Web Fetch

Fetch a known URL and return the page as markdown, plain text, raw HTML, or a downloaded image file using the local fetch.js helper. Use this whenever the us...

Security Scan

The skill does what it claims — fetch a provided URL and return or save its contents — and its code, instructions, and requirements are consistent with that purpose.

Purpose & Capability

Name/description match the provided code and runtime instructions: the skill fetches a given URL and returns markdown/text/html or saves images. No unrelated credentials, binaries, or external services are requested.

Instruction Scope

SKILL.md restricts use to fetching known URLs and shows only the node ./fetch.js CLI; the instructions do not request extra env vars or read other system state. Note: the tool can write downloaded images to a user-supplied path (or a temp file) — choose the output path carefully to avoid overwriting local files.

Install Mechanism

No install spec or remote downloads are used. The bundled fetch.js is the only code; nothing is fetched from external installers or obscure URLs at install time.

Credentials

The skill requires no environment variables or credentials. It performs network requests only to the user-provided URL and writes image output only to a specified path or temp directory.

Persistence & Privilege

always is false and the skill does not request persistent or system-wide privileges. It does write image files to the filesystem when asked (tempdir by default), which is expected behavior.

Assessment

This skill is internally consistent and implements a straightforward HTTP fetcher. Before using it: (1) be careful when specifying --output to avoid overwriting important files (it will write binary data to whatever path you provide); (2) only fetch URLs you trust — the tool will download content up to 5MB and save images to disk; (3) ensure your Node runtime supports global fetch and AbortSignal.timeout (modern Node 18+); and (4) if you need the agent to fetch pages from behind authentication, note the skill does not handle credentials. If any of these behaviors are unacceptable, do not enable the skill.

Use this skill when the target URL is already known and the job is retrieval, not search.

Common cases:

• fetch a documentation page the user already linked
• pull an article into readable markdown or plain text before summarizing it
• inspect raw HTML to debug markup, metadata, or page structure
• download a remote image to a local path for later inspection or processing

If the user needs help finding the right page first, use web-search before this skill.

Command

Run:

node ./fetch.js --url "https://example.com"

Optional flags:

• --format markdown|text|html
• --timeout <seconds>
• --output <path> for image responses

Choose the right format

Default to markdown unless the user clearly wants something else.

• markdown: best for readable docs, articles, and summarization
• text: best when the user wants the cleanest plain-text extraction
• html: best when inspecting source markup, metadata, links, embeds, or page structure
• --output <path>: use when the response is an image and you want a stable saved file path instead of a temp file

Examples:

node ./fetch.js --url "https://example.com/docs" --format markdown

node ./fetch.js --url "https://example.com/page" --format text --timeout 20

node ./fetch.js --url "https://example.com/page" --format html

node ./fetch.js --url "https://example.com/logo.png" --output /tmp/logo.png

How to work with the result

The CLI prints a <web_fetch> block.

For text-like responses it includes:

• title
• url
• mime
• format
• content

For image responses it includes:

• title
• url
• mime
• image

After fetching:

1. Read the returned fields carefully.
2. Use only the parts relevant to the user's request.
3. Summarize or quote concise excerpts instead of dumping the whole page unless the user asked for the full content.
4. If you fetched HTML, explain notable structure or metadata rather than pasting large raw blobs unless needed.
5. If you downloaded an image, tell the user the saved file path.

Behavior and limits

Keep these in mind while using the tool:

• http:// URLs are tried as https:// first, then retried as plain HTTP if needed.
• Default timeout is 30 seconds; maximum is 120 seconds.
• Responses larger than 5MB are rejected.
• Browser-like headers are used, with a retry path for some Cloudflare 403 challenge responses.
• SVG is treated as text, not as a binary image attachment.

Failure handling

If the fetch fails:

• verify the URL is complete and starts with http:// or https://
• retry with a different format only if the user's goal changed
• if the page is too large, blocked, or times out, tell the user plainly and suggest a narrower target URL or an alternate source