Jevry
Let Jev select grounded browser actions while language models handle planning and explanation.
Let Jev select grounded browser actions while language models handle planning and explanation.
Jevry sends the user goal and observed page state to Jev. For a simple navigation request, one Choice classifies the intent as SIMPLE_NAVIGATION or OTHER, while a second Choice selects an observed link or NONE. Code validates both answers against the offered IDs before navigation. Its direct website planning path separately asks typed Choice questions about task intent, site scope, and each unparsed request clause.
Classify whether the entire user request requires only opening one existing page, or needs other work or remains uncertain.
The user goal and requested destination, plus an untrusted snapshot of the current page (URL, title, text) and a bounded list of observed links. Link IDs are assigned as link_1, link_2, and so on from accepted page links; their exact values depend on the page. This chart shows only the fixed intent question.
Collect the current URL, title, visible text, and observed navigation links. Accept only safe HTTP(S) link destinations and bound the candidate list.
Send the user goal, requested destination, page snapshot, and observed links. Ask the fixed navigation_intent Choice and a navigation_target Choice whose criteria include NONE and the page-specific link IDs.
Check that returned choices match the offered criteria. Continue only for SIMPLE_NAVIGATION with a valid offered link; then open that selected URL.
The navigation plan distinguishes the user’s intent from the destination and constrains destination choices to links actually observed on the page. A separate direct-read path decomposes request clauses and checks each against explicit supported categories before it forms a plan.