# I’m building a macOS AI app around Ollama, with a focus on making smaller models more reliable — retained record

> Retained reddit record from reddit:ollama.

Canonical: https://highsignal.app/data/records/952cce11d4a46198

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:ollama)  published 2026-10-19 18:26 ingested 2026-10-20 02:39

[original source ↗](https://www.reddit.com/r/ollama/comments/1xa4s9r/im_building_a_macos_ai_app_around_ollama_with_a/)

## retained text



```
One thing I’ve noticed while researching local AI agents is that model intelligence isn’t always the biggest limitation.  
Sometimes the model identifies the problem correctly but keeps inspecting files, repeats tool calls, loses context, or never finishes the task.  
I’m building an execution system called Cortex to investigate whether better tool handling, progress detection, and recovery mechanisms can reduce those failures.  
Cortex is part of Solar, a macOS AI app I’m developing for local and cloud models.  
An important design goal is keeping the execution system lightweight, especially for smaller quantized models. I don’t want to bury a 9B model under huge system prompts and dozens of unnecessary tools.  
These are goals we’re still testing, not benchmark-proven improvements.  
I’m also looking for early beta testers who regularly use Ollama and would be interested in evaluating Solar as it develops.  
Beta waitlist, for anyone interested:  
https://solar-waitlist.vercel.app  
What’s the smallest local model you’ve found genuinely useful for agentic tasks?
```



document fetched 2026-10-20 00:41 · canonical https://reddit.com/r/ollama/comments/1xa4s9r/im_building_a_macos_ai_app_around_ollama_with_a
