Kuri – Zig based agent-browser alternative
TL;DR Highlight
Kuri, a 464KB browser automation tool built with Zig, cuts token costs in AI agent loops by eliminating Node.js dependencies.
Who Should Read
Developers integrating web browser automation into LLM agents who are frustrated by the weight or token waste of Node.js-based tools like Playwright/Puppeteer.
Core Mechanics
- Kuri is a browser automation tool written in Zig, completely independent of Node.js. Its binary size is only 464KB, and cold starts are as fast as approximately 3ms.
- It directly controls Chrome's CDP (Chrome DevTools Protocol), communicating with Chrome without a separate runtime. However, Chrome itself must be running somewhere.
- Its core design philosophy is 'for agent loops, not QA engineers.' It's optimized for a cycle of reading page state → minimizing tokens → reliably clicking with stable references → and moving to the next step.
- Compared to Vercel's agent-browser, Kuri used 16% fewer tokens for the entire agent workflow (go→snap→click→snap→eval) on a Google Flights (SIN→TPE) route. (kuri: 4,110 tokens vs agent-browser: 4,880 tokens).
- Several snapshot modes are available, with the `--interactive` mode being the most efficient for agent loops, using only 1,927 tokens compared to the compact mode's 4,328 tokens.
- However, a full JSON snapshot (`--json`) uses 31,280 tokens—7.2 times more than compact—and lightpanda's semantic_tree is similarly expensive at 26,244 tokens. lightpanda also has the additional drawback of sometimes producing empty DOMs because it doesn't execute JavaScript.
- It includes built-in features like A11y (accessibility tree) snapshots, HAR (HTTP Archive) recording, a standalone fetcher, an interactive terminal browser, and security testing.
- Benchmarks were measured using the same Chrome session and the same tiktoken cl100k_base tokenizer, and can be directly reproduced using `./bench/token_benchmark.sh`.
Evidence
- "Reports indicate that the installation script (install.sh) and installation via bun return 404 errors, suggesting the project's infrastructure is not yet fully stable. A comment pointed out that the benchmark in the README is self-published, making it difficult to trust the 16% token reduction claim without independent verification. A user noted that while kuri-fetch advertises itself as 'standalone,' it still requires Chrome to be running somewhere, functioning merely as a wrapper around CDP and not being truly standalone. A user previously using brow.sh (a text-based browser) for page fetching found Kuri more interesting but was somewhat disappointed after confirming its Chrome dependency."
How to Apply
- "If you're implementing tasks where an LLM agent needs to read web pages (price comparison, information gathering, etc.), try Kuri's `snap --interactive` mode instead of Playwright. You can read the same page with fewer than half the tokens, reducing API costs. The token savings compound in multi-step agent loops (page navigation → snapshot → click → snapshot → judgment). Kuri's benefits increase with the number of steps, making it particularly suitable for automating complex web tasks with 10 or more steps. If you've abandoned serverless or lightweight container deployments of Node.js-based browser automation due to binary size issues, consider Kuri's 464KB single binary. The 3ms cold start is practical even in environments like Lambda. For automated security testing, leverage Kuri's built-in security testing features and HAR recording. HAR files allow you to debug the agent's HTTP requests later."
Code Example
snippet
# Recreate token benchmark directly
./bench/token_benchmark.sh
# Basic snapshot (compact mode, 4,328 tokens)
kuri snap
# Optimal mode for agent loops (1,927 tokens — 0.4x compact)
kuri snap --interactive
# Full JSON dump (31,280 tokens — for debugging)
kuri snap --json
# Standalone page fetcher (but Chrome must be running)
kuri-fetch https://example.comTerminology
CDPChrome DevTools Protocol, which allows external programs to remotely control the Chrome browser. Playwright and Puppeteer also use this internally.
A11y snapshotAn abbreviation for Accessibility tree snapshot, which is a text extraction of the page structure created by the browser for screen readers. It's much more concise than the entire HTML and is better suited for passing to LLMs.
HARAn abbreviation for HTTP Archive, a JSON-formatted file that records all HTTP requests/responses made by the browser. It's used for network debugging and replay testing.
tiktokenAn OpenAI-created tokenizer library that calculates how many tokens a text will be divided into in GPT-based models. cl100k_base is the tokenizer setting used in GPT-4.
cold startThe initial startup time of a program. Node.js-based tools can take hundreds of milliseconds to seconds due to runtime loading, while Kuri claims 3ms.
agent loopThe cycle in which an AI agent 'perceives the situation → makes a decision → takes action → checks the result.' Token efficiency is high because tokens are consumed with each loop, so the more tokens saved, the more steps can be taken at the same cost.