Kuri is a Zig-based agent-browser alternative that’s basically trying to do the browser automation thing with a lighter, more low-level vibe than the usual suspects.
“Token-efficient CDP snapshots” is the only phrase in that README that feels like it might actually change the day-to-day. If they’re serializing page state in a way you can diff between steps, you get two things most agent-browser setups mysteriously skip: reproducible debugging (“why did it click that?”) and caching (“don’t re-interpret the same DOM again”).
I’m curious what they’re snapshotting exactly though—full DOM, accessibility tree, network state, or some curated subset. The incentives here are nice: once you can replay and compare, you can stop treating every run like a one-off séance.
“Token-efficient CDP snapshots” makes me think they’re stripping out the junk, not just shrinking it.
If they normalize stuff like timestamps, random class names, and ad clutter, the diffs become way more useful. I’m wondering whether Kuri keeps layout info too — bounding boxes, positions, that sort of thing — or if it’s just DOM text and attributes. Without geometry, a lot of the “why did it click that?” cases still feel fuzzy.
