Nativ runs frontier open models locally on your Mac, MIT-licensed
Built on mlx-vlm, it ships a localhost OpenAI-compatible server to plug coding agents into — no cloud, no accounts. Plus: Claude Code caps runaway subagents.
Copy markdown
Frontier models, fully offline on Apple Silicon
Nativ (v0.0.1, MIT-licensed) is a native macOS app that runs MLX vision-language models — Google's Gemma, Cohere's North Mini Code, Liquid's LFM2.5-VL — entirely on-device, with no accounts, subscriptions, or cloud. Built by mlx-vlm maintainer Prince Canuma, it hit the Hacker News front page and a Simon Willison write-up within a day of its July 20 release.
A local OpenAI-compatible endpoint for your agents
Beyond chat, Nativ runs a localhost server that speaks the OpenAI API, so you can point Cursor, Claude Code, or your own scripts at a private on-device model instead of a paid cloud endpoint. That makes a fully offline, zero-API-bill coding-and-vision assistant a realistic weekend build on any M-series Mac.
Claude Code puts a leash on subagents
Claude Code 2.1.217 (July 21) now caps concurrently-running subagents at 20 (override via CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS) and stops subagents from spawning their own nested subagents by default — a real behavior change if your workflows fan out agents. It also adds :shortcode: emoji autocomplete and fixes a memory leak from truncated MCP tool outputs.