Ouroboros

Ouroboros watches your app like a real user, catches what you missed, and fixes it, automatically. The loop never stops

Comment

Inspiration

Vibe coding. Vibe code central. The world is running on engineers who can operate, manage, and develop with coding agents at a large scale. It was on stage when a young engineer came clean and admitted his workflow consisted of 5 coding agents going off at the same time. As much as vibe coding lowers the drawbridge for excited developers and people interested in software, a threshold that still holds us back is the strenuous task of debugging exponentially larger amounts of code at once. With UX struggles heavy on our mind, the lack of tools for visual debugging, and exciting technology like Browser Use, we wanted to tackle this obstacle of optimizing the vibe coding workflow.

What it does

Ouroboros thins the line between constant prompting, running generated code, and testing said code. This CLI app works with coding agents, Claude in this case, to spin up the localhost for a project, make it reachable to Browser Use via Cloudflare, and test prospective user efforts, scan for visual bugs that would fly under script radars, and give feedback on inaccessible components. After the Browser has had a chance to gauge, it sends the feedback back to Claude for instant repairs, creating a loop of development and fixes until the feedback stops.

How we built it

We built the app as a CLI tool to be used by coding agents like Gemini and Claude. It's a Python codebase to best script and utilize the Browser Use API. We used Cloudflare to create proxies to make development environments visible to Browser Use in the cloud, and ngrok as a backup if Cloudflare decided to rate limit us. Finally, a little command line action to keep the orders flowing, environments up, and bond interactions between AI agents.

Challenges we ran into

Making proxies for localhost through integrated Cloudflare was a headache as it kept rate-limiting us, but we developed an ngrok backup to handle missing tunnels for agents. As we developed more features, and the tasks for Browser Use became more abstract, we needed to stretch Browser Use agents to bounds we struggled to keep up with; synced terminating sessions, persistent workspaces, and opening clients on their own. Through our efforts, extensive readings of the docs, and abuse prompting, we pulled through to create a workflow that was effective.

Accomplishments that we're proud of

We're proud of how far we managed to dig under the hood for development. So many layers and abstractions we really got to understand and experiment with to make this tool that is totally going to be a go-to for our next front-end project.

What we learned

After everything, we learned about atomic tools for web interactions, how LLMs are capable but lack that (except Browser Use), and you can never trust a proxy; you never know if it's going to blow up, give you a bad gateway, or work splendidly.

What's next for Ouroboros

We want to make this an installable CLI tool to be used in development by other coding agents besides Claude to boost vibe code fiends or within IDEs to make the lives of developers who are getting into vibe coding easier.

Built With

Share this project:

Updates