents and applications search the live web and ground their responses in current information. According to the company's developer changelog, the point is to stop models from guessing at URLs or falling back on whatever was captured before their training cutoff.

The announcement, which surfaced on the Hacker News front page, positions the API as plumbing for grounded AI: one endpoint, several search backends, and management handled through Cloudflare's existing AI tooling.

One endpoint, three providers

Behind the API sit three search providers at launch: Ceramic.ai, Exa and Linkup. Developers choose the provider per request, which makes it possible to route different queries to whichever backend suits them without maintaining three separate integrations.

Cloudflare also notes that all three providers support Zero Data Retention for requests made through its platform, and that each has committed to the company's verified bot crawling standards — a nod to the ongoing friction between AI crawlers and website operators over how content is gathered.

Routing and billing through AI Gateway

The service runs through AI Gateway, Cloudflare's control plane for AI traffic, and that has two practical effects for developers. First, search requests appear in gateway logs alongside other AI traffic, giving teams one place to observe what their agents are actually searching for. Second, billing is folded into AI Gateway credits: requests are charged at each provider's list API price, with no markup added by Cloudflare. Teams that already have a relationship with a search vendor can supply their own provider API key instead.

How you call it

Two integration paths are documented. Over the REST API, you send a POST request to the websearch endpoint under your Cloudflare account, authenticated with a bearer token. The JSON body carries the query, the provider name (for example "ceramic"), a result limit, and gateway options pointing at an ID such as "default".

If you run on Cloudflare Workers, there is a native AI binding instead: a call to env.AI.websearch() with the same shape — gateway ID, query, provider and limit — and the response parsed as JSON. Cloudflare's example passes "exa" as the provider and asks for five results, but the parameters are interchangeable across the supported backends.

Why it matters

Retrieval grounding has become the default architecture for anything resembling an agent. A model that cannot check the live web will answer from stale training data, and prompt engineering alone does not fix that. Cloudflare's move lowers the friction of wiring search into that loop.

The multi-provider design is the more interesting part. Search quality varies between backends and query types, and a brokered endpoint makes it cheap to compare providers per request and then watch the results in a single log stream. For teams already routing model calls through AI Gateway, search becomes just another instrumented call rather than a new system to operate.

The verified-crawling commitment matters beyond the API itself. If agents are going to fetch the web at scale, publishers want to know who is doing the fetching and under what rules; tying provider access to Cloudflare's bot standards pushes at least some of that accountability upstream.

The caveats are straightforward: this is a beta, the provider pool is three names deep, and list-price billing means costs still scale with usage. But as agent infrastructure — one endpoint, observable, with swappable backends — it is a sensible addition, and it signals where Cloudflare expects the developer workload to go.