web-fetch

name

web-fetch

description

Fetch content directly from URLs and return it as markdown, text, or raw HTML. Use when the user wants to read a webpage, inspect a URL, pull docs from a site, extract page content from a link, or fetch web content without opening a browser. Prefer this over browser automation when a normal HTTP fetch is enough. Supports GitHub blob URLs and sites that return LLM-optimized markdown for Accept: text/markdown.

Web Fetch

Use this skill when the user gives you a URL and wants the page contents or information derived from that page.

Prefer this skill over browser automation when:

• the page is publicly fetchable
• the user wants content, not interaction
• JavaScript rendering is not obviously required

Use browser tooling instead when the page requires login, heavy client-side rendering, form interaction, clicks, or debugging in a real browser.

How to fetch

Run the wrapper script relative to this skill directory (requires bun or node):

./web-fetch <url> [format]

Formats:

• markdown — preferred default for most reading/summarization tasks
• text — use when the user wants plain extracted text
• html — use only when the user explicitly wants raw markup or you need to inspect the source

Examples:

./web-fetch https://example.com markdown
./web-fetch https://example.com text
./web-fetch https://example.com html

Behavior

The fetcher already does the following:

• requests Accept: text/markdown, text/html;q=0.9 first, so sites that support agent-friendly markdown can return it directly
• falls back to converting HTML when the server does not return markdown
• converts GitHub .../blob/... URLs to raw content URLs automatically
• enforces a timeout and truncates very large responses

How to respond

• If the user asked to fetch or read a page, return the fetched content in the most useful format for the task.
• If the user asked a question about the page, fetch it first, then answer using the fetched content.
• Mention fetch failures clearly instead of silently falling back to guesses.
• If a page appears to require browser interaction or client-side rendering, say so and switch to browser tooling rather than pretending the fetch result is complete.