---
name: meta-search
description: Recovers web content when searches fail with 403, 429, 422 errors, blocked sites, or empty results. Runs a bundled multi-provider script (works without API keys), then falls back to manual strategies.
allowed-tools:
  - Bash(node *)
  - WebSearch
  - WebFetch
---

# Meta Search

Recover web content when standard tools fail. Orchestrates multi-service search and extraction via a bundled script, with manual fallback strategies.

## Recovery workflow

### Step 1: Run the recovery script

```bash
node "${CLAUDE_SKILL_DIR}/scripts/search.mjs" "<query-or-url>"
```

- **Query** → SearXNG (if `SEARCH_PLUS_SEARXNG_URL` is set) → TinyFish → Tavily → Brave → Exa → Jina Search (each only if its key is set) → Firecrawl (keyless). Prints a markdown list of results.
- **URL** → Direct markdown fetch (asks the origin for `Accept: text/markdown`; only used when it actually serves markdown) → Tavily Extract (if key) → Jina Reader → TinyFish Fetch (if key) → Firecrawl → Defuddle → Wayback Machine snapshot. Prints the page as markdown.

Exit code 0: use the stdout directly. Non-zero: stderr lists each service tried and why it failed; proceed to Step 2.

### Step 2: Manual recovery with built-in tools

Apply the strategy matching the error type:

**403 Forbidden**
1. Retry with `web_fetch` using the URL directly
2. Search for the page title or key terms instead
3. Try the archived copy: `https://web.archive.org/web/2/<URL>`

**429 Rate Limited**
1. Wait briefly, then retry
2. Simplify the query to essential keywords
3. Switch approach: if searching, try fetching a known URL instead

**422 Validation / "Did 0 searches..."**
1. Remove special characters and quotes from the query
2. Shorten to essential keywords only
3. Split compound queries into separate searches

**451 / blocked domain**
1. Search with domain exclusion: `"<query> -site:<blocked-domain>"`
2. Search for alternatives: `"<query>" alternative OR mirror`

**ECONNREFUSED / Timeout**
1. Retry after a brief pause
2. Try `web_fetch` with a different URL for the same content

**Empty Results**
1. Broaden the query — remove restrictive terms
2. Try different phrasing or synonyms

### Step 3: Report results

- On success: return the recovered content
- On partial success: return what was found, note what remains inaccessible
- On failure: report which strategies were tried and why they failed

## Sandbox compatibility

The extraction script makes outbound requests to external services. If Claude Code's sandbox is enabled, these domains must be in the `allowedDomains` list:

```jsonc
// ~/.claude/settings.json
{
  "sandbox": {
    "network": {
      "allowedDomains": [
        "api.tavily.com",
        "api.search.brave.com",
        "api.exa.ai",
        "s.jina.ai",
        "r.jina.ai",
        "api.jina.ai",
        "api.firecrawl.dev",
        "api.search.tinyfish.ai",
        "api.fetch.tinyfish.ai",
        "defuddle.md",
        "archive.org"
      ]
    }
  }
}
```

Without these, all extraction services will fail with `fetch failed` and the script will fall through to Step 2 (manual recovery).

Keyless services (Jina Reader, Firecrawl, Defuddle, Wayback) are included in the list above.

The direct markdown fetch (first step for URLs) connects to the target site itself and ignores proxy settings, because its SSRF guard must check DNS answers at connect time. In a proxied sandbox that step always fails (`direct:` in the `Tried:` list) and the chain continues with the providers above; this is expected.

## Limitations

- Cannot bypass CAPTCHA or advanced bot protection
- Some paywalled content remains inaccessible
- Cache/archive services may have stale content
- Without API keys, web search depends on Firecrawl's keyless tier (rate-limited per IP); set any provider key for reliability