Been working on an open source project for a while called Ironsmith. The gist is you can create highly specific macOS apps for yourself with just a prompt, and one of my main goals from the beginning was to get it to work with low end models like the Apple foundation and the Gemma series.
It uses a custom agentic loop tailor made to work with small models with limited context. This means you can create very simple apps entirely on device with a Mac as limited as an 8gb MacBook Air.
I found that the secret sauce to making this work was just have the model generate the entire app in one go, and then run a bajillion formatting, linting and deterministic repairs until it makes something compilable. Turns out these little models are pretty decent at writing full apps if you fix all of their hallucinations and syntax errors.
That being said you will get higher quality apps and less chances for errors the better the model you build with. I find that Gemma 4 26b a4b gives the best balance here, but it does require at least 24gb memory.
You can use Ollama out of the box and also use all of your favorite local providers via an OpenAI compatible API. ChatGPT, Claude and Gemini are also available to connect to if you want to provide your own API key (and I recommend it if you plan on making anything more complex).
Security wise, every app is sandboxed and hardened by default (though you can turn it off), and all sensitive permissions like camera and microphone have to be explicitly enabled.
Give it a try and let me know what you think or if you run into and issues!
Show HN: Ironsmith – Generate personal native Mac apps with local LLMs · HackerTrans
It uses a custom agentic loop tailor made to work with small models with limited context. This means you can create very simple apps entirely on device with a Mac as limited as an 8gb MacBook Air.
I found that the secret sauce to making this work was just have the model generate the entire app in one go, and then run a bajillion formatting, linting and deterministic repairs until it makes something compilable. Turns out these little models are pretty decent at writing full apps if you fix all of their hallucinations and syntax errors.
That being said you will get higher quality apps and less chances for errors the better the model you build with. I find that Gemma 4 26b a4b gives the best balance here, but it does require at least 24gb memory.
You can use Ollama out of the box and also use all of your favorite local providers via an OpenAI compatible API. ChatGPT, Claude and Gemini are also available to connect to if you want to provide your own API key (and I recommend it if you plan on making anything more complex).
Security wise, every app is sandboxed and hardened by default (though you can turn it off), and all sensitive permissions like camera and microphone have to be explicitly enabled.
Give it a try and let me know what you think or if you run into and issues!