Agents no longer chew raw HTML: Cloudflare folds fetching and Markdown conversion into one tool

An agent can now read webpages and documents like any other tool call, receiving not just the body but also the final URL, HTTP status, and content type. Previously such reading required the agent to handle page formats itself, or the developer to build a separate fetching and conversion pipeline. The tool ships in the Agents SDK, requests Markdown directly from sites that support Markdown for Agents, and sends the rest through Workers AI Markdown Conversion before the model sees it.

Long pages are not handed to the model all at once; they come back in smaller sections with an offset the model can use to continue reading. This changes the case where an entire page had to enter the context in one pass. Developers control which hosts the agent may fetch, response size, timeouts, and redirects.

When the agent also has the web_search tool, the model is told to use web_fetch to read the pages that web_search returns. Search and reading become one flow rather than two separate steps. The tool works with any harness, the SDK includes AI SDK and TanStack AI adapters, and fetching without a model is possible through fetchWeb(), windowWebFetchPage(), and renderWebFetchPage() from agents/webfetch, with no Durable Object or browser required.

The tool targets developers using the Cloudflare Agents SDK, requires v0.28.0 or greater, and must be added to a Pi Durable extension registry before the harness opens. It reads webpages, PDFs, and supported Office documents, and developers control allowed hosts, response size, timeouts, and redirects.

The source gives no data on fetch success rates, conversion quality, or latency, and does not say which Office formats count as supported.