ketch — Search, Code Search and Scraping in One Binary

Why This Matters

Giving an agent research tools usually means wiring up several providers, each with its own auth and response shape. ketch puts them behind one stateless binary. An operator sets the backend once, and every agent just calls ketch search or ketch scrape without knowing which provider answered.
It comes from George Dikeakos ( 1broseidon ), who also wrote the tree-sitter code navigation CLI cymbal . It landed in Homebrew core on September 16, 2026 at version 0.16.2. MIT licensed, written in Go. It releases quickly: the details below reflect 0.20.4, released and checked November 3, 2026.

What It Actually Does

  • Web search. ketch search works with no key. The default auto backend falls back through hosted keyless providers, and prefers anything you've configured. --multi queries several at once and fuses the rankings.
  • Code search. ketch code greps public open-source repos through Grep, Sourcegraph, or GitHub.
  • Library docs. ketch docs queries Context7.
  • Scrape and crawl. Clean markdown from HTML and text-based PDFs, with an automatic fallback to headless Chrome for JavaScript-rendered pages.
  • Tags. Bookmarks that group sources by project and survive cache clears.
  • Agent wiring. --json everywhere, documented exit codes, a ketch doctor health check, a bundled skill, and an MCP server ( ketch mcp serve ).

What to Check First

Keyless is convenient, but auto means your queries go to a chain of third-party services, so check which one served a result. The output's backend: field reports it, and setting search_order pins the list.
The README also says private and internal HTTP(S) addresses remain fetchable from the machine running ketch . An agent pointed at an untrusted URL could reach your local network, so that matters if ketch runs near anything internal. And cookie support ( --cookie-file ) comes with a plain note that respecting a site's terms is the operator's responsibility. Authenticated pages also get cached locally unless you pass --no-cache .

Where It Fits in Our Workflow

This overlaps with our crawl-fetch-ingest skill, though less neatly than a list of backends suggests. We haven't run it.
That skill routes each URL by need: Jina Reader for page text, Firecrawl with a JSON schema for structured extraction, Tavily for discovery, and OpenGraph.io for metadata. In ketch , Firecrawl and Tavily are search backends. Its scraper is its own extraction pipeline, and the README describes nothing like Firecrawl's schema-filled extraction, which the skill relies on to pull {name, title, bio, headshot} from a team page. So ketch doesn't replace that step.
Where it could help is the cheaper end. It might stand in for Tavily when we're finding a firm's homepage, or for Jina Reader when we only need the text of a page. Its keyless search could also serve as a backup when a keyed service runs out of credits. The skill's checkpoints, human confirmation gates, and per-firm cache would stay as they are: ketch 's page cache defaults to a 72-hour TTL, which is a different job from keeping a firm's data across decks.
Scanned PDFs need OCR, so image-heavy pitch decks may not extract with the built-in parser. We'd test that before relying on it for a deck.
ketch code is the separate surface worth trying, since it answers "how do other projects use this API?" with no setup at all.

Install

bash
brew install ketch
ketch doctor
ketch scrape https://go.dev/doc/effective_go