I've been working on Vesta, a native SwiftUI app for macOS that lets you run AI models locally on Apple Silicon — or connect to 31+ cloud inference providers though APIs. The approach of this app is different that LMStudio, Jan and others. They are great. This app also gives acces to Apple's on-device AI model. I'm disappointed that Apple hasn't evolved it since it's not actually terrible. But they limit the context size of it (hard coded)

This is also an experiement on if Coding agents can build an app from scratch. You be the judge. I can assure you however that it wasn't a 'one shot' build. Many millions of tokens burned! Over time I've seen very measurable progress of Claude Code as it evolves. I hope that we can achieve unthetered and local coding AI of this quality soon! This is something I'm prediciting for 2026.

The best bang for the buck as been the Qwen3-VL models for me. Even though they tend to get in repetitive loops sometimes. Known issue.

I chose a more simplistic UI and a different way to interact with the App itself using natural language for those who hate GUI navigation.

To download and view screenshots of the capabilities:

Just Visit - https://kruks.ai/

My github: https://github.com/scouzi1966

This distribution: https://github.com/scouzi1966/vesta-mac-dist

  What makes it different:

  - Natural Language Interface (NLI) with Agentic Sidekick — chat with the app system. Only tested with Claude Code — more to come

Tell Agentic Sidekick to set things up for you instead of using the GUI

The agent can have a conversation with any othe model - entertaining to have 2 models discuss about the meaning of life!

MCP can be activated to allow any other external MCP client using it with ephemeral tokens generated in app for security (I have not tested all the degrees of freedom here!)

MCP can deeply search the conversation history through backend SQL

  - 5 backends in one app — Apple Intelligence (Foundation Models), MLX, llama.cpp, OpenAI, HuggingFace. Switch between them

  - HuggingFace Explorer — I am not affiliated with HuggingFace but combined with the $9/month Pro subscription makes it interesting to explore HF's inference services (this is rough around the edges but it is evolving)

  - Vision/VLM — drag an image into chat, get analysis from local or cloud models

  - 33+ MCP tools — the AI can control the app itself (load models, switch backends, check status) - Agentic Sidekick feature

  - TTS with 45+ voices (Kokoro) + speech-to-text (WhisperKit) + Marvis to mimic your own voice — all on-device

  - Image & video generation — FLUX, Stable Diffusion, Wan2.2, HunyuanVideo with HuggingFace Inference service

  - Proper rendering — LaTeX/KaTeX, syntax-highlighted code blocks, markdown tables

  It's not Electron. It's not a wrapper around an API. It's a real macOS app built with SwiftUI, Metal, llama.cpp library and Swift MLX, HuggingFace Swift SDK — designed for M1/M2/M3/M4/M5.

  Runs on macOS 26+.

  Install:

  brew install --cask scouzi1966/afm/vesta-mac

  Or grab the DMG: https://kruks.ai

  Would love feedback — especially from anyone running local models on Apple Silicon.