Michael JohnsonOn October 2, 2026, Cloudflare introduced Web Search API through AI Gateway—an interface for web search by AI agents and applications. At launch, the company connected Ceramic.ai, Exa, and Linkup. The official overview documentation, checked on October 8 and updated on October 2, identifies the product as an open beta.
For developers, the key change is the ability to add search to an existing model workflow. According to the AI Gateway integration documentation, search requests share logs, analytics, billing, and access controls with model requests. This makes it possible to treat search as a distinct, observable step in an agent’s task execution.
Why an agent needs a search query
In its announcement, Cloudflare describes a practical problem: an agent may guess the address of a page it needs, only for a request to the wrong URL to return a 404. A search tool offers another way forward: formulate a query, get the URLs of relevant materials, and use the information found to prepare an answer.
A second use case is dealing with changes that occur after a model is trained. Cloudflare specifically mentions recent events, changing APIs, and new tool releases. For an agent that helps write code, a useful query might search for documentation on a specific library version or a recently introduced parameter. In this process, search supplies material for context, while the application decides which results to pass to the model.
A practical takeaway for teams: save the query wording and URL alongside each retrieved passage. That way, when checking an answer, you can trace which material the agent relied on. For version-specific tasks, it is also useful to include the product name and required version explicitly in the search query.
How to connect the new API
The documentation describes two ways to make calls: the REST API from an external backend and an AI binding inside Cloudflare Workers. The provider is selected with the provider parameter. All three services return results in a common format: titles, URLs, and descriptions. This makes it easier for the application to switch providers when processing responses.
A common format is useful for building your own search-tool adapter. Its job is to accept an agent’s query, call the selected provider, and pass the chosen results to the model. At the application level, decide separately when a result’s description is enough and when the found page needs to be read to verify a specific claim.
The launch announcement describes built-in AI Gateway server-side tools as a future capability. Cloudflare says web search will be one of the first such tools. The launch example assumes custom orchestration: handling the tool call and passing the search response into the next model request.
What terms have been confirmed
According to the billing terms in the documentation, requests can be paid for with AI Gateway credits at the provider’s public API price, with no additional Cloudflare markup. Bring Your Own Key is also supported, allowing the use of a search provider’s own API key. With this option, teams should decide in advance where they will track usage and reconcile charges.
The reviewed announcement and overview page do not provide specific monetary rates or numerical request quotas. Therefore, calculate the integration budget after checking the selected provider’s pricing. It is useful to build acceptance tests around your own tasks: the share of relevant sources, response time, number of search requests per completed task, and error handling.
Data retention is another consideration. In its description of Zero Data Retention, Cloudflare promises to identify partners that support ZDR. For a specific search route, check the provider’s retention terms and AI Gateway logging settings separately: the announcement also confirms that search queries are recorded in the observability system.
Providers also agree to follow Cloudflare’s Verified Bots requirements: identify their bots, respect robots.txt and crawl directives, maintain a reasonable request rate, and honor website owners’ preferences. The Web Search API overview confirms that every result must include a link to its source.
For developers, the launch provides a clear way to connect web search to an agent and a shared framework for tracking it. Start your evaluation with a small set of real tasks, an explicit provider choice, and links saved in responses. Those decisions will determine how well the materials found help the application do its specific work.