Submission + - AI Agent Writes, Debugs, and Ships a Web App Entirely Through a Browser (dev.to)
domonus writes: An open source project has demonstrated what appears to be the first instance of an AI agent building a complete web application entirely through a browser-based IDE, with zero access to a local filesystem, terminal, or development toolchain. The agent navigated to StackBlitz, forked a Vite project, wrote 7,000 characters of JavaScript and CSS through the Monaco editor API, triggered hot-reload via the WebContainer dev server, and produced a working interactive timer app — all by manipulating a headless Chromium browser.
From the writeup: "Every AI coding tool today works the same way: the agent reads and writes files on your local filesystem, runs commands in your terminal, and uses language servers for feedback. This demo inverts that assumption. The agent had no filesystem access, no terminal, no language server, no local state. Everything existed in a browser tab."
The project, vscreen, is open source and gives AI agents access to a real Chromium instance via MCP. The implication is that any browser-based development environment — StackBlitz, CodeSandbox, Replit, GitHub Codespaces — is now a potential agent workspace without any API integration.
From the writeup: "Every AI coding tool today works the same way: the agent reads and writes files on your local filesystem, runs commands in your terminal, and uses language servers for feedback. This demo inverts that assumption. The agent had no filesystem access, no terminal, no language server, no local state. Everything existed in a browser tab."
The project, vscreen, is open source and gives AI agents access to a real Chromium instance via MCP. The implication is that any browser-based development environment — StackBlitz, CodeSandbox, Replit, GitHub Codespaces — is now a potential agent workspace without any API integration.
