Broad web search across the SearXNG instance's full enabled engine pool.

Use this for general-purpose queries when you don't have a specific source preference. Returns a merged, deduplicated result list from across this instance's enabled engines — actual yield depends on which engines respond at query time.

Knobs: • pages=N (max 5) — fetch multiple pages and merge for more results in one call. • format='full' — full result objects with metadata; default 'compact' is token-efficient. • time_range — narrow to recent results (note: not all engines support this — see field doc).

When to reach for a different tool: • search_on_engines — you want specific sources (e.g. just ArXiv + PubMed) • search_by_category — you want all engines in a category (e.g. all science)

Search using ONLY the specified engines (e.g. engines: ["arxiv", "duckduckgo"]).

Use when you have a specific source preference instead of the broad search tool.

NOTE: this instance's live engine/category list is currently unavailable — SearXNG was unreachable when the server started. Names you pass are forwarded and validated once SearXNG is reachable; restart the server after SearXNG is up to see the enabled list here.

Multi-page fanout (pages=N) and other knobs work the same as the broad search tool.

Search within one or more categories — runs every engine tagged with each (e.g. categories: ["science"]).

Use when you want broad coverage of a content type without enumerating engines.

NOTE: this instance's live engine/category list is currently unavailable — SearXNG was unreachable when the server started. Names you pass are forwarded and validated once SearXNG is reachable; restart the server after SearXNG is up to see the enabled list here.

Multi-page fanout (pages=N) and other knobs work the same as the broad search tool.

Fetch a URL and convert its HTML content to clean Markdown.

Use after search (or its variants) when you have a URL and want the actual page text, not just the search snippet. Lightweight HTTP + HTML→Markdown — handles ~80% of the static-HTML web (Wikipedia, docs sites, blogs, news, GitHub READMEs).

What this DOES NOT handle: • JavaScript-rendered pages (React/Vue/Angular SPAs) — content loads after the initial HTML, which we don't execute. Returns minimal or empty markdown for these. • Bot-protected pages (Cloudflare challenge, captcha) — typically fail with HTTP 403/503. • Binary resources (PDF, images, archives) — returns an explanatory hint instead of garbled bytes.

For those cases, fall back to a Chromium-backed reader (e.g. Crawl4AI exposed via the SearXNG-Compose reader profile).

Token-efficient extraction modes (priority order — first one set wins): • readHeadings:true — returns ONLY the heading list (hierarchical TOC). Cheapest survey of a long page. • section:'' — returns content under first matching heading, up to the next same-or-higher heading. Use after readHeadings to jump. • paragraphRange:'3-7' — 1-indexed paragraph slice; supports 'N' (single) or 'N-M' (range). • startChar + maxLength — character window pagination. Response includes total_length and truncated so you can plan follow-up calls.

If no extraction mode is set, returns the full Markdown.

Recommended workflow for long pages:

1. web_url_read(url, readHeadings: true) ← TOC scan

2. web_url_read(url, section: '') ← targeted read Far more token-efficient than fetching the full page up front.