Nativ

Local AI, native to your Mac.

Chat, serve, monitor, and connect MLX models from one macOS app.

Nativ is a native macOS workspace for running AI models locally on Apple silicon. It bundles an mlx-vlm server, finds compatible models in your Hugging Face cache (honoring HF_HUB_CACHE and HF_HOME), and wraps the whole experience in a polished SwiftUI app.

Use Nativ as a private chat app, a model manager, a performance dashboard, or an OpenAI- and Anthropic-compatible local inference server for the tools you already use.

Documentation

Full feature and contributor documentation lives in Docs/ — start there for Chat, Models, Voice, Routines, Integrations, and the Developer reference, plus guides for extensions, kits, and the MCP catalog.

What Nativ can do

Feature	What you get
Local chat and vision	Streaming conversations, image attachments, reasoning output, response metrics, and persistent chat history.
Image generation and editing	Generate and edit images locally with compatible MLX image models in a dedicated Images tab.
Model library	Discover installed MLX models, browse and download compatible models from Hugging Face with fit warnings for your memory, inspect capabilities, switch models, or remove old ones. Preload separate language, image-generation, speech, and embedding models at once, with a warning if the combination would exceed your Mac's memory.
Performance analytics	Track request volume, token usage, time to first token, decode speed, model performance, and recent activity.
System monitor	Inspect live per-core CPU load, GPU utilization, unified memory and swap pressure, disk throughput, capacity, SMART health, and thermal and power sensors.
Local APIs	OpenAI-compatible chat, Responses, image, audio, embeddings, and model endpoints, plus Anthropic Messages endpoints.
Coding-tool integrations	Configure and launch terminal coding agents — Codex, Claude Code, Pi, Hermes, OpenCode, Aider, Goose, Crush, Qwen Code, OpenClaw — and set up editors — VS Code, Cursor, Zed, JetBrains, Cline, Continue — against models served by Nativ. See Docs/features/integrations.md for per-tool setup.
Developer workspace	Set the server host and port, add a Hugging Face token for gated models, inspect runtime details, copy endpoint URLs, search and filter live server logs, and monitor server health.
Menu bar controls	Start or stop the server, change the loaded model, check serving statistics, open the main app without breaking focus, or pin multiple live CPU, GPU, and RAM percentages and mini graphs.
Extension platform	Install, disable, remove, and restore independently versioned capabilities. Audio ships as the first included extension and contributes its own pages, commands, shortcuts, settings, and permission declarations.
Audio extension	Use private local audio capabilities, including voice dictation in any app with either a pointer-following waveform or a camera-cutout pill with a reactive gradient orb and timer. Review transcript history, track words per minute, total words, time saved, and streaks, choose an installed speech model, and customize the record and retry shortcuts.
Artifacts	Browse every image and document produced or uploaded across chats, with filters, sorting, and semantic search over your own artifacts.
MCP servers	Connect Model Context Protocol servers and expose their tools to chat, with per-tool consent.
Routines	Schedule prompts to run on their own and collect the results.
Chat tools	Let the model generate and edit images, list or switch models, and read server and system stats, each behind a consent prompt.
Advanced inference controls	Tune sampling, thinking budgets, structured output, KV-cache quantization, prefix caching, and speculative decoding.

Inference runs on your Mac after a model has been downloaded. Model downloads and first-time build dependencies still require network access.

How it works

flowchart LR
    A["Nativ · SwiftUI app"] --> B["NativServerKit"]
    B --> C["Bundled mlx-vlm server"]
    C --> D["MLX runtime"]
    D --> E["Local models · Apple unified memory"]
    F["Apps and coding agents"] -->|"localhost API"| C

NativServerKit owns the embedded Python distribution and server lifecycle. The app adds model discovery, chat, analytics, configuration, integrations, logs, menu bar controls, and software updates around that runtime.

Requirements

To run the app:

• A Mac with Apple silicon.
• macOS 26 or newer.
• Enough unified memory for the model you choose.
• Optional: a Hugging Face token (set in the app or via HF_TOKEN) to download gated models.

To build from source, you will also need:

• Xcode with the macOS 26 SDK.
• xcodegen.
• Python 3.
• Network access to GitHub Releases and PyPI while the embedded Python bundle is first assembled or refreshed.

Get started

Download a release

Download the latest DMG from GitHub Releases, drag Nativ to Applications, and launch it. Nativ uses Sparkle for subsequent in-app updates.

