web-search（网络搜索）

Web Search & Scrape

You have three tools for web access. Use them in combination based on what the task needs.

The Stack

SearXNG — Search Engine

Local meta-search aggregating 25+ engines (Google, Bing, DuckDuckGo, Brave, etc). No tracking, no rate limits, JSON API.

# Basic search
curl -s "http://localhost:8888/search?q=QUERY&format=json" | python3 -c "
import json, sys
data = json.load(sys.stdin)
for r in data.get('results', [])[:10]:
    print(r.get('title', ''))
    print(r.get('url', ''))
    print(r.get('content', '')[:200])
    print()
"

Category search — append &categories= with: general, news, images, files, science, it, music, videos

# News search
curl -s "http://localhost:8888/search?q=QUERY&format=json&categories=news"

# Multiple categories
curl -s "http://localhost:8888/search?q=QUERY&format=json&categories=science,it"

Pagination — append &pageno=2 (or 3, 4, etc) for more results.

Lightpanda — Fast Headless Fetch

Built in Zig. 10x faster than Chrome, tiny memory footprint. Use this as the default for fetching page content.

# Fetch as markdown (best for reading/summarizing)
lightpanda fetch --dump markdown https://example.com

# Fetch as HTML (when you need structure)
lightpanda fetch --dump html https://example.com

# Semantic tree (useful for understanding page layout)
lightpanda fetch --dump semantic_tree https://example.com

# Strip unnecessary elements
lightpanda fetch --dump markdown --strip_mode js,css https://example.com

# Include iframe content
lightpanda fetch --dump markdown --with_frames https://example.com

Agent-Browser — Full Browser Automation

Playwright-based. Use when Lightpanda can't handle the page (JS-heavy SPAs, login-required pages, dynamic content, form interactions).

# Open and snapshot
agent-browser open https://example.com
agent-browser wait --load networkidle
agent-browser snapshot -i

# Get text content
agent-browser get text body

# Interact with elements
agent-browser fill @e1 "search query"
agent-browser click @e2

# Screenshot for visual inspection
agent-browser screenshot --annotate

# Always close when done
agent-browser close

Decision Guide

Need to find something? → SearXNG first. Always.

Need page content? → Lightpanda. It's fast, it returns clean markdown, and it handles 90% of pages.

Lightpanda returns garbage or empty content? → The page probably needs JavaScript to render. Switch to Agent-Browser.

Need to log in, fill forms, click through flows? → Agent-Browser. Save auth state for reuse:

agent-browser state save auth.json
# Later:
agent-browser state load auth.json

The web-search CLI

There's also a unified CLI at ~/.agents/tools/web-search (also available as web-search on PATH) that chains these together:

# Search only
web-search "hospice compliance CMS 2026"

# Search + scrape top results
web-search "hospice compliance CMS 2026" --scrape -n 3

# Fetch a single URL
web-search --fetch https://example.com

# Use Agent-Browser for JS-heavy pages
web-search --fetch https://spa-app.com --browser

# News search + scrape
web-search "CMS hospice updates" --categories news --scrape

Common Patterns

Research a topic

# 1. Search
curl -s "http://localhost:8888/search?q=topic+here&format=json" > /tmp/results.json

# 2. Review results, pick the best URLs

# 3. Fetch the good ones
lightpanda fetch --dump markdown https://good-result.com

Get current/breaking info

# News category + recent results
curl -s "http://localhost:8888/search?q=topic&format=json&categories=news"

Deep scrape multiple pages

# Search, extract URLs, fetch each
curl -s "http://localhost:8888/search?q=topic&format=json" | \
  python3 -c "import json,sys; [print(r['url']) for r in json.load(sys.stdin)['results'][:5]]" | \
  while read url; do
    echo "=== $url ==="
    lightpanda fetch --dump markdown "$url" 2>/dev/null
  done

Handle a stubborn JS-heavy page

# Lightpanda returned nothing useful? Switch to agent-browser
agent-browser open https://stubborn-spa.com
agent-browser wait --load networkidle
agent-browser get text body > /tmp/page-content.txt
agent-browser close

Important Notes

• SearXNG runs at http://localhost:8888. If it's down, check: docker ps | grep searxng and restart with docker start searxng
• Lightpanda is at /opt/homebrew/bin/lightpanda
• Agent-Browser is at /opt/homebrew/bin/agent-browser (v0.21.1)
• The web-search CLI is at ~/.agents/tools/web-search and symlinked to /opt/homebrew/bin/web-search
• When SearXNG returns results, the content field has a snippet — often enough to answer simple factual questions without fetching the full page
• For URL encoding in curl, use python: python3 -c "import urllib.parse; print(urllib.parse.quote('my query'))"

Bundled Resources

This skill includes everything needed to rebuild or troubleshoot the stack:

• scripts/web-search — The unified CLI script (also installed at ~/.agents/tools/web-search)
• references/infrastructure.md — Full infrastructure docs: binary locations, SearXNG API reference, container management, OrbStack setup, troubleshooting guide. Read this if something breaks or you need to reconfigure.
• references/searxng-settings.yml — SearXNG config (engines, formats, API settings). Edit and copy to ~/.agents/searxng/config/settings.yml then docker restart searxng to apply changes.

Related Skills

• [[agent-browser]] — full browser automation for JS-heavy pages and form interaction
• [[human-browser]] — stealth browsing with residential proxies for bot-protected sites
• [[seo]] — SEO audits and optimization that complement web research