Get answers, not links: search + read the top pages in one API call

The usual agent search loop is three round-trips: search, pick a result, scrape it — and pay (or wait) three times. /research collapses the whole job into one call: it searches the live web and reads the top pages as clean markdown.

One call, content out

curl "https://superhighway.walls.sh/research?q=what+is+the+x402+payment+protocol&pages=2"

The response carries both the ranked results and the pages, already read:

{
  "query": "what is the x402 payment protocol",
  "count": 10,
  "results": [ { "title": "...", "url": "...", "description": "..." }, ... ],
  "pages": [
    { "url": "https://x402.org", "title": "x402", "markdown": "# x402\n...", "length": 8421 },
    { "url": "https://github.com/...", "title": "...", "markdown": "...", "length": 12044 }
  ]
}

pages is 1–3 (default 2). Price is $0.005 per call — via x402 pay-per-call or an API-key plan.

Restrict reading to specific sources

Add domains= to tell /research to read pages only from a list of trusted sources:

curl "https://superhighway.walls.sh/research?q=drug+interaction+aspirin+ibuprofen&domains=nih.gov,pubmed.ncbi.nlm.nih.gov"

The domains= parameter is a comma-separated list of domain names (e.g. nih.gov,pubmed.ncbi.nlm.nih.gov). When set, /research prefers pages from those domains over the general ranking — if no matching pages appear in the results, it falls back to the full pool so the call never returns empty.

Useful patterns:

• Medical research — domains=nih.gov,pubmed.ncbi.nlm.nih.gov,mayoclinic.org
• Legal research — domains=law.cornell.edu,supreme.justia.com
• Academic / peer-reviewed — domains=scholar.google.com,arxiv.org,semanticscholar.org
• Developer docs — domains=docs.python.org,developer.mozilla.org

It delivers readable pages, not just "top N URLs"

On the open web, the literal top N results are often unreadable for a machine: video players, social walls, login gates, pages that time out. /research is built around the contract an agent actually needs — N readable pages from the top of the ranking:

• Unreadable hosts are skipped — video platforms, social networks, and app stores never consume a paid page slot.
• Failed or thin pages are backfilled — if a chosen page 403s, times out, or returns a near-empty shell, the next readable candidate is read instead. Pages with real content come first in the response.
• Latency is budgeted — each page read races a fixed time budget, so a slow site can't stall the call. Typical responses land in 2–4 seconds.

Per-page failures are reported in an error field rather than failing the call, so the agent always gets whatever the web could deliver, transparently.

When to use which tool

All five ship in the MCP server too — one command to install. Full parameter reference: /openapi.json.