Nativ – Local AI on macOS

What it is

Nativ is a native macOS application (written in SwiftUI) that lets you run large‑language‑model, vision, and speech models locally on Apple‑silicon Macs. It bundles the open‑source mlx‑vlm server (which uses the MLX runtime) and provides a polished UI for chatting, image generation, audio dictation, model management, performance monitoring, and exposing OpenAI‑compatible APIs to other tools.

Core features (as described in the README)

Feature	What you get
Local chat & vision	Conversational UI with streaming responses, image attachments, metrics, and persistent history.
Multiple windows	Separate workspaces that share the same inference server and model.
Shared work pane	Open docs, code, HTML, or websites beside chat; agents can edit or interact with the same pane.
Image generation & editing	Run compatible MLX image models in an “Images” tab.
Model library	Discover installed MLX models, browse Hugging Face, see memory‑fit warnings, store cache on external APFS, preload multiple model types, and manage them.
Performance analytics	Track request volume, token usage, latency, decode speed, etc.
System monitor	Live per‑core CPU, GPU, memory, disk, SMART health, thermal and power sensors.
Local APIs	OpenAI‑compatible /v1/chat/completions, /v1/models, image, audio, embeddings; Anthropic‑compatible /v1/messages.
Coding‑tool integrations	Pre‑configured agents (Claude, Codex, etc.) and editors (VS Code, Cursor, Zed, JetBrains, etc.) can talk to the local server.
Developer workspace	Server host/port, Hugging Face token, runtime details, log view, health monitor.
Menu‑bar controls	Quick start/stop, model switch, live CPU/GPU/RAM mini‑graphs.
Extension platform	Install/disable versioned extensions; first‑party audio extension adds dictation UI, shortcuts, analytics.
Artifacts browser	Browse, filter, and semantically search images/documents produced in chats.
MCP servers	Connect Model Context Protocol servers for tool‑level extensions.
Scheduled tasks	Run saved prompts on a schedule with kits, tools, and skills.
Advanced inference controls	Sampling, budget, structured output, KV‑cache quantisation, speculative decoding, etc.

How it works (high‑level architecture)

Nativ (SwiftUI app) → NativServerKit → bundled mlx‑vlm server → MLX runtime → local Apple‑silicon models

• NativServerKit ships an embedded Python runtime and manages the lifecycle of the mlx‑vlm server.
• The server runs inference using the MLX library, which executes directly on the Mac’s unified memory and GPU.
• The SwiftUI front‑end talks to the server over localhost HTTP, exposing the same endpoints that OpenAI/Anthropic services provide.

Getting started

• Pre‑built – Download the DMG from the latest GitHub release or install via Homebrew (brew install --cask nativ). The app updates itself with Sparkle.
• First launch – Choose or download a compatible model, optionally set an API key, grant microphone/accessibility permissions, and start chatting.
• From source – Requires Xcode 26+, Swift 5, xcodegen, Python 3, and the Metal toolchain. Build with make xcode-generate && make xcode-build. The first build assembles the relocatable Python bundle and installs pinned mlx‑vlm dependencies.

System requirements

• Apple‑silicon Mac (M1/M2/…)
• macOS 26 or newer
• Sufficient unified memory for the chosen model
• Optional Hugging Face token for gated models

Extensibility

• Extensions – Packages that add UI pages, commands, shortcuts, and permissions. The built‑in Audio extension provides voice dictation.
• Kits & MCP – Developers can ship additional capabilities (e.g., custom tools) that the chat can invoke.

Typical use‑cases

1. Private local chat assistant – No internet, full control over data.
2. Local API server – Run curl or configure VS Code, Cursor, etc., to use the OpenAI‑compatible endpoints.
3. Image generation – Generate or edit images without leaving the app.
4. Voice dictation – Capture audio, transcribe locally, and insert into any app.
5. Performance monitoring – See how models utilise CPU/GPU and adjust settings.

---

All details above are taken directly from the repository’s README.