search（搜索）

Search — Run a Query, Return Structured Results

For any task that needs "the top N results for X." Read chrome-bridge first.

Why not a search API

The chrome-bridge gives you a real logged-in Chrome session with full cookies, so queries return personalized / region-appropriate results and aren't rate-limited the way a raw API key would be. No API signup, no tokens.

How this skill is laid out

Each capability has a focused JS file under scripts/. To use one:

1. Read the script file to load it as a string.
2. Pass the contents as the code argument to mcp__chrome-bridge__execute_script.
3. The script returns a structured object — the header comment at the top of each file documents pre-conditions and return shape.

Capability map

Workflow

Primary: Google

navigate(url='https://www.google.com/search?q=' + urlencode(query))
# wait ~3s

# Detect & clear EU consent wall if present:
u = execute_script(code="(()=>location.href)()")
if 'consent.google.com' in u:
    dom_click('button[aria-label*="Reject all" i]')   # try the dedicated button first
    # fall through to the text-matching script if that did nothing:
    execute_script(code=<contents of scripts/consent_click.js>)
    # wait ~3s for redirect back to /search

results = execute_script(code=<contents of scripts/google_results.js>)
# If any result has empty snippet and you need them all populated:
results = execute_script(code=<contents of scripts/google_results_fallback.js>)

If Google's primary extractor still returns empty:

• Fall back to page_markdown and pick out result-looking links heuristically
• Fall back to DuckDuckGo (below)

Google search operators

Pass operators in the query string:

• site:example.com — restrict to a domain
• filetype:pdf — restrict to a file type
• "exact phrase" — exact match
• after:2025-01-01 — recency filter
• -word — exclude a term

Fallback: DuckDuckGo HTML endpoint

Useful when Google shows CAPTCHA, consent loops, or you need a no-JS path:

navigate(url='https://html.duckduckgo.com/html/?q=' + urlencode(query))
# wait ~2s
results = execute_script(code=<contents of scripts/ddg_results.js>)
# Or if you need real URLs (not the //duckduckgo.com/l/?uddg= wrapper):
results = execute_script(code=<contents of scripts/ddg_results_decode.js>)

Fallback: Bing

navigate(url='https://www.bing.com/search?q=' + urlencode(query))
# wait ~3s
results = execute_script(code=<contents of scripts/bing_results.js>)

Specialty searches

• Academic: https://scholar.google.com/scholar?q=... — results under .gs_ri, title at .gs_rt a, metadata at .gs_a, snippet at .gs_rs.
• News: https://news.google.com/search?q=... — SPA, may need 5+ seconds to render.
• Images: https://www.google.com/search?q=...&tbm=isch — extract img[src^="http"] with alt text.

Return shape

[
  { "title": "...", "url": "https://...", "snippet": "..." }
]

Always include a snippet even if short — it's what lets the caller rank without re-fetching.

Common failures