If you have been trying to run a serious language model on a Mac without giving Anthropic or OpenAI a copy of every prompt, the menu of options just got wider. Prince Canuma, the developer who maintains MLX-VLM, shipped Nativ on July 20, 2026: a native macOS chat app for local models that doubles as a localhost API server for Claude Code, Codex, Pi, OpenCode, and Hermes. The first release, v0.0.1, is MIT-licensed, written mostly in Swift, and runs entirely on Apple Silicon.
For self-hosters this matters because Nativ ships as a native macOS app built directly on top of MLX, the open-source machine learning framework Apple ships for its own chips. The story behind the app is also worth knowing: the developer behind Nativ is also the maintainer of MLX-VLM, one of the more widely used Python packages for running vision and omni models on Apple Silicon, so the “newcomer” is already a known quantity in the local-AI ecosystem.
What Nativ actually does
The project page describes Nativ as a desktop workspace for running AI on a Mac, built to “[r]un AI models locally on your Mac” without accounts, subscriptions, or cloud calls (Nativ project site). The app is a SwiftUI shell around an embedded MLX-VLM Python server. When you launch it, Nativ scans your Hugging Face cache for MLX-format models you have already downloaded, picks up anything compatible, and exposes them through two interfaces: a streaming chat UI with markdown, code highlighting, image input, and per-message telemetry (tokens/sec, memory pressure, thermal state, time-to-first-token), and a localhost HTTP server.
That localhost server is the feature local-AI builders will care about most. The default endpoint is http://127.0.0.1:8080, and the GitHub repository lists OpenAI-compatible routes (/v1/chat/completions, /v1/responses, /v1/models) and an Anthropic-compatible /v1/messages route, plus /health, /metrics, and model-cache controls. That means a CLI coding agent configured to talk to Anthropic or OpenAI can be repointed at Nativ and use whatever model you have on disk instead of a hosted endpoint.
Simon Willison, who wrote about Nativ on July 21, 2026, found the Hugging Face cache detection worked on his own machine: “The app picked up MLX models I had already tried that were present in my Hugging Face cache directory.” That kind of zero-config reuse is rare in the local-LLM tooling market.
The developer, in one paragraph
Prince Canuma publishes as “Blaizzy” on GitHub. His MLX-VLM repository is a package for “inference and fine-tuning of Vision Language Models (VLMs) and Omni Models (VLMs with audio and video support) on your Mac using MLX.” It has 5.2k stars, 676 forks, and at the time of writing sits at v0.6.6. Canuma’s other repositories include mlx-audio (“A text-to-speech (TTS), speech-to-text (STT) and speech-to-speech (STS) library built on Apple’s MLX framework”) and the Swift wrapper mlx-audio-swift (“A modular Swift SDK for audio processing with MLX on Apple Silicon”). Local-AI users who already pull models from mlx-community on Hugging Face will recognize the lineage.
What runs on it
The Nativ project site currently lists three “featured” models with public benchmarks the app has captured live on Apple Silicon: Gemma 4 E2B Instruct (10.28 GB, 128k context, vision and audio), North Mini Code from Cohere (19.38 GB, 500k context, code and tools), and LFM2.5-VL 1.6B from Liquid AI (3.20 GB, 128k context, vision and language). The app does not hard-code this list, though; it discovers any MLX-format model already on your disk, and its GitHub roadmap says support for dedicated audio-only and image-generation-only models is “coming soon.”
On the hardware side, the system requirements are deliberately narrow: a Mac with Apple silicon, macOS 26 or newer, and enough unified memory for the chosen model. Apple silicon shares RAM between the CPU and GPU, which is why MLX can run multi-billion-parameter models on a MacBook without a discrete graphics card. The single piece of guidance the project gives is “enough unified memory for the chosen model,” with no minimum figure stated.
The coding-agent angle
The single most useful detail in Nativ for an intelligibberish reader is that you can keep using Claude Code or Codex without sending your prompts to Anthropic or OpenAI. The GitHub readme lists integrations with “Pi, Codex, Claude Code, Hermes, OpenCode.” Point any of those tools at http://127.0.0.1:8080 and it will route requests to the local model instead. The README also lets you optionally generate an API key to protect the server’s management endpoints (Nativ GitHub).
This is the practical privacy escape hatch the local-AI audience has been asking for. You keep the same coding-agent workflow, the same slash commands, the same habit patterns, but the data stays on the laptop. For proprietary code, security audits, or anything covered by NDA, the trade is now: a slower model running on your own M-series chip in exchange for not paying a cloud provider to log your prompts.
There is one real limitation worth flagging. The MLX backend in Nativ is built on the MLX-VLM engine, and MLX-VLM does not yet implement every sampling method the broader local-LLM stack does. Quality-sensitive inference that depends on a wider menu of samplers (such as top-n-sigma, the novel sampling method introduced on arXiv in late 2024) may still be better served by running a GGUF model through llama.cpp or Ollama. Nativ is best understood as the MLX-native option for Apple Silicon, not a one-stop replacement for the whole local-LLM stack.
What This Means
Nativ is not the first local-LLM GUI on macOS. LM Studio has owned that slot for two years, Ollama has the CLI crowd, and oMLX is the open-source option that several self-hosters already recommend. What Nativ adds is a single binary, written in Swift, that combines (a) a chat UI with live telemetry, (b) OpenAI- and Anthropic-compatible endpoints for coding agents, and (c) a maintained MLX backend from the developer whose Python package LM Studio already relies on for vision work. If you have an M-series Mac and you would rather not run an Electron app or a Docker container just to talk to a local model, Nativ is now the most direct path.
For the privacy-focused intelligibberish audience, the deeper point is that the local-AI toolchain on Apple Silicon is finally catching up to what Linux and Windows users have had for a year. The Nativ v0.0.1 release on July 20 is an early first build; when Simon Willison looked at it the same day, the architecture was in place even though the project had only a few hours of commit history behind it.
The Bottom Line
If you own an Apple Silicon Mac and you want a native, MIT-licensed chat app that doubles as a localhost API for Claude Code or Codex, install Nativ v0.0.1, point an existing MLX model at it, and reroute your coding agent to 127.0.0.1:8080. Your prompts stay on your machine, the project is open source, and the same MLX-VLM engine is already doing vision work inside LM Studio.