---

name: local-web-search description: > Free, private, real-time web search for OpenClaw — zero API keys required. Powered by self-hosted SearXNG + Scrapling anti-bot engine. Multi-engine parallel search (Bing/DuckDuckGo/Google/Startpage/Qwant), intent-aware Agent Reach query expansion, three-tier Browse/Viewing (Fetcher → StealthyFetcher → DynamicFetcher for Cloudflare/JS sites), cross-engine anti-hallucination validation, and automatic public fallback. homepage: https://github.com/wd041216-bit/openclaw-free-web-search metadata: clawdbot: emoji: "🔍" requires: env: [] files: ["scripts/*"]

Local Free Web Search v3.0

Use this skill when the user needs current or real-time web information. Powered by Scrapling (anti-bot) + SearXNG (self-hosted search). Zero API keys. Zero cost. Runs entirely locally.

---

External Endpoints

Endpoint	Data Sent	Purpose
http://127.0.0.1:18080 (local)	Search query string only	Local SearXNG instance
https://searx.be (fallback only)	Search query string only	Public fallback when local SearXNG is down
Any URL passed to browse_page.py	HTTP GET request only	Fetch page content for reading

No personal data, no credentials, no conversation history is ever sent to any endpoint.

---

Security & Privacy

• All search queries go to your local SearXNG instance by default — no third-party tracking
• Public fallback (searx.be) is only used when local service is unavailable, and only receives the raw query string
• browse_page.py makes standard HTTP GET requests to URLs you explicitly pass — no data is posted
• Scrapling runs entirely locally — no cloud API calls, no telemetry
• No API keys required or stored
• No conversation history or personal data leaves your machine

Trust Statement: This skill sends search queries to your local SearXNG instance (default) or searx.be (fallback). Page content is fetched via standard HTTP GET. No personal data is transmitted. Only install if you trust the public SearXNG instance at searx.be as a fallback.

---

Model Invocation Note

This skill is invoked autonomously by the agent when a query requires live web information. You can disable autonomous invocation by removing this skill from your workspace. The agent will only use this skill when it determines real-time information is needed.

---

Tool 1 — Web Search

python3 ~/.openclaw/workspace/skills/local-web-search/scripts/search_local_web.py \
  --query "YOUR QUERY" \
  --intent general \
  --limit 5

Intent options (controls engine selection + query expansion):

Intent	Best for
general	Default, mixed queries
factual	Facts, definitions, official docs
news	Latest events, breaking news
research	Papers, GitHub, technical depth
tutorial	How-to guides, code examples
comparison	A vs B, pros/cons
privacy	Sensitive queries (ddg/startpage/qwant only)

Additional flags:

Flag	Description
--engines bing,duckduckgo,...	Override engine selection
--freshness hour|day|week|month|year	Filter by recency
--max-age-days N	Downrank results older than N days
--browse	Auto-fetch top result with browse_page.py
--no-expand	Disable Agent Reach query expansion
--json	Machine-readable JSON output

---

Tool 2 — Browse/Viewing (read full page)

python3 ~/.openclaw/workspace/skills/local-web-search/scripts/browse_page.py \
  --url "https://example.com/article" \
  --max-words 600

Fetcher modes (use --mode flag):

Mode	Fetcher	Use case
auto	Tier 1 → 2 → 3	Default — tries fast first
fast	Fetcher	Normal sites
stealth	StealthyFetcher	Cloudflare / anti-bot sites
dynamic	DynamicFetcher	Heavy JS / SPA sites

Returns: title, published date, word count, confidence (HIGH/MEDIUM/LOW), full extracted text, and anti-hallucination advisory.

---

Recommended Workflow

1. Run search_local_web.py — review results by Score and [cross-validated] tag
2. Run browse_page.py on the top URL — check Confidence level
3. If Confidence is LOW (paywall/blocked) — retry with --mode stealth or try next URL
4. Answer only after reading HIGH-confidence page content
5. Never state facts from snippets alone

---

Rules

• Always use --intent to match the query type for best results.
• When local SearXNG is unavailable, both scripts automatically fall back to searx.be.
• If the fallback also fails, tell the user to start local SearXNG:

cd "$(cat ~/.openclaw/workspace/skills/local-web-search/.project_root)" && ./start_local_search.sh

• Do NOT invent search results if all sources fail.
• search_local_web.py and browse_page.py are complementary: search first, browse second.
• Prefer [cross-validated] results (appeared in multiple engines) for factual claims.
• For sites behind Cloudflare or requiring JS, use browse_page.py --mode stealth.