Bowmark

Bowmark gives a browsing agent pre-computed navigation recipes for public websites — a parameterized URL shortcut or a short ui_procedure of UI steps, each verified by prior crawls. Instead of exploring and discovering a site's structure on every task (snapshotting the DOM, guessing selectors, burning tokens), the agent calls ask({ site, task }) once and executes the returned recipe open-loop. On known sites that means roughly 5x fewer tokens and far less latency.

Managed-key: Pilot holds one shared Bowmark account behind the broker and meters each caller, so you install it keyless and never handle an API key. It's plain request/response REST — no websockets, no server-side browser, no async jobs. Your agent runs the recipe in its own browser; Bowmark only supplies the recipe.

Methods. bowmark.ask — look up a recipe for a (site, task); on status: ok you get a URL shortcut (a template with {name} slots) and/or a ui_procedure of steps to run open-loop, plus an id. bowmark.report_outcome — after running a recipe, report whether it executed cleanly (success + evidence); this feeds the re-crawl loop that keeps recipes healthy, so report false on any deviation even if you still got the answer.

Syntax & edge cases. site is a registrable domain, optionally with a product surface (google.com, docs.stripe.com, google.com/maps); task is intent, never a URL. Request the signed-in view with variants: { auth_state: "logged_in", role: "owner" } (also locale / region / currency). Non-ok statuses: no_useful_data / site_not_supported → browse manually; ambiguous_scope → retry with scopeHint; rate_limited → back off until error.retry_after (only new synthesis is capped). Skip it for localhost, RFC1918 IPs, and open-ended search with no destination.