An open-source local AI app for Mac just went viral, and its manifesto calls out every competitor

Nativ's Manifesto Exposes the Open Source Local AI Divide

Nativ is an MIT-licensed macOS app that runs AI models from Google, Cohere, and Liquid AI entirely on Apple Silicon, with no accounts, no subscriptions, and no cloud 1. It handles chat, vision, audio, and code. It connects to coding agents including Claude Code and Codex through a local endpoint, letting developers swap models without changing their workflow 1. Under the hood, it runs on MLX, Apple's open machine learning framework, tuned for M-series unified memory 1.

But the reason Nativ reached the front page of Hacker News this week was not its feature list. It was the manifesto.

In its own documentation, Nativ calls competing local-AI apps "proprietary shells built on top of open-source engines they don't own" 1. The accusation is pointed: commercial local-AI products keep their interface closed, add a paywall, and count on users never examining what is underneath 1. Nativ's counter-offer is full transparency. Every line of code, every model loader, every telemetry chart is readable, forkable, and open to contributions 1.

The manifesto articulates a tension that has been building in the local-AI space for over a year: open weights and open infrastructure are not the same thing.

You can release a model under a permissive license and still wrap its execution in proprietary software that you control. The model runs on the user's hardware. The software layer between the user and that model is closed, metered, and owned by someone else. That distinction barely mattered when local AI was a niche hobby. It matters now that the local-AI ecosystem is attracting investment and a growing user base that adopted these tools specifically for privacy.

Nativ sits at one end of this spectrum. Its entire stack is open: the MLX framework it builds on, the open-weight models it runs, and the application code itself are all published under permissive licenses 1. At the other end are apps that use the same open-source inference engines but enclose their interface, add subscription tiers, and keep their code private. Both run models on your Mac. Only one lets you verify how.

For builders, the question is concrete: contribute to projects that publish their entire stack under MIT, or fork an open engine, close the UI, and charge for the wrapper? Nativ frames the first path as community-owned and the second as extractive 1. Whether you accept that framing or not, the fork-versus-contribute decision is getting harder to avoid as the ecosystem matures and the financial incentives to close-source grow.

For investors, the manifesto raises a structural question that due diligence will increasingly need to answer. If a local-AI product's core value is a graphical interface on top of an inference engine it did not build and does not own, what is the defensible technology? The open-source layer beneath most local-AI software is freely available to anyone with a GitHub account. A polished interface and smart distribution can build a business. They cannot build a moat against a competitor doing the exact same thing.

For users, the stakes are the most personal. The entire pitch of local AI is that your data never leaves your machine. But when the software mediating between you and the model is closed-source, you are trusting a developer's word that no telemetry or data forwarding is happening out of sight. Nativ pledges "no dark pattern that turns your prompts into training data" 1. The implication is stark: a closed-source app cannot offer the same verifiable guarantee, because nobody outside the company can inspect the code.

Whether Nativ sustains a contributor base large enough to maintain parity with commercial alternatives is almost beside the point. What matters is the line it drew. A fully open-source local-AI app, built on an open framework, running open models, has publicly named the gap between the open-weights movement and the closed-infrastructure products that profit from it. That gap is where the real contest over local AI will be decided.

The open-weights movement proved that frontier-grade models can run free on consumer hardware. The question Nativ's manifesto forces is whether the software layer that runs them will be free too.