Our electron app browsing the HH25 site
Cascade using our MCP server!

Lumen: AI-Powered Browser Testing for Programming AI

Inspiration

With the advent of AI coding assistants like Cursor, the quality of code is seriously deteriorating. 40% of new code at Google is AI-generated, and nearly 70% of new websites are all vibe-coded.

Google and these websites have something in common, though. They don't work. In 2025, Google Cloud outages tripled, and the vibe-coded websites, well, we know they don't always work.

But what if these AIs could see and test what they build? We wanted to bridge this gap and give AI the power of sight and interaction.

What it does

Lumen gives your Programming AI complete control over your browser to automatically test and validate its creations. By simply calling our tool, software development AI's can:

Take Full Browser Control: We literally take over your browser - clicking buttons, filling forms, navigating pages, scrolling, and interacting with every element on your site
Design Feedback: Makes websites look professional and polished with AI-powered visual analysis that catches layout issues, contrast problems, and design inconsistencies
Functionality Testing: Ensures all features work as intended by automatically testing user flows, form submissions, navigation, and interactive elements
Accessibility Auditing: Identifies accessibility issues like poor contrast, missing alt text, and keyboard navigation problems
Error Detection: Finds broken links, JavaScript errors, and UI bugs that humans might miss

The beauty of Lumen is that it's completely free! We custom-engineered a way for OpenRouter to interact with Stagehand, allowing for zero-cost testing. This means your AI could work uninterrupted for hours with no human intervention, with no worry of the extra costs.

How we built it

We built Lumen using a powerful combination of technologies:

Stagehand: Browser automation framework that gives us complete control over Chrome/Chromium
OpenRouter API: AI vision models for analyzing screenshots and providing intelligent feedback
Electron Dashboard: Real-time UI showing live browser interactions and AI analysis
Custom MCP Integration: Model Context Protocol implementation for seamless IDE integration
WebSocket Communication: Real-time data streaming between browser, AI, and dashboard

The core innovation is our AI Agent Mode - instead of just taking screenshots, we actually program the AI to navigate and interact with websites autonomously, just like a human user would.

Challenges we ran into

Browser Control Complexity: Getting reliable browser automation that works across different websites and interaction patterns
AI Vision Accuracy: Ensuring the AI could accurately identify UI elements and provide meaningful design feedback
Real-time Performance: Balancing detailed analysis with fast execution for seamless developer experience
Cross-platform Compatibility: Making sure the tool works consistently across different operating systems and browser versions
Cost Optimization: Finding ways to provide comprehensive testing without expensive API calls

Accomplishments that we're proud of

Zero-Cost AI Testing: Achieved completely free browser automation and visual analysis
Real Browser Control: Successfully implemented full browser takeover with click, type, scroll, and navigation capabilities
Live Dashboard: Created an Electron app that streams real-time browser interactions with live AI critique
MCP Integration: Built Model Context Protocol support for seamless IDE integration
Comprehensive Analysis: Developed AI-powered critique system that catches design, accessibility, and functionality issues
Professional Results: Generated detailed reports with screenshots, terminal output, and actionable feedback

What we learned

Pivot quickly: We spent a lot of time on a Python implementation of this despite knowing there was an easier way (~4 hours), partly because of the sunk cost.
Real-time interaction: Live browser control with immediate feedback creates a much more engaging development experience

What's next for Lumen: Bringing sight to AI

We're turning Lumen into a full MCP (Model Context Protocol) tool! This integration will allow users on any major IDE to set up our browser testing capabilities with just a click of a button, just like you would do on the App Store! As a long-term outlook, we truly believe this is a game-changing tool in democratizing web development and computer science for everyone. We want to reach as many developers as possible and anticipate further refining our product with features like: