anon-research

Anonymous internet research through Tor. Every search and every page fetch uses a fresh Tor circuit (different exit IP), so target sites can't link your requests to each other or to you.

When to use

Invoke this skill when the user wants to:

• Research a topic without target sites profiling them
• Look something up where the search itself is sensitive
• Pull content from multiple sites without those sites correlating the requests as coming from one source

Do not use this skill when:

• The user wants speed or rich JS-rendered content (use firecrawl instead)
• The user is logging in to something (anonymity is meaningless once you authenticate)
• The user asked for a specific search engine that blocks Tor (e.g., Google) — fetches will return BLOCKED

Quick start

# One-time setup (installs Tor if missing, starts the daemon):
~/.claude/skills/anon-research/scripts/anon.py setup

# Verify circuit isolation works (shows two different exit IPs):
~/.claude/skills/anon-research/scripts/anon.py status

# Search the web:
~/.claude/skills/anon-research/scripts/anon.py search "site:nature.com mrna vaccine"

# Fetch a specific page as readable markdown:
~/.claude/skills/anon-research/scripts/anon.py fetch https://example.org/article

Commands

Privacy model

Every request:

• Goes through Tor SOCKS5 (socks5h://127.0.0.1:9050) with a unique random SOCKS auth pair, which forces Tor to build a fresh circuit (different guard, middle, and exit relays). Sites cannot link requests by IP.
• Sends DNS through Tor (socks5h, not socks5) — no DNS leaks to your ISP's resolver.
• Uses HTTPS only by default; HTTP URLs print a warning.
• Carries no cookies (no session jar between calls), no Referer, and a generic Firefox-on-Linux User-Agent (a common, low-fingerprint choice — not the Tor Browser UA, which without the matching JS fingerprint is its own tell).

Fail-closed behavior

If a site blocks Tor exits (Cloudflare challenge, HTTP 403/429/451, common captcha bodies), fetch returns BLOCKED: with the reason and exits 2. It never silently retries over clearnet. This is by design — silent fallback would deanonymize you without warning.

If a fetch is blocked and the user really wants the content, they can either (a) try a different source, (b) use the regular firecrawl skill explicitly (accepting the privacy tradeoff), or (c) wait and retry — Tor exit IPs rotate and another circuit may work.

Limitations

• No JavaScript — curl-based; SPAs and JS-heavy sites won't render. For JS, the user has to use firecrawl (not anonymous).
• Many sites block Tor — Cloudflare, Google, news paywalls, etc. Returns BLOCKED. Try alternative sources.
• Don't authenticate — logging into anything from this skill destroys the anonymity property. Don't do it.
• Slower than direct HTTP — Tor circuit setup adds 5–15s per request, and every request here builds a new circuit (the privacy/speed tradeoff).
• ISP visibility — your ISP can see you're using Tor (not what you're doing). Mitigation (Tor bridges) is out of scope for v1.

How it works (one line)

curl --proxy socks5h://$(openssl rand -hex 8):x@127.0.0.1:9050 ... — the random SOCKS user triggers Tor's IsolateSOCKSAuth and gives every request its own circuit.

See also

• firecrawl-search / firecrawl-scrape — fast, JS-capable, but traceable
• WebSearch — built-in, fast, but goes through Anthropic
• enterprise-search — for internal knowledge bases, not the open web