AI-powered interfaces, creative tools, and macOS utilities today.
LobeHub — Your Chief Agent Operator
Describe a goal, and LobeHub assembles the right agents, runs tasks in parallel across models, and reports back via Slack, Discord, or Telegram. 273K+ skills, 51K+ MCPs. Less tab-switching, more outcomes.
Shadow V2 — AI Interface for Your Mac
Sees your screen, hears your voice, and runs custom prompts on a keyboard shortcut. Quick Reply drafts emails from context, Voice Typing turns speech into clean text. Every skill is editable. Local transcription, no bot in calls.
Moody — Music-Reactive Mac Wallpaper
Six smart modes that react to your music (Spotify/Apple Music), weather, Focus mode, time of day, and even CPU load. 4K wallpapers, multi-monitor, SwiftUI native. macOS 14.6+.
Krea 2 — Image Model for Style Control
Krea's first foundation image model, built around aesthetics and creative control. Style references, moodboards, and strength controls let you guide the final result instead of relying on vague style words.
QuickRight — Missing Finder Right-Click Features
Create new files instantly, true Cut & Paste, open Terminal or Warp in current folder, copy file paths, compress images, extract colors, check hashes — all from Finder's context menu.
Agentspan — Open-Source Durable Agent Runtime
MIT-licensed server and SDK for running AI agents as durable workflows. Crash recovery, human-in-the-loop approvals, guardrails, tool history, and observability. Framework and model agnostic.
Musk's OpenAI lawsuit is over, Qwen 3.7 surfaces, and Cloudflare shares their experience with Anthropic's Mythos.
A jury in Oakland ruled that Musk waited too long to sue, finding his claims time-barred by the statute of limitations after just 90 minutes of deliberation. Judge Yvonne Gonzalez Rogers agreed with the jury's advisory verdict.
Musk had alleged OpenAI "stole a charity" when it shifted to a for-profit structure. His attorneys confirmed an appeal to the Ninth Circuit. Musk called the ruling "a free license to loot charities."
Qwen 3.7 Spotted on Qwen Chat
Community testers found Qwen3.7-Max-Preview and Qwen3.7-Plus-Preview appearing in Qwen Chat ahead of Alibaba Cloud's May 20 summit. No official weights yet, but early impressions are circulating. Open-weight release expected soon.
Gemini 3.2 Flash Solves IMO 2025 Problem 6
Community testing suggests Google's Gemini 3.2 Flash can solve IMO 2025 Problem 6 — previously only GPT-5.5 Pro could do this without scaffolding. A Flash-tier model matching frontier reasoning on elite math is notable.
Dario Amodei: AI Will Bring High GDP Growth and High Unemployment Simultaneously
Anthropic's CEO predicts 5-10% GDP growth alongside 10%+ unemployment — "a combination never seen before." Says half of entry-level white-collar jobs could vanish within five years while the economy booms.
Cloudflare Publishes Mythos Preview Audit Results
After running Anthropic's Mythos Preview against 50+ internal repos, Cloudflare shared an honest breakdown. The model reasons about exploit chains like a senior researcher, but generic "find all vulns" prompts don't work at scale — you need narrow, parallel investigations.
SmallCode: 87% Coding Benchmarks with a 4B-Active Model
A terminal-native coding agent designed from the ground up for small local models. Parses tool calls from JSON/YAML/XML, auto-repairs mistakes, decomposes tasks into atomic steps. 87/100 benchmark tasks pass with Gemma 4 (4B active) vs ~75% for OpenCode with 14B models.
Quick hits.
BitCPM4-CANN: New 1-Bit Models from OpenBMB
Three BitNet models (1B, 3B, 8B) designed for efficient inference. Waiting for llama.cpp support.
Linus Torvalds on "Unmanageable" AI Bug Reports
Linux maintainers struggling with flood of AI-generated bug reports that are plausible but wrong.
ByteDance drops a unified 3B model, and ComfyUI gets interactive warping.
Lance by ByteDance: Unified 3B Model for Image & Video
A lightweight 3B-parameter model that handles image generation, video generation, image editing, video editing, and multimodal understanding in a single framework. Trained from scratch on 128 A100s. Apache 2.0 license. Open weights on HuggingFace.
Olm Liquify: Photoshop-Style Warping Inside ComfyUI
New custom node that brings interactive real-time image warping directly into ComfyUI. Useful for cleanup, proportions, face adjustments, and stylization tweaks without leaving your workflow.
LTX Director Tutorial: Advanced LTX 2.3 Videos in ComfyUI
Full walkthrough covering setup and all main features of the LTX Director node for creating advanced video content.
Native macOS editors, self-hosted terminals, and mesh VPNs.
Editorio: Native macOS Markdown & Code Editor, Free Forever
Built entirely with AppKit — no Electron, no web views. Markdown preview with CommonMark/GFM, syntax highlighting for 180+ languages, Sublime-style minimap, export to PDF. Completely free, no in-app purchases.
macUSB v2.2: All-in-One USB Creator for Mac
Now supports Linux and Windows USB installers alongside macOS. Auto-detects distribution, architecture, and version. Handles Apple Silicon legacy signing issues. Free and open-source.
Cotypist Pricing Revealed: $8/mo or $12/mo
The local AI autocomplete for Mac now has pricing: Free tier, Plus at $8/mo, Pro at $12/mo. Runs entirely on-device (Apple Silicon only), predicts your next words inline, and catches typos. 25% launch discount on first bill.
Terminator: Open-Source SSH Client with Self-Hosted Sync
Free alternative to Termius with a self-hostable sync server. Keep your SSH connections, keys, and configs in sync across machines without paying for cloud sync. Open-source desktop app + server.
Nylon: Self-Hosted Mesh VPN with Latency-Aware Routing
Connects cloud servers, mobile devices, and workstations in one unified VPN. Routes via the lowest-latency LAN path when on the same network, falls back to WAN when needed. Built for game streaming and latency-sensitive work.
AutoShelf: Auto-Organize Files with Rules
Watch folders, match conditions (extension, name, date), move or rename automatically. Free with 1 rule, $19.99 one-time Pro.