AI Browser Development Story: How a Failed Project Found New Life

Source: Dev.to

Background

I work as the frontend lead for AI initiatives at my company. In the first half of 2024 I was maintaining a lukewarm AI project, and like many professionals I needed a standout effort for a good performance review. My colleague Tan, the product manager for AI applications, and I started a two‑month exploration into a more ambitious idea.

Initial Concept

After reviewing competitor apps we settled on a “Jarvis‑style” assistant:

• The user asks a question.
• The agent autonomously gathers information, calls various tools, and delivers an answer.

The concept was inspired by Fellou’s AI browser, whose interactions and UI impressed me.

Technical Challenges

Controlling a Browser

The core problem was programmatically controlling a browser to automate web tasks. I quickly abandoned the company’s legacy C++ client and, with a frontend background, turned to Electron for a desktop app.

I tried to combine several tools:

• browser‑use – a Python service for browser automation.
• Playwright – a separate automation engine.

I asked Cursor (an AI assistant) whether Electron + browser‑use + Playwright could be integrated. The answer was “Yes,” but after many attempts the integration proved impossible. I spent $200 on Cursor credits without getting a working demo, learning that AI can accelerate work only in domains you already understand.

Finding an Alternative Stack

Research led me to the Eko framework (open‑sourced by the Fellou team). I discovered that Eko could run inside Electron with some compatibility work.

To embed a local frontend service, I chose Next.js. After several failed attempts with AI assistance, I manually followed the Next.js documentation and succeeded in embedding it in Electron using a custom server—similar to the approach used in the open‑source project bolt.diy.

Building the Prototype

• I learned React fundamentals (my previous stack was Vue).
• I studied Electron’s main/renderer processes and IPC communication.
• I made the Eko framework compatible with Electron to control web pages in view windows.

Agents Implemented

I experimented with screen‑sharing (inspired by Fellou) and fully voice‑driven input, but both approaches had significant limitations, so they were dropped.

The result was a “Frankenstein” demo that combined the three agents and functioned well enough for a presentation.

Project Rejection

When I demoed the prototype to leadership, the project was rejected despite the two months of work. I was reassigned to other company AI initiatives, and the demo sat on my hard drive.

Open‑Sourcing the Project

A month later I wondered whether others were facing similar challenges. Using Claude Code, I rebuilt a clean open‑source version over a weekend, this time without corporate constraints.

The repository is now public:

👉 GitHub:

Features Implemented (Late Nights & Weekends)

• History task playback
• Human interaction capability
• Voice input support
• Multi‑language internationalization
• Agent configuration system
• Scheduled tasks
• Toolbox page

Insights

• AI is not omnipotent. It speeds up work in familiar domains but cannot replace a solid understanding of the underlying technology.
• The $200 spent on Cursor bought a valuable lesson: when dealing with unfamiliar tech, you still need to learn it yourself.
• Open‑sourcing preserves knowledge that would otherwise gather dust and can help others avoid the same pitfalls.

Future Work

• 🎨 Theme customization (dark mode)
• 🔌 Plugin marketplace (MCP tool ecosystem)
• 📊 Visual workflow editor
• 🌐 Additional agents (Shell, Email, Notion, …)

Contributing

If you’re interested:

• ⭐ Star the repository – it helps visibility.
• 🐞 File issues or submit pull requests.
• 💬 Join discussions in the GitHub Issues/Discussions.
• 📥 Download and try the tool, then provide feedback.

From an internal, performance‑driven project that was rejected, to an open‑source effort rebuilt over a weekend, this “failed” project has found new life. Open source isn’t just about sharing code—it’s about sharing experience and ideas.