Kuri is a new browser automation tool built in Zig, aimed squarely at AI agents. Created by systems engineer justrach (known for work on the Mach game engine), it ships as a 464 KB binary with a roughly 3ms cold start. Compare that to Vercel's agent-browser, which weighs 6 MB as a CLI binary and needs 33 MB of npm packages to install. Kuri doesn't use Node.js at all.
The core pitch is token efficiency. According to benchmarks published in the project's GitHub repository, a full agent loop on Google Flights costs 4,110 tokens with Kuri versus 4,880 with agent-browser. That's a 16% reduction per cycle.
The savings come from an "interactive" snapshot filter that returns only essential page elements (links, buttons) with short references, instead of dumping the entire DOM at your model. When you're running multi-step agent tasks, those savings compound fast. Action responses are flat JSON, so a click costs 9 tokens, back costs 5.
Kuri offers four modes: a CDP server for Chrome automation, a standalone fetcher called kuri-fetch that doesn't need Chrome (it uses QuickJS for SSR-style DOM rendering at ~2 MB), an interactive terminal browser, and an agentic CLI. For agents that just need to read a page and extract information, skipping Chrome entirely is a real architectural advantage.
Most browser automation tools were built for QA engineers, not agent loops. Playwright and its ecosystem add latency and bloat that don't make sense when your agent just wants to read, click, and move on.
The project sits at 54 GitHub stars, so it's early. But justrach's argument is straightforward and worth considering. The performance numbers are self-published, but the benchmarks include reproduction scripts, so you can verify them yourself.