On first launch:

1. Choose an installed language model, download a recommended one, or continue with load-on-demand.
2. Optionally generate an API key to protect the server's management endpoints.
3. Grant microphone, accessibility, and screen recording permissions up front, or change them later in Settings.
4. Open Models to download or select a compatible model.
5. Start chatting, inspect analytics, or connect one of the supported coding tools.

Nativ asks for Accessibility permission so it can detect the default Control + Option + Command double-tap outside the app, and for Microphone permission the first time you record. Recordings are saved temporarily as .wav files with matching .txt transcripts and can be opened from Show Voice Recordings in the menu-bar menu. Raw audio is deleted automatically after five minutes (or immediately when Nativ quits), while transcript files remain available. Press Fn + R to transcribe the newest available audio again and insert it at the current cursor. The transcript remains on the clipboard. If no speech-to-text model is installed, Nativ links directly to filtered speech-model discovery. Open Audio to inspect dictation analytics and history, select an installed speech-to-text model, choose the capture animation, or change either global shortcut.

Build from source

brew install xcodegen
make xcode-generate
make xcode-run

The first build can take a while because NativServerKit creates a relocatable Python runtime and installs the pinned mlx-vlm server dependencies into the framework resources. Later builds reuse the bundle until an input changes.

Local builds are signed with the Apple Development identity configured in Configuration/Signing.xcconfig (or the ignored Configuration/Signing.local.xcconfig override). Keep the identity's login keychain unlocked while building. The stable signer-bound identity lets macOS keep Accessibility permission across rebuilds instead of treating each binary as a different app.

To test an unreleased mlx-audio checkout, point the build at its local path:

MLX_AUDIO_SOURCE_PATH=/path/to/mlx-audio make xcode-build

The local package participates in dependency resolution, so its pyproject.toml metadata replaces the published package metadata. A sibling ../mlx-audio checkout on the main branch is detected automatically, matching the existing local mlx-vlm behavior.

Use Nativ as a local API server

By default, the app exposes its server at http://127.0.0.1:8080. You can change the host and port in the Developer page, which also lists every available endpoint and lets you copy URLs directly.

For example, with a model selected:

curl http://127.0.0.1:8080/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "your-model-id",
    "messages": [{"role": "user", "content": "Why is the sky blue?"}],
    "stream": false
  }'

If you enabled a server API key, also send it as a Bearer token:

-H 'Authorization: Bearer your-api-key'

The server includes:

• OpenAI-compatible /v1/chat/completions, /v1/responses, /v1/models, image, and audio routes.
• Anthropic-compatible /v1/messages and token-counting routes.
• /health, /metrics, cache statistics, cache reset, and model unload endpoints.

Project layout

Sources/
├── Nativ/                       # SwiftUI application
│   ├── Features/
│   │   ├── Artifacts/
│   │   ├── Chat/
│   │   ├── Dashboard/
│   │   ├── Developer/
│   │   ├── Extensions/         # Extension registry, broker, and management UI
│   │   ├── ImageGeneration/
│   │   ├── Integrations/
│   │   ├── IssueReport/
│   │   ├── MCP/
│   │   ├── Models/
│   │   ├── Routines/
│   │   ├── Settings/
│   │   ├── Shared/
│   │   ├── VoiceCapture/
│   │   └── SystemMonitor/
│   ├── Assets.xcassets/
│   ├── ModelProviderIcons/
│   └── Utilities/
├── NativExtensionSDK/           # Versioned extension manifests and XPC contracts
└── NativServerKit/              # Embedded server and Swift clients
Extensions/
└── VoiceDictation/              # Audio extension (legacy internal target name)
PythonDistribution/
├── Launcher/                    # Relocatable server launcher
├── Requirements/                # Pinned Python dependencies
└── Scripts/                     # Bundle assembly and verification
Configuration/                   # App metadata and signing settings
Design/                          # Brand source files and README artwork
scripts/                         # Archive, signing, notarization, and release tools
project.yml                      # XcodeGen project definition

See Docs/extending/extensions.md for the extension package format, lifecycle, permission model, and the steps for adding another first-party extension.

Development

Build and smoke tests

Generate and build the Xcode project:

make xcode-generate
make xcode-build

Verify that the bundled executable can launch and print mlx_vlm.server help:

make xcode-smoke

Exercise the long-running process lifecycle and /metrics readiness:

make xcode-lifecycle-smoke

To generate a few real requests and compare metrics before and after:

scripts/run_metrics_queries.py

The first request may take longer while its model downloads and loads.

---

Built for fast, local inference on Apple silicon.