Cloak Fetch

cloak-fetch — bot-protection / WAF fetch fallback

When a normal HTTP fetcher fails because the target site uses bot protection or WAF screening, route the same URL through CloakBrowser (a real Chromium with anti-bot patches at the C++ level and a genuine TLS/JA3 fingerprint) and return clean markdown via trafilatura. The agent never has to tell the user "the page is unfetchable" — it can try this skill first.

The underlying engine doesn't care which vendor is blocking the request — Cloudflare, DataDome, Akamai, PerimeterX, Imperva, Distil, Kasada, AWS WAF, or a no-name WAF rule. They all eventually let a real browser through, and CloakBrowser is a real browser.

When to trigger

Use immediately after any of these symptoms from a recent fetch tool call:

Vendor-specific signatures

If the failure string roughly matches one of these, it's a CloakBrowser case — don't waste another round-trip on a plain HTTP client:

Also trigger preemptively when the user asks to fetch a page on a domain known to live behind one of these stacks — publishers (science.org, nature.com, sciencedirect.com, jstor.org), news (nytimes.com, bloomberg.com, ft.com, wsj.com), retail (nike.com, adidas.com, sephora.com), travel (kayak.com, southwest.com), financial (most broker portals). Going straight to this skill avoids a guaranteed-to-fail plain-HTTP round trip.

When NOT to trigger

In these cases, report the actual failure to the user instead of masking it.

How to invoke

One command — the wrapper picks the right Python, runs the headless browser, extracts the main content with trafilatura, and writes clean markdown to stdout:

<SKILL_DIR>/cloak_fetch.sh "<URL>"

Where <SKILL_DIR> is wherever this skill is installed. Common locations:

• Claude Code: ~/.claude/skills/cloak-fetch
• OpenClaw: ~/.openclaw/skills/cloak-fetch
• Codex: ~/.codex/skills/cloak-fetch
• Project-local: .claude/skills/cloak-fetch or skills/cloak-fetch

A portable invocation that finds the skill across these locations:

for d in \
  "$HOME/.claude/skills/cloak-fetch" \
  "$HOME/.openclaw/skills/cloak-fetch" \
  "$HOME/.codex/skills/cloak-fetch" \
  ".claude/skills/cloak-fetch" \
  "skills/cloak-fetch"; do
  if [ -x "$d/cloak_fetch.sh" ]; then
    SKILL_DIR="$d"; break
  fi
done

"$SKILL_DIR/cloak_fetch.sh" "https://www.science.org/content/page/information-authors-research-articles"

The wrapper streams clean markdown on stdout. Save to a file with > out.md or pipe directly into further processing.

Behavior

• Headless by default — no browser window opens.
• Latency: ~20–40 s per call (browser launch + page render + content settle).
• Output: clean markdown via trafilatura. Page chrome, navigation, ads, and cookie banners stripped. Headings, lists, links, and code blocks preserved. When trafilatura can't isolate a main content node, the wrapper falls back to emitting the rendered HTML so the agent still has something to read.
• Failure modes: exits non-zero with a message on stderr if no cloakbrowser-enabled Python is found, the browser couldn't reach the URL, or CloakBrowser couldn't pass the challenge. Surface the failure honestly to the user — do not fabricate page content.

Configuration

For more advanced tuning (headless toggle, content-selector wait list, settle timing), edit cloak_fetch.py directly — see comments in that file.

Example end-to-end

User: "Get the authors info from https://www.science.org/content/page/information-authors-research-articles"

1. Agent tries WebFetch — gets HTTP 403 Forbidden.
2. Agent recognises the 403 → invokes this skill:

   ~/.claude/skills/cloak-fetch/cloak_fetch.sh \
     "https://www.science.org/content/page/information-authors-research-articles"
   

3. ~25 s later, ~26 KB of clean markdown lands on stdout.
4. Agent answers the user from the markdown — no need to mention the underlying fetch took two attempts.

Related

• cloakFetch hook — same fallback wired up as a Claude Code PostToolUse hook (fully automatic, no agent decision required). Use the hook on Claude Code; use this skill on agents that lack a hook system (Codex, OpenCode, OpenClaw, etc.).

Source: Agents365-ai/cloakFetch — distributed by TomeVault.