fresh-fetch

fresh-fetch

Stop answering web-content questions from stale training data. Always go to the source.

When this skill applies

Fire on:

• Any URL in the user's message that they appear to be asking about (not just sharing)
• Any question about content, docs, APIs, libraries, prices, dates, versions, or behavior that lives on the public web — even with no URL in the message

Skip on:

• URL in a code block / fenced sample (it's code, not a question)
• URL is localhost, 127.0.0.1, or a private IP — note "looks local, no fetch attempted" and continue
• User wants you to act on the URL via another tool (open this PR, git push to <url>, query this linear ticket) — use the action tool, not the fetcher
• URL was produced by you in a prior turn — don't infinite-loop on links you just printed

The procedure

Run this top to bottom. Stop at the first step that succeeds.

1. In-session cache check

If you've already fetched this URL in the current session within the last 15 minutes (WebFetch's built-in cache window), reuse it.

Status line: [fresh-fetch] <url> — cached Nm ago

2. Specialized MCP

Scan the available tools list for an MCP that natively serves the URL's host. Match on capability suffix, never on the per-user server UUID prefix.

URL host	Look for a tool matching
github.com / *.github.com	mcp__*github*
linear.app	linear issue/doc tools (mcp__*__get_issue, *__list_issues, etc.)
supabase.com/docs	*search_docs* on a supabase server
vercel.com/docs	*search_vercel_documentation*
render.com	mcp__*render*
slack.com / *.slack.com	mcp__*slack*
(others)	grep for the host's brand name in the tool list

If you find a match, use it. Status line: [fresh-fetch] <url> — source: <mcp-server-name>

If no specialized MCP matches, continue to step 3.

3. WebFetch

Call WebFetch with the URL and a prompt that extracts what the user actually needs.

• Success → Status line: [fresh-fetch] <url> — source: WebFetch (fresh, Ts)
• Failure (robots.txt, 403/4xx, "javascript required", auth wall) → continue to step 4

4. Chrome MCP

Check whether the Chrome MCP is installed. Use ToolSearch with query claude-in-chrome (or similar). If tools matching mcp__*claude-in-chrome* appear, load them, navigate to the URL, and extract the page content.

Status line: [fresh-fetch] <url> — WebFetch failed (<reason>) → Chrome MCP

If Chrome MCP is not available, continue to step 5.

5. Playwright

Check whether Playwright is installed and the chromium binary exists:

npx -y playwright --version 2>&1 | head -1
test -d "$(npx -y playwright install --dry-run chromium 2>&1 | grep -oE '/[^ ]+chromium[^ ]*' | head -1)" && echo "chromium present" || echo "chromium missing"

If both succeed, run a small inline extract script:

node -e "
const { chromium } = require('playwright');
(async () => {
  const browser = await chromium.launch();
  const page = await browser.newPage();
  await page.goto(process.argv[1], { waitUntil: 'networkidle', timeout: 30000 });
  console.log(await page.evaluate(() => document.body.innerText));
  await browser.close();
})();
" "<URL>"

Status line: [fresh-fetch] <url> — source: Playwright

If Playwright is not installed or chromium binary is missing, continue to step 6.

6. Refuse and offer install

Do not guess from training data. Print exactly:

[fresh-fetch] Could not fetch <url>.

WebFetch:    <actual reason: blocked by robots.txt | 403 | JS-only | timeout>
Chrome MCP:  not connected
Playwright:  not installed

I won't guess from training data — too high a chance of being wrong.

Options:

1. Chrome MCP — recommended; uses your existing Chrome session, handles
   auth/cookies/JS automatically.
   Install: visit https://claude.ai/chrome and follow the prompts to install
   the official Claude for Chrome extension (requires Pro/Team/Enterprise plan).

2. Playwright — headless, no auth state, works in CI.
   Install: npm install -g playwright && npx playwright install chromium

Reply "chrome" (after installing the extension, I'll wait for confirmation)
or "playwright" (I'll run the install for you with your OK).

Then wait.

• "chrome" → wait for the user to confirm the extension is installed, then restart this procedure from step 2.
• "playwright" → confirm before running (this is a risky action — global npm install + browser download). On user OK, run:

  npm install -g playwright && npx playwright install chromium
  

  Then restart this procedure from step 5.
• "neither" / something else → stop. Tell the user you can't proceed without a fetcher.

Edge cases

Case	Behavior
Multiple URLs in one message	Run the procedure for each URL in parallel; one status line per URL
Auth-walled URL (Notion, Linear with no MCP, Figma, internal company tools)	WebFetch will fail with 401/403 — jump straight to Chrome MCP since only it has the user's session
PDF, image, or other non-HTML	Try WebFetch once. On failure, hand the URL back to the user with a note ("PDF — Chrome/Playwright won't help extract this; consider downloading it locally"). Do not escalate to browser fetchers for binary content.
URL produced by you in a prior turn	Skip — assume the user is referencing it, not asking about it

Status line format

Always one short line, then the answer. Never more, never less.

[fresh-fetch] react.dev/learn — source: WebFetch (fresh, 1.2s)
[fresh-fetch] github.com/foo/bar — source: github MCP
[fresh-fetch] notion.so/abc — WebFetch failed (auth wall) → Chrome MCP
[fresh-fetch] linear.app/team/X/issue/1 — cached 4m ago

The status line is required even when the fetch is trivial. It's how the user knows what just happened and gives them a chance to redirect.

What this skill is not

• Not a screenshot tool — only text/markdown content extraction.
• Not a persistent cross-session cache — relies on WebFetch's existing 15-minute in-session cache.
• Not a replacement for action tools — if the user wants to push code, open a PR, or run a query through a connected MCP, use that tool instead of fetching the URL.