I am building an open-source software-dev agent. Couldn’t think of a better name than LFG. Goal is simple, up the world to get things done.
When it comes to automated, web based dev agents, there are a few solid products already in the market – Replit, Loveable, Bolt among the top ones. Then you have Cursor and Windsurf doing great work as agentic IDEs.
So why another tool?
Software, already powers most of the modern economy. It solves lot of problems, and it is cheap. With AI, it will get even more entrenched into our daily lives. To a degree where I believe creating new software might become analogous to opening a new tab to browse.
For such a critical infrastructure, there has to be an open-source version that doesn’t belong to a single company. A handful of players should not be controlling something as critical as this.
So, that’s why the LFG project. It’ll be fully open under the MIT license.
From what I have seen and built, this isn’t that hard to put together. The real magic isn’t in the tools, but in the LLMs underneath. I’ve been deep in AI tooling the past two years, and even I’m constantly surprised by what Claude, ChatGPT, and Gemini are pulling off now. Things are only going get better from here. And soon even the open-source LLMs will catchup with the mainstream.
Basic requirements:
A dev agent, that takes in your requirements, does the research, and helps build what is needed. I will be adding a little bit of my own flavor. I have a few ideas on how to make the dev agent work better than the current ones out there.
Most tools today jump straight into code generation. That works fine for simple stuff. But if your project has any meaningful complexity, you need structure.
Here’s my approach:
- First, generate a high-level PRD.
- Then, break it down into features and user personas.
- Use that to create proper tickets for backend and frontend.
- Then start writing code.
This structure helps track what worked and what didn’t. It also creates a clear reference for future changes. And this would also allow you to handover something to human if you ever need to.
BTW, none of this is new. This is how teams already build software. It might feel like more work up front, especially compared to just watching UI magically appear. But this should get better results.
Wish me luck!