Web Search Plus

Web Search Plus

Web Search Plus gives you two MCP tools. Both return sources, never answers. You read the sources and synthesize; the server does not generate truth claims.

• web_search — multi-provider search with automatic routing. Returns ranked results with URL, title, snippet, provider, and routing metadata.
• web_extract — fetches one or more URLs and returns bounded markdown (or HTML), with optional query-ranked spans for long pages.

When to prefer these over the built-in search

• The user wants to see which sources say something, or wants a quality/routing report.
• The query is provider-shaped: docs/API lookups, arXiv, CVE/security, GitHub/OSS discovery, local/shopping, news with a recency window.
• The user wants a specific provider or a specific locale.
• You need the full text of a page, not a snippet. Use web_extract.
• The user says "wsp", "web search plus", or names a provider.

For quick, one-line factual lookups the built-in search is fine. Do not run both for the same question unless the user asks to compare.

web_search

Required: query.

Useful parameters:

• provider — auto (default) or one of serper, brave, tavily, linkup, exa, firecrawl, parallel, you, searxng, keenable, serpbase, querit, donsetch, octen, tinyfish. Only providers with a configured key work; leave auto unless the user asks.
• count — results, default 5, max 20.
• freshness — day, week, month, year for recency.
• search_type — news for a news vertical.
• country / language — locale overrides.
• include_domains / exclude_domains — allow or deny lists.
• mode — research queries up to three providers concurrently, dedupes, and extracts the top sources. Slower and costs more; use for deep questions.
• quality_report — true to include routing class, provider scores, and authority signals. Useful when the user asks why a result ranked where it did.

Example:

{ "query": "Hermes Agent 0.21 release notes", "freshness": "month", "count": 5 }

web_extract

Required: urls (list).

Useful parameters:

• provider — auto (default) or tavily, exa, linkup, parallel, firecrawl, you, keenable, serper, donsetch.
• format — markdown (default) or html.
• render_js — true for JavaScript-heavy pages when the provider supports it.
• spans + spans_query — return the passages most relevant to a query instead of the whole page. Use this for long documents.

Extraction refuses private or internal targets (loopback, RFC1918, cloud metadata hosts). That is deliberate; do not try to work around it.

Example:

{ "urls": ["https://example.com/docs/setup"], "spans": true, "spans_query": "environment variables" }

Working pattern

1. web_search with a precise query. Read the result list, not just the top hit.
2. web_extract the one or two URLs that actually answer the question.
3. Cite the URLs you used. Say when sources disagree.
4. If web_search returns an error that no provider is configured, tell the user to run uvx --from web-search-plus-mcp==4.1.1 web-search-plus-mcp status and set at least one provider key (see Setup). Do not guess results.

Setup (first run)

The MCP server starts through uvx, so uv must be installed. Provider keys are read from the environment; the plugin ships none.

Recommended starter set: You.com + Serper + Linkup.

export YOU_API_KEY=...      # fast current and multilingual search
export SERPER_API_KEY=...   # Google-style facts, news, local, shopping
export LINKUP_API_KEY=...   # source-grounded results and clean extraction

Optional: TAVILY_API_KEY, EXA_API_KEY, FIRECRAWL_API_KEY, PARALLEL_API_KEY, BRAVE_API_KEY, KEENABLE_API_KEY, SEARXNG_INSTANCE_URL.

Check what is configured:

uvx --from web-search-plus-mcp==4.1.1 web-search-plus-mcp status

Write a starter .env template:

uvx --from web-search-plus-mcp==4.1.1 web-search-plus-mcp setup --preset starter

Full provider list, routing rules, and tool reference: https://github.com/robbyczgw-cla/web-search-plus-mcp