Atlas is a native macOS app that unifies the AI models I use daily. Instead of juggling browser tabs, it brings cloud providers (OpenAI, Anthropic, Gemini) and local models (Ollama, Apple Intelligence) into a single, locally-persisted chat interface.
It operates on a Bring-Your-Own-Key (BYOK) model, saving credentials securely in the macOS Keychain.
Abstracting the Provider
The core idea is simple: “I want to talk to different AIs without opening four different apps.”
By normalizing wildly different API behaviors into a single UI state, the view layer treats a cloud stream from Anthropic exactly the same as a local on-device model.
- 100% Local Persistence: Built on SwiftData. Zero cloud sync; everything lives exclusively on-device.
- Unified Routing: Abstracts complex model APIs behind a single view layer.
- Reasoning & Web Grounding: Native support for chain-of-thought metadata and live web search context.
Evolution & Timeline
Atlas is a highly polished daily driver, but it began as an experimental Python CLI named Orion. After Orion proved the value of smart routing, I abandoned the terminal and poured its best ideas into this native SwiftUI app over a rapid 40-hour sprint:
- Jul 2024: Early Python web UI experiments for local LLMs.
- Mar 2026: Built Orion, a CLI testing smart routing and web search.
- Apr 2 (Morning): Abandoned the CLI; built the initial native Mac UI and Gemini integration.
- Apr 2 (Afternoon): Expanded to Claude, OpenAI, Ollama, and Apple Intelligence. Added SwiftData persistence.
- Apr 3: Added native LaTeX math rendering. Rewrote the architecture using dependency injection to manage the rapid scope creep, and ported Orion’s web search to Swift.
- Apr 17: Finalized routing logic and stabilized at v1.0.