Web search with OpenSERP

OpenSERP returns real results pages from named engines, with original ranks, plus page content as Markdown. Use the MCP tools (search, mega_search, any_search, extract, batch_extract, list_engines). Without MCP, call the HTTP API on http://localhost:7000 (or $OPENSERP_BASE_URL).

Workflow

1. Search one engine first. search with engine: "google" for general queries. Use yandex for Russian-language, baidu for Chinese-language queries. Keep limit at 10 unless you need more.
2. Pick sources by rank and domain. Prefer official docs, release notes and primary sources over aggregators. Skip type: "ad".
3. Read with one batch_extract call for the 2-4 URLs you need, not repeated extract calls. Avoid archive or changelog index pages: they can be hundreds of KB. Pick the single post or section page instead.
4. Answer with links. Cite each claim with the URL you read. Mention the engine and rank when ranking matters (SEO questions).

When a search fails

• captcha detected, blocked or rate_limited from one engine: do not retry the same engine more than once. Use any_search with engines: ["google", "duckduckgo", "bing"], or search on another engine.
• Results that do not match the query (unrelated brands, dictionary pages, a single keyword) mean the engine served a degraded page. Treat them as a failure and switch engine; do not use them as evidence.
• An empty result set is not proof that nothing exists. Rephrase once or try another engine.
• In batch_extract, an item with metadata.error failed; the others are still valid.

HTTP fallback

curl "http://localhost:7000/google/search?text=QUERY&limit=10"
curl "http://localhost:7000/mega/search?text=QUERY&engines=google,duckduckgo&mode=any"
curl -X POST "http://localhost:7000/extract/batch" -H "Content-Type: application/json" \
  -d '{"urls":["https://example.com/a","https://example.com/b"],"mode":"auto"}'

extract=N adds Markdown for up to N organic results (1-5) in one call. Ads and answer rows are skipped.