Most AI search calls an API and summarises whatever it is handed. Palette opens each page in a real tab, in a real engine, with JavaScript running, then reads what actually rendered. Pages that block scrapers work here. A PDF gets looked at rather than skipped.
The tabs it uses sit off screen, out of your way. They never enter your history or your session, and they are closed when it is done.

Measured on a real run, 24 September 2026: four planned queries, 27 pages found, six read, two of them PDFs, 53 citations against six sources.

No account. No sync. No telemetry. Your history, bookmarks and open tabs are files on your Mac; passwords live in the macOS keychain, encrypted by the system.
The research agent thinks on Palette's own models, served from Palette's own infrastructure. Your question and the pages it read are never handed to a third-party search API or a third-party model.
Trackers and ad networks are stopped at the network layer, so there is nothing to render and nothing to slow you down.
Click a cookie banner or a newsletter overlay and it is gone, on that site, before the page draws a frame.
One keystroke strips a page down to the article.
Paste a Chrome Web Store link. They run on WebKit's own extension engine, the one Safari uses.
Across the top or down the left. Pin the ones you keep open all day and they shrink to their icon.
Lift a video out of the page into a window that stays above everything, including other apps.
Free. About 8 MB. Opens instantly, because the engine is already on your Mac.
$ python3 research/agent.py "KMMLU와 CLIcK은 무엇을 측정하나?" planning · KMMLU 한국어 벤치마크 누가 만들었나 · CLIcK Korean benchmark paper authors searching ddg: 10 results naver: 12 results 27 distinct pages found reading read 2,679 chars KMMLU: Measuring Massive… read 83,514 chars KMMLU paper read 0 chars arxiv.org/pdf/2507.08924 mostly pictures, looking at it writing from 6 sources report ready · 53 citations