I've been working with AI to code for a few years now, from my early experiments with Claude where I copy/pasted code around, to the agentic frameworks we got last year like Claude Code. It's crazy to think it's only been a year!

I did experiment with Claude Code, Codex, and Gemini all three for a few months in 2025 while I was developing Dunil's Hold. They ran some great deals last fall, double usage and such, and I had all three churning away on it.

I've since started serious development of a local coding agent using an on-device sLM. I have the first version on my github now, it was an early experiment, and I am currently on my 5th version which I plan to release once in a releaseable state.

Tom5 is a multi-agent framework, using mainly the same LLM in various roles, and I have started experimenting with using Apple's Foundation Model in limited roles. Apple has released a Python API so you can interface with the model on your Mac, and it can do tool calling but in my tests tool calling isn't production ready.

My developer machine is a M2 Pro MacBook Pro with 16 gigs of RAM, nothing amazing, and I've been using 2B/4B/8B models with various levels of success. In my use case I have found Qwen to be the best at coding at these weights, though to be honest I have only compared Qwen against Gemma. I'm currently using Qwen 3.5 4B.

Tom3, another private build, was the first version I experimented with editing files and multi-agent worksflows with the same LLM. The edit_file tool would launch a sub-agent that performed the edit, keeping the main agents context smaller. Context management is key, and I'm finding even small models can perform well given the right context.

Tom3 can't one-shot a demo game like frontier models, but can work through it with your help. Tom5 hopes to change that. In working on 3, and a bit of 4 which was a side-quest into a TUI version, and a lot of reading, I've reimagined the architecture and started development of Tom5 which I hope to make work till completion of long tasks.

I had started development on Tom5 when I heard of Apple's Foundation Model API in Python and quickly had Tom3 write me a test script so I could see how well the Apple model performs. In my tests it says it is a model from OpenAI, but I can't find anything concrete on model specs. Here is a link to the docs

Tool calling was flaky, it sometimes worked sometimes threw an obscure error from the internals, so I'm just using it for text generation. We will see how well it performs once it reaches testing in the Tom5 framework filling in roles like verifier. I've been testing it by feeding it good and bad code to see if it catches errors and its general output.