From Product Hunt
Claude Code gets a menu bar companion, and several sharp native Mac utilities launched today.
Vibedock — Toggle Claude Code MCP Servers from Your Menu Bar
Every MCP server you configure injects tool definitions into your context window on every message. Vibedock sits in your macOS menu bar and lets you toggle MCPs on/off per project with one click, automatically restarting Claude sessions. Saves tokens and reduces context bloat.
Memdex — Cross-Model AI Memory
Chrome extension that auto-captures your chats across ChatGPT, Claude, and Gemini, stores them encrypted locally (IndexedDB), and injects relevant context into new conversations like Grammarly for memory. One memory across all AI tools, no cloud.
RetroMac — CRT, VHS & Game Boy Shaders for Your Mac
30+ real-time shader presets (CRT, PAL, NTSC, VHS, retro LCD) applied to your whole screen or a single window. Swap your Dock for classic Mac/Windows looks. Free with optional one-time upgrade.
FinderLock — Lock Files in Finder with Touch ID
Native macOS app for locking individual files and folders with AES-256 encryption. Unlike FileVault (full-disk), this protects per-file. Double-click to unlock, auto-lock on idle. No account, no cloud, one-time purchase.
Coca 2.0 — Keep Your Mac (and Status) Awake
Menu bar utility to prevent sleep, display dimming, and screen savers. New "Deep Control" feature keeps you green in Teams/Slack by subtly moving the mouse when idle. Closed-lid mode, Shortcuts support. Free.
From Reddit
Anthropic's Mythos hits a milestone, OpenAI Codex learns to work while your Mac sleeps, and llama.cpp gains agentic tools.
Anthropic says its Mythos model has already discovered more than 10,000 vulnerabilities through Project Glasswing, its automated security research program. Meanwhile, Mythos 1 has been spotted as an available model in Claude Code, suggesting a broader rollout is underway.
OpenAI Codex Can Now Control Your Mac Even When It's Locked
Codex gained the ability to run tasks on macOS with the screen locked. It can execute code, navigate files, and complete multi-step workflows without requiring the user to stay at the keyboard. A step toward always-on AI computer use.
GPT-5.5's "Secret Sauce" Appears to Be Caveman-Mode Thinking
A user captured GPT-5.5 leaking its internal reasoning trace during a normal conversation. The thinking style reads remarkably like the "caveman mode" prompting fad from months ago — terse, compressed reasoning tokens. Suggests OpenAI may have trained on this pattern for better token efficiency.
llama.cpp Server Now Has Built-In Agentic Tools
The llama.cpp server quietly added native tool support via an experimental --tools flag: read_file, file_glob_search, grep_search, exec_shell, and edit_file. Essentially Claude Code-style agentic capabilities running locally with any GGUF model.
llama.cpp b9297: NVFP4 + MTP Now Work Together
The latest release combines NVIDIA FP4 quantization with multi-token prediction for significant speed gains on supported hardware.
Update: Command A+ (218B MoE) Now Runs on Apple Silicon via MLX
Following the May 20 release, a community member ported Cohere's 218B/25B-active MoE model to mlx-lm. Sliding window attention, sigmoid routing, Apache 2.0.
From Reddit
New ComfyUI workflows for style transfer and resolution upscaling, plus Kling reaches Hollywood production.
Untwisting RoPE: Training-Free Style Transfer in ComfyUI
ComfyUI implementation of a new technique that enables style transfer without any training or fine-tuning. By manipulating rotary position embeddings, it transfers artistic styles from a reference image to generated output. Currently supports Z-Image Turbo.
SEGA: Resolution Extrapolation for Diffusion Transformers
New paper introducing Spectral-Energy Guided Attention that enables generating images at resolutions far beyond training resolution. Works with existing DiT models without retraining. Results show clean upscaling from 1024px base to much higher resolutions.
Kling AI Used in "House of David" — First Hollywood Show to Go Public About AI Video
House of David, which hit #1 on Prime Video and reached 44M viewers, is the first Hollywood production to openly discuss using AI video generation (Kling) at industrial scale in its production process.
LongCat Video Avatar 1.5 — Open-Weight Talking Head Model
Meituan released an upgraded audio-driven avatar model. Commercial-grade quality, supports audio-to-video, image-to-video, and video continuation.
From Reddit
A Liquid Glass video player, self-hosted tools, and an open-source graphics editor update.
Vidi — Native macOS Video Player with Liquid Glass Design
New video player built around macOS Tahoe's Liquid Glass. Plays everything (MKV, AVI, HEVC, HDR, Dolby Vision), with spatial audio, PiP, AirPlay, and Chromecast. Apple Silicon optimized, free on the App Store with a Pro tier for casting.
Homarr v1.62 — 40-60% Less Memory
The self-hosted dashboard app addressed its biggest criticism. By merging Node.js processes and tweaking Next.js settings, memory usage dropped by half. Performance and speed unchanged.
VaulTLS — Self-Hosted Certificate Management
Lightweight web app that spins up your own CA on first boot. Clean UI for issuing and managing TLS certificates for internal services. Docker compose, done. No more self-signed cert headaches for homelab HTTPS.
Graphite — Open-Source 2D Graphics Editor, May 2026 Update
Major update with vector blending modes, gradient overhaul, draggable panels, and 500+ changes. Rust-based, runs in the browser. Aiming to be a procedural, non-destructive alternative to Illustrator.
PureMac — Open-Source CleanMyMac Alternative
Native SwiftUI, zero telemetry, MIT licensed. Cleans system/user caches, Xcode/Homebrew junk, large files. Signed and notarized.