Web Fetch Tool
A tool that retrieves the full content of a specific URL for an AI agent to read.
What Is a Web Fetch Tool
A web fetch tool is a capability that lets an AI agent retrieve the content of a specific web page or URL, as opposed to searching for pages in the first place. Given a URL, the tool loads the page and returns its content to the agent, typically as plain text, so the agent can read, summarize, or extract information from it as part of a larger task.
How It Works
The agent calls the web fetch tool with a target URL, usually one it found through a web search tool or one provided directly by the user. The tool then requests that page and processes the response before handing it back to the agent. Simple pages made mostly of static HTML can be fetched and converted to plain text directly. Pages that rely heavily on JavaScript to render their content, such as many modern web applications, require the fetch to be performed through a headless browser that executes the page's scripts before extracting the resulting text, since a raw HTML request would return an empty or incomplete shell.
Web Fetch vs Web Search
The two tools are complementary and are often used in sequence. Web search answers the question of what pages exist for a given query, returning a short list of candidates. Web fetch answers the question of what a specific page actually says, returning the full content of one URL. An agent researching a topic typically searches first to find candidate sources, then fetches the most relevant ones to read their full content before forming an answer.
Agenhood and the Web Fetch Tool
Agenhood's Vanilla driver ships a web fetch tool with two modes. By default it retrieves a page and returns its plain text content, which is fast and sufficient for most static pages and documentation. For JavaScript-heavy pages that do not render their content without a browser, the tool can instead route the request through a headless Chromium instance, which loads the page, executes its scripts, and returns the fully rendered text. Both modes execute inside the agent's sandboxed Docker container, so fetched content and any side effects of loading a page are contained within the same security boundary as the rest of the agent's actions.