First package built with AI-assistance

[AlejandroGomezFrieiro]

I ended up building https://github.com/AlejandroGomezFrieiro/noogle_mcp_server

This is a simple mcp server for querying noogle.dev for Nix functions. Not sure it'll be very easy to use, or even useful, but I got something out of it. It also helped me get into uv2nix, which is something I wanted to try for a while.

Some learnings:
Higher-order thinking is important. Developing a step-by-step plan before any code is written. Also use the edit-test workflow a lot to check that things work. It is really like guiding another engineer that knows how to code to some degree, but needs strict guidance.

I used context7 to fetch documentation of fastmcp and, honestly, most of the package was build fairly quickly. After that it was a matter of fine-tuning the code to output what I wanted. Really like that the fetch was also cached!

Tests were also built by codecompanion, asked to output a specific markdown from the html example. It was a relatively nice loop to build the tests, and then loop.

Although it only took me a few hours to get it done, I have to say babysitting an AI can be frustrating, in particular when the buffer editing tool tends to fail with the free version of gemini.

I ended up fixing the last few issues with the html->markdown parsing (at least that I found), so it definitely was more of an assisted workflow than "vibe coding".

Even with all of that, it was a fun weekend project that clarified for me how I might use LLMs more in the future. Still not quite convinced I'll use them a lot, but at least it was a good learning experience.

Thanks to all the developers who made codecompanion!