← All examples Levirge Search · examples

Agent-friendly content: search and direct fetch

A captured run from question to cited answer on an open site: ranked structured search results, the page as clean markdown capped with max_bytes, and the answer with quotable sources — next to what an ordinary tool returns instead.

1

The research question

"What's new in Deno 2.9 — anything that changes how we ship?" An open, unprotected site this time — the difference isn't access, it's what reaches the agent. Every capture on this page is from 8 Aug 2026, from Claude Code; payloads verbatim, long bodies excerpted.

2

The ordinary tool — a paraphrase, not the page

The host's native fetch works on open sites, but it answers with a small model's paraphrase of the page: nothing to quote, no control over length, and the original wording is gone before the agent sees it. On a protected site the same call returns nothing at all — the blocked portal.

native WebFetch paraphrase · uncited
// host's built-in fetch — captured verbatim, excerpt
WebFetch("https://deno.com/blog/v2.9", "What's new in Deno 2.9?")

# What's New in Deno 2.9
**Desktop Applications**
The release introduces `deno desktop`, an
experimental feature for building native
desktop apps … compiles to a single binary
with no Electron boilerplate …
**Performance Gains**
Startup time has been halved (17.3ms vs 34.2ms) …

✗ a restatement — no verbatim source to
  quote or check, no size control
3

Find the sources — structured results

search returns ranked results as structured JSON — title, URL, snippet — that a model picks from programmatically. No result page to scrape, no markup to wade through.

search · find the announcement ranked results
// find sources
search("Deno 2.9 new features release notes")

[ { "title": "Deno 2.9 | Deno",
    "url": "https://deno.com/blog/v2.9",
    "content": "Deno 2.9 is here, headlined by
      deno desktop, a new way to build native
      desktop applications…" },
  { "title": "Releases · denoland/deno - GitHub",
    "url": "https://github.com/denoland/deno/releases",
    "content": "2.9.4 / 2026.07.23 · feat(desktop):
      enable --hmr for React Router…" },
  … more results ]
4

Read the top hit — the page itself, capped

fetch returns the page as clean markdown: the content without the nav, cookie banners or markup soup. The article arrives capped by max_bytes — server-side, with the truncation declared. An open site takes the [via direct] fast path; a protected site gets the full browser, same call shape.

fetch · top result, max_bytes [via direct]
// read it, capped at 4 KB
fetch("https://deno.com/blog/v2.9", max_bytes: 4000)

[via direct]
# Deno 2.9
June 25, 2026 — Bartek Iwańczuk

Deno 2.9 is here, headlined by `deno desktop`…
no Electron boilerplate and a single binary
at the end… `deno install` now reads npm,
pnpm, yarn, and Bun lockfiles directly…

[truncated at max_bytes=4000 of 45127]
The truncation marker is verbatim: 4 KB delivered of a 45 KB article. Long reads never flood the agent's context.
5

The answer the agent can now give

The quoted phrases are the article's own words — checkable against the source — not a model's restatement of them.

agent answer cited
Deno 2.9's headline is deno desktop — native desktop
apps from a Deno project, "no Electron boilerplate and
a single binary at the end", with --hmr hot reload in
dev. Migration got easier too: deno install now reads
npm, pnpm, yarn, and Bun lockfiles directly.

Source: deno.com/blog/v2.9 (Bartek Iwańczuk,
25 Jun 2026), read 8 Aug 2026 [via direct] — quotes
verbatim from the article; 4 KB of 45 KB was enough.
6

Measured improvement, and where this came from

Context: 4 KB (~1k tokens) delivered of a 45,127-byte article, cap declared in the payload — never a silent truncation. Citability: verbatim markdown means the answer quotes the source; a paraphrase can't be checked against anything.

provenance 8 Aug 2026
client      Claude Code over MCP
route       [via direct] — open site, no browser needed
cross-check same URL returned 44 KB [via direct] in the
            queued-reads run the same morning
retention   repeat reads within 24 h come from cache