The tool searches the public web through the Cloudflare Web Search API, which Cloudflare says is powered by ceramic.ai, Exa, and Linkup.
An agent using the tool supplies a query and a maximum result count. It receives a short numbered list containing titles, URLs, and trimmed descriptions. When web_fetch is also available, the model is instructed to use it to read pages returned by web_search.
Cloudflare says the feature requires Agents SDK v0.27.0 or later. Its example installs webSearchTool in a Pi Durable extension registry before opening the harness. The SDK also offers AI SDK and TanStack AI adapters, plus createAIWebSearch() and createHTTPWebSearch() for searches run without a model, such as scheduled jobs.
The changelog does not state pricing, rate limits, regional availability, result-quality guarantees, or the separate terms and limits of the named search providers. Source: https://developers.cloudflare.com/changelog/post/2026-10-12-web-search