frane/vibesurfer

<h1 align="center">vibesurfer (<code>vs</code>)</h1> <p align="center"><strong>A real browser for your local AI agent.</strong></p> <p align="center"> <a href="https://github.com/frane/vibesurfer/actions/workflows/ci.yml"><img alt="ci" src="https://img.shields.io/github/actions/workflow/status/frane/vibesurfer/ci.yml?branch=main&label=ci&style=flat-square"></a> <a href="https://github.com/frane/vibesurfer/actions/workflows/engine-tests.yml"><img alt="engine-tests" src="https://img.shields.io/github/actions/workflow/status/frane/vibesurfer/engine-tests.yml?branch=main&label=engine-tests&style=flat-square"></a> <a href="https://github.com/frane/vibesurfer/releases/latest"><img alt="release" src="https://img.shields.io/github/v/release/frane/vibesurfer?style=flat-square"></a> <a href="https://github.com/frane/vibesurfer/blob/main/LICENSE"><img alt="license" src="https://img.shields.io/badge/license-Apache_2.0-blue?style=flat-square"></a> </p> <p align="center"> <img src="https://github.com/frane/vibesurfer/raw/main/docs/demo-claude.gif" alt="Claude Code using vibesurfer" width="560"> </p>

Why

I wanted agents to test web apps via the browser. Everything I tried (Playwright, Puppeteer, anything else that wraps CDP) was too heavy and too unstable. CDP drops sessions. Playwright crashes on long runs. Chrome gets fatter every release. None of that is the actual problem though. CDP and Chrome were designed for humans staring at DevTools. They were never designed for an agent stuck in a while loop.

An agent pays per token. It blocks per response. It can't deal with the event firehose, and a 4kb DOM dump on every read burns the context budget fast. The Hacker News front page through Playwright is about 2000 input tokens before the agent has done anything. Through vibesurfer it's around 50.

vibesurfer is a native browser daemon in Rust. Reads return state tokens and tree deltas instead of the full DOM. Writes check the token. If anything moved between the re