Cloudflare rolled out a beta Web Search API directly inside AI Gateway in early October. It lets an agent or application inject fresh web results into a model’s context at inference time, without wiring up a separate search vendor. The launch landed during the company’s Birthday Week and addresses a familiar gap: a model frozen at its training cutoff knows nothing about what happened since.
Three search providers, one bill
The feature relies on three partners Cloudflare picked for web indexing: Ceramic.ai, Exa, and Linkup. Each request names the provider explicitly, giving a choice between general-purpose search and engines tuned for semantic retrieval or agent workflows. Billing runs through the AI Gateway credits already used for inference, at the partners’ list pricing with no markup added by Cloudflare.
| Provider | Positioning | Access |
|---|---|---|
| Ceramic.ai | General-purpose web search for agents | REST, Workers binding |
| Exa | Semantic search tuned for RAG | REST, Workers binding |
| Linkup | Structured results with source citations | REST, Workers binding |
Wiring it up as a developer
Two access paths are offered: a classic REST endpoint authenticated with a bearer token, and a native binding for Workers. The latter skips an extra HTTP round trip for applications already hosted on the platform.
const results = await env.AI.websearch({
gatewayId: "default",
query: "latest stable Laravel release",
provider: "exa",
limit: 5,
});
Every web search request runs through the same observability logs as a standard AI Gateway inference call.
Governance and data handling
Teams can bring their own keys (BYOK) with each provider instead of relying on Cloudflare’s managed account. Every partner is tagged with its data-retention status, and all three committed to honoring robots.txt and Cloudflare’s verified-bot requirements.
What to remember
Cloudflare turns web search into an infrastructure primitive consumable the same way as a model call, which simplifies building agents able to check a recent fact without the team managing a separate contract and invoice with a search vendor. Paired with protocols like the one described in the July 2026 MCP specification, this primitive makes it easier to ship editorial agents that produce verified content rather than static answers. The Workers binding integration meaningfully lowers the barrier to entry for projects already hosted on the platform.
On projects where real-time search is duct-taped together with homemade scraping or a poorly rate-limited Google Search API key, this kind of primitive changes the equation: one endpoint, one invoice, no more juggling three separate vendor contracts. Whether Exa or Linkup results hold up against a plain Google search on tricky queries still needs a real-world test before plugging it into a production agent. — Simon Janvier
Further reading: the full official announcement on Cloudflare’s blog.
