nc://web-search-plus

Web Search Plus: multi-provider web search and clean page reading, as two tools the model calls. A Nodecode rewrite of robbyczgw-cla/hermes-web-search-plus by robbyczgw-cla.

The two tools are the original's:

• web_search_plus — search the web. With provider: "auto" the query's shape picks the first engine (Brave for a general question, Serper for security, shopping and news, Exa for documentation and papers) and the next configured engine answers when one fails or finds nothing. Recency (freshness), the news vertical, include_domains/exclude_domains and mode: "research" (search, then read the top pages) are the original's parameters.
• web_extract_plus — read pages as clean text, falling back from Tavily to Firecrawl to Exa to a plain fetch here that strips the HTML. render_js asks Firecrawl first, because it renders JavaScript.

A person can call the same functions from eval: (web-search-plus:search "…") and (web-search-plus:extract "https://…") taking a decoded JSON object; the model learns them from (help :web-search-plus).

Get it

Tell the agent "add the web-search-plus cell", or run nodecode add nc://web-search-plus.

Set it up

One provider key is enough. Put it in the web-search-plus section of the config — in /settings, on the settings page, or by asking the agent to set it — or in the environment variable the key names:

{
  "web-search-plus": {
    // One engine is enough; more give the fallback somewhere to go.
    "brave_api_key": "…",          // BRAVE_API_KEY   https://api.search.brave.com/app/keys
    "serper_api_key": "…",         // SERPER_API_KEY  https://serper.dev/api-key
    "tavily_api_key": "…",         // TAVILY_API_KEY  https://app.tavily.com
    "exa_api_key": "…",            // EXA_API_KEY     https://dashboard.exa.ai/api-keys
    "firecrawl_api_key": "…",      // FIRECRAWL_API_KEY  extraction, and JavaScript pages
    // A self-hosted instance needs no key:
    "searxng_url": "https://searx.example.org",   // SEARXNG_INSTANCE_URL
    "provider": "auto",            // or brave, serper, tavily, exa, searxng
    "extract_provider": "auto",    // or tavily, firecrawl, exa, native
    "max_results": 5,
    "max_chars": 6000,
    "timeout": 20
  }
}

With no key and no searxng_url the folder refuses to start, in one line naming what to set.

Use it

Ask for a search, or for a page to be read. The answer is one JSON object:

{"query": "…", "provider": "brave", "count": 2,
 "results": [{"title": "…", "url": "…", "snippet": "…", "age": "…"}],
 "notes": ["serper: HTTP 401: …"]}

notes carry what happened on the way — which engine was skipped and why, that a domain list was applied here, that an engine has no news vertical — and never replace an answer. mode: "research" adds pages, the first pages' text. web_extract_plus answers pages (url, title, chars, truncated, text) and notes; a page nothing could read is a note, not a failure.

What the rewrite leaves out

The original speaks to fourteen search engines and eight extractors. This cell keeps the ones whose API is one HTTP request — Brave, Serper, Tavily, Exa, SearXNG for search, Tavily, Firecrawl, Exa and a plain fetch for extraction — and leaves out You.com, Linkup, Parallel, Keenable, Serpbase, Querit, TinyFish, the Firecrawl search vertical and the DonSeTch CLI. Also left out: the search cache, the quality report, the routing diagnostics and health statistics, the semantic-span ranking, markdown output (every page is text), and the native Hermes web-provider bridge, which has no meaning outside Hermes.

Tests

cd cells && NODECODE_SRC=… .github/test-cell.sh tools/nodecode-web-search-plus

Every provider is a stubbed nlk:http, so the suite needs no network and no key: the lifecycle, a missing key refusing, a search per engine, the fallback, the routing, the domain filter, research mode, the extraction chain, a page cut to its limit, and a url scheme that is never fetched.

Licence

MIT, as the original. See LICENSE and NOTICE.