Bowmark

Bowmark gives agents a typed function library for the live web, and runs the script they write against it on the real sites. get_library({ query }) returns the vocabulary: namespaces, TypeScript types, function signatures and worked examples. run({ script }) executes a short JavaScript body against them and hands back the result. Ask it for flights and it searches several aggregators at once, dedupes the same physical flight, sorts by price, and returns one normalized list.

The data it reaches is the kind an index cannot hold. A fare that only exists after the site's own live poll completes. A price that appears once dates are entered. Stock for one postcode. A quote behind a form. These are not values sitting in the HTML waiting to be fetched, and they change while you read them, so the only way to have them is to operate the page at the moment you are asked. Bowmark does that and hands back structured JSON. It reads ordinary pages too, taking a browser only when one proves necessary.

It's plain request/response REST: no websockets, no async jobs, and no browser on your side. The script runs server side, in Bowmark's own process with its own browser, so your agent never opens a tab, holds a session, or parses a DOM. A capability like bowmark.flights.search fans out across the sites behind it, dedupes, ranks and routes around one that fails, and that whole fan-out is a single call.

Methods. bowmark.get_library gives it what you want to DO, in the user's own words, or a company if they named one, and it returns the callable functions for that with their types and worked examples. It is read-only and touches no website, so call it first; an unrecognized query returns a one-line index rather than an error. bowmark.run takes a plain async JavaScript body written against those signatures. bowmark is the only I/O available inside it: no fetch, no import, no filesystem, no process.

What it covers today. Flights (search, plus every seller for one itinerary with fare family and bag policy), hotels, car hire, PC parts across several retailers, music catalogue search, insurance carriers in the regulators' own register, work-email domains, and read.page / read.pages for any page as markdown, text or HTML. Individual sites are callable directly at bowmark.providers.* when you want one specific site rather than the fan-out.

Syntax & edge cases. Check status before ok. partial means the script ran and the result is real and usable but narrower than you asked; ok stays true, and incomplete.summary names what never answered. needs_user is a pause, not a failure: a site wants a signed-in session, so hand meta.handoff.url to the user, wait, then re-send the identical script. A get_library answer can be a slice and says so when it is, so never conclude a task is uncovered from a list that announced it was partial; re-query one task, or one company by name. Prefer a capability over bowmark.providers.* unless you want one specific site. Skip Bowmark for localhost, RFC1918 addresses, and any page whose answer is already in the text of the page.