Search the web AND read the top sources in one shot.

Runs SearXNG metasearch, dives into the top max_sources pages with a real browser, distills each page down to passages relevant to query, and returns one merged digest. Best first call for any research question. Args: query: research question or keywords (RU/EN both fine) max_sources: how many top hits to read in full (1-6) max_results: how many search hits to consider (1-20) per_source_chars: per-source character budget (300-8000) time_range: "day" | "week" | "month" | "year" category: searxng category, e.g. "general", "news", "science", "it" language: result language, e.g. "ru", "en", "ru-RU" refresh: ignore cache and re-fetch search results and pages

Search the web via SearXNG metasearch; return a compact ranked link list.

Returns title, URL and a short snippet per hit — no page content. Empty or blocked results are retried automatically with other engine sets. To actually read pages, call read_url; to do both at once, call deep_research. Args: query: search query (natural language or keywords) max_results: 1-20 time_range: "day" | "week" | "month" | "year" category: e.g. "general", "news", "science", "it", "files" engines: comma-separated engine names, e.g. "google,bing,duckduckgo" language: e.g. "ru", "en", "ru-RU" refresh: ignore cache and re-run the search as_json: return pure machine-readable JSON {query, count, hits[], answer?} instead of the human-friendly list (no footer line)

Read one web page; return its main content as clean, budgeted markdown.

JS-rendered pages are handled by a real headless browser. Boilerplate (nav, footer, ads) is stripped; if query is given, only passages relevant to it are returned. Pages are cached — re-reads with a different query are instant and cost no network. Args: url: absolute http(s) URL query: optional focus; return only passages relevant to it max_chars: output character budget (300-50000) refresh: ignore cache and re-fetch the page find: search the cached RAW text for this exact substring (case-insensitive): returns matches with counters and context, no network needed. Requires the page to have been read before; combine with query for first reads.

Fetch up-to-date official documentation for a library and distill it under your question.

Context7-style, but local and unlimited: the docs site is resolved from a built-in index (or one live web search), fetched from the primary source, and distilled to passages relevant to query. Repeated questions about the same library are instant, offline and free (raw-page cache). Args: library: library name, e.g. "fastapi", "react", "postgresql", "crawl4ai" query: your concrete question about the library (used for distillation) max_chars: output character budget (300-20000) refresh: re-fetch the docs page even if cached subpages: when the docs home is navigational, follow this many same-site subpages ranked by query relevance (0 disables) version: pin docs to this version (tag, e.g. "0.115.0", "v3", branch name). Works for GitHub-backed libraries: docs come from that exact tag on raw.githubusercontent.com. Doc sites are shown at their latest with an honest note; wrong/missing tag on GitHub also falls back to latest with a note in the answer

Read several known web pages in one call under one shared character budget.

Pages are fetched in parallel (JS-rendered, boilerplate-stripped) and the combined total_chars budget is split evenly between the pages that came back. Prefer this over N read_url calls when you already hold the URLs: one round-trip, one budget, and a failed page costs one line instead of a failed call. Args: urls: 1-10 absolute http(s) URLs; duplicates (after utm/fragment cleanup) are merged, extras beyond 10 are reported in a Skipped line query: optional focus; each page is distilled to passages relevant to it total_chars: combined output budget across all sections (300-30000) refresh: ignore cache and re-fetch every page

Deterministically verify a claim against web sources; return a verdict with quoted passages.

No LLM involved: sources are read (yours via urls, or found by a web search on the claim), distilled under the claim, and scored lexically — polar markers (with negation handling) decide SUPPORTED / CONTRADICTED / UNCLEAR / MISSING-EVIDENCE. Best for checking a fact, assertion or rumour when you need a reproducible verdict with citations, not a narrative. Args: claim: the statement to verify, in your own words (RU/EN both fine) urls: optional 1-10 http(s) URLs to check against; without them the sources are found by a web search on the claim max_sources: how many sources to consider (1-10) refresh: ignore cache and re-fetch the search results and pages