Web search router

Skill phntm7/skills/skills/web-search-router

Use before web search, extraction, crawling, entity lookup, or monitoring when Linkup, Parallel, and Tavily are available; it selects the best engine by capability, reliability, and cost.From its SKILL.md

Install

npx -y skills add phntm7/skills --skill web-search-router

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

6.0 KB, as published. Nobody here has run it

Web search CLI router

Last verified: 2026-08-05. Routing rules distilled from a 25-query benchmark (5 categories × 3 engines) plus capability tests — evidence in references/benchmark-findings.md.

Three CLIs, three specialists. Per-command usage lives in the sibling skills: linkup-search, parallel-search, tavily-search.

Linkup — answer engine: composed cited answers, LinkedIn extraction, single-call sequential deep search, structured JSON at search tier.
Parallel — structured-data platform: primary-source-biased index, licensed entity data (Tracxn), tabular enrichment, monitors, cheapest research tiers.
Tavily — content-acquisition toolkit: news/finance topic pipelines, crawl/map, query-focused extraction, image search.

On plain "search the web" the three are near-equal (benchmark totals 40 / 39.5 / 43 of 50) — routing matters at the edges, not the average.

Routing table

Task	Use	Command sketch
Quick fact to ground a reply	Parallel turbo ($0.001) or Linkup fast (composed answer, $0.006)	`parallel-cli search -q "..." --mode turbo`
General agent search step	Parallel basic (programming) / Tavily (docs lookup) — close call, any works	see sibling skills
News, current events	Tavily — first choice (5/5 in our single-run benchmark)	`tvly search "..." --topic news --time-range week`
Regional news (e.g. Ukraine-local)	Tavily news topic without `--country`, query in the local language; cross-check with Linkup/Parallel general search	locale drift is Tavily's weak spot
Live numbers: rates, prices	Tavily or Linkup (numbers appear in snippets); NOT Parallel raw search	`tvly search "..." --topic finance --include-answer advanced`
LinkedIn profile / posts / comments	Linkup only	`linkup search "{linkedin_url} Return the profile details."`
Entity lists, firmographics, people search	Parallel (licensed data)	`parallel-cli findall entity-search "..." -t companies`
Enrich CSV/JSON rows with web data	Parallel	`parallel-cli enrich run --data '...' --intent "..."`
Crawl a docs site / map site URLs	Tavily	`tvly map URL` → `tvly crawl URL --select-paths ...`
Find page → scrape → extract, one call	Linkup deep ($0.055)	`linkup search "First find... Then scrape... Then return..." --depth deep`
Specific fact inside a big/JS page	Tavily extract `--query` or Parallel fetch `--objective`	`tvly extract URL --query "..."`
Full page as markdown (incl. PDF)	any — all three parse PDFs well	cheapest: Linkup fetch $0.001
Structured JSON, one-shot	Linkup `--schema` (search tier, ~3 s, $0.006)	fresher but slower: Parallel enrich
Image search	Tavily (`--include-image-descriptions`)	Linkup weaker, Parallel none
Continuous monitoring / webhooks	Parallel monitor (only option); recurring cost — confirm with user first	`parallel-cli monitor create "..." --frequency 6h`
Ukrainian / regional queries	Linkup (live trackers, Telegram) or Parallel (official service sites); query in the local language	Tavily third — locale drift
Offloaded research, scoped question	Parallel `core-fast` ($0.025, ~1 min) or Tavily `mini` ($0.03–0.88 dynamic, best source discipline)	Linkup research is 10× pricier at entry
Offloaded research, broad report	Parallel `pro`/`ultra` or Tavily `pro`; Linkup `investigate`/`research` for EU/LinkedIn-heavy topics	confirm cost with user first
Bulk/batch lookups	Linkup tasks (CLI, 100/batch); Parallel's Task Group API is not exposed by `parallel-cli` (its CLI bulk path is `enrich` for tabular rows)	Tavily CLI has no batch

Failure modes — route around these

Tavily is keyword-literal: ambiguous entities fail hard ("Rust async traits" → the Rust video game; "Нова пошта" → Wikipedia "Nova"). Always disambiguate in the query ("Rust language ..."). It can also drift to wrong-country sites on regional queries and returns duplicate content across mirror domains. No publish dates outside news topic.
Parallel's --after-date is a hard filter over sparse metadata — sparse queries return zero results (observed 3×). For recency, omit it and sort by publish_date yourself (often null on general results). Its excerpts often point at the page with the number (quote pages) without quoting the number — follow up with fetch. Weakest on Ukrainian/regional content.
Linkup's news-ish results are LinkedIn posts — fresh (24–48 h) and they do carry the story, but they're social posts, not citable sources. Fine for "what's the buzz", wrong for citations. Raw search-results can also surface years-stale pages for live data — use sourced-answer there. Index favors content sites/blogs over primary docs.

Cost cheat sheet (per call)

	Linkup	Parallel	Tavily
Search	$0.005–0.006 (deep $0.05)	$0.001–0.005 incl. 10 results, +$0.001/extra result	$0.008 (advanced $0.016)
Fetch/extract	$0.001–0.005	$0.001 per URL	~$0.0016/URL
Research	$0.25–2.50	$0.005–2.40	$0.03–2.00 (dynamic)
Free tier	$20/mo credits	$5/mo + 5k req	1,000 credits/mo

Patterns

Escalate stepwise: cheap search → advanced/deep search → research at the lowest sufficient tier. Never open with research for something a half-cent search answers.
Cross-check load-bearing facts across two engines (~1¢ total) — their indexes differ enough that agreement is meaningful.
Combine specialists: Tavily news headline → Parallel entity data on the company → Linkup LinkedIn posts for the people involved.
Only one CLI installed? Its sibling skill stands alone — this router only matters when there's a choice.