web-search

Web search

Web retrieval runs inside codemode scripts. Search, fetch, and filter in the same script, and return only what the answer needs: titles, URLs, dates, and the decisive passages. Codemode cuts output past 10,000 tokens from the middle, so a script that returns whole results loses part of what it retrieved.

Need	Route
Discover sources	tools.mcp__perplexity__perplexity_search
Read known URLs	tools.mcp__exa__web_fetch_exa
GitHub issues, PRs, code, or repositories	gh search through tools.bash
Login, forms, clicks, or visual state	the agent-browser skill

GitHub allows 30 search requests a minute across gh search commands, so start with two or three targeted queries and read the items they find.

Both tools resolve to an MCP result, including failures, which set isError. Its content[0].text holds the Markdown output. describeTool() lists the parameters this skill leaves out.

Search

Call perplexity_search({ query, max_results, max_tokens_per_page, search_domain_filter, search_recency_filter }). The defaults, 10 results with up to 1,024 tokens of page excerpt each, return about 40,000 characters, so set both to what the question needs; max_tokens_per_page accepts 256 to 2,048.

Put identifiers, versions, dates, and source types in the query, and run independent queries in parallel within one script. Restrict official sources and papers with search_domain_filter, for example arxiv.org and openreview.net, and time-sensitive questions with search_recency_filter. Keep the default web search type, which the benchmarks behind this setup measured.

Results are numbered N. **Title** blocks, each with a URL: line and a page excerpt. Split them and keep the lines that bear on the question:

const terms = /rate limit|429|retry-after/i;
const r = await tools.mcp__perplexity__perplexity_search({
  query,
  max_results: 5,
});
return r.content[0].text
  .split(/\n(?=\d+\. \*\*)/)
  .slice(1)
  .map((result) =>
    result
      .split("\n")
      .filter((line, i) => i < 2 || terms.test(line))
      .join("\n"),
  );

Fetch the page when its excerpt stops short of the decisive passage.

Fetch

Call web_fetch_exa({ urls, maxCharacters }) with every URL in one batch. It returns the pages in one text, each starting with a # Title line and a URL: line, so split it with /\n(?=# .*\nURL: )/. Set maxCharacters to the length the question needs, since the default of 3,000 characters per page truncates long documents.

Fetch each page once. For a long page, store() its text and return its headings, then load() it in later scripts to return the sections the answer needs.

A batch leaves out pages it could not fetch, so match the returned URL: lines against the request. Check a missing page's status code with curl -sL -o /dev/null -w '%{http_code}': report a 404 as gone, and read any other failure with agent-browser. Use agent-browser as well for an empty or shell page, and when the page's current state matters, such as prices, status, or a change published today, since Exa may serve a cached copy.

Establish the answer

Read the decisive passage in the source that owns the claim, in its search excerpt or its fetched page. Excerpts from other sites and extracted fields are leads, not sufficient evidence by themselves.

Cite canonical URLs for material claims. Add another source only when it could change a disputed, time-sensitive, or consequential conclusion. Mark inference, contradictions, and material evidence gaps explicitly.

A negative search establishes only what those queries found. Refine scope or identifiers before concluding that evidence is unavailable. Finish once the inspected evidence adequately answers the request; further searching needs a concrete unresolved question.