From Product Hunt

Coding agent tools, AR glasses for devs, and a Mac security auditor stood out today.

Spotlight by Backplanes — Session reports for Claude Code & Codex Air traffic control for AI agents. Automatically generates reports from your coding sessions — tracks files accessed, commands executed, external services reached, and scope drift. Free for individuals and teams.
▲ 356 · backplanes.com
Screen Charm — Polished screen recordings in minutes Smart zoom, motion blur, webcam overlay, background music, and a built-in editor. Export up to 4K. One-time purchase, no subscription.
▲ 265 · screencharm.com
Gemini 3.5 Live Translate — Real-time speech translation in 70+ languages Google's new audio model translates live speech while preserving intonation and pacing. Works in noisy environments with overlapping voices. Live on Android/iOS, coming to Google Meet.
▲ 214 · blog.google
iArt.ai — AI agent for video and motion graphics Creates promos, explainers, kinetic type, and motion graphics with audio. Chat to refine and ship. Aims to replace After Effects and CapCut workflows.
▲ 181 · iart.ai
Monako Glass — AR glasses for hands-free coding agents 48g wearable running Buildroot Linux with waveguide display, bone conduction mic, and gesture input. Run Claude Code or Codex hands-free. Reservation-only, ships July-August 2026.
▲ 147 · monako.ai
Hero Studio Photos — AI product photography from a single snap Snap one photo and get clean studio shots from every angle, including on-model images for clothing. Comes with an API and MCP server for developer integration.
▲ 111 · herostuff.com
fort — One command to audit and fix your Mac's security Checks 15+ settings (FileVault, SIP, firewall, SSH, AirDrop, Gatekeeper), reports a score, and fixes most issues. Single binary, no telemetry, MIT licensed. brew install djadmin/tap/fort
▲ 102 · github.io
LayerProof Vellum — Visual canvas for AI image pipelines Drag-and-drop subject, background, and effects together on a canvas, and Vellum fuses them into a pixel-perfect final asset. Controlled AI image production instead of random generations.
▲ 101 · layerproof.app
From Reddit

Anthropic dropped Fable 5 — their most powerful public model yet. Google open-sourced a text diffusion model that generates blocks of text in parallel instead of one token at a time.

DiffusionGemma — Google's 26B open model generates text via diffusion, not tokens Instead of one token at a time, DiffusionGemma refines 256-token blocks in parallel — like image diffusion but for text. 26B MoE (3.8B active), Apache 2.0, fits in 18GB VRAM. 1000+ tok/s on H100. NