April 22, 2026AgentsInfrastructureOpen Source

Kuri is a 464KB Zig browser that thinks in tokens, not pixels

Kuri just shipped on Hacker News and the pitch is refreshingly unsexy. A browser automation tool written in Zig, 464KB binary, 3ms cold start, zero Node.js dependency, built specifically so an agent loop costs fewer tokens. On a Google Flights test, a full Kuri loop runs 4,110 tokens versus 4,880 for the standard agent-browser, a 16% cycle reduction. That sounds boring until you multiply by a million agent runs.

The design choice is worth stealing. Most browser tools hand an agent a verbose JSON dump of the page state and burn context window on layout metadata the agent never uses. Kuri uses accessibility tree snapshots with refs, the agent reads a compact, stable representation and acts on named references instead of CSS selectors that break. Four modes ship in the binary, CDP server, standalone fetcher, terminal browser, and an agentic CLI. You pick the smallest surface for the job.

This is the second serious Zig-native browser play in the agent space alongside Lightpanda, and the pattern is clear now. Chrome is built for humans, the DOM is built for humans, Playwright and Puppeteer are built for humans. When the user of the browser is an LLM paying per token, the winning browser is one that was designed to be read by an LLM from day one. Kuri is one of the first to commit to that all the way down to the language choice.

The practical question is whether token efficiency becomes a real buying factor for agent infra. Right now most builders pick Playwright because it works, not because it is cheap. Kuri is betting that when a production agent runs 100K browser cycles a day, 16% token savings compounds into real dollars and that someone eventually does the math. If they are right, the space for human-shaped browser tools in agent pipelines shrinks fast.

Repo at https://github.com/justrach/kuri.
← Previous
DecisionBox Enterprise bets air-gapped is the only serious enterprise agent pitch
Next β†’
AgentSPEX argues that Python is the wrong shape for agent workflows
← Back to all articles

Comments

Loading...
>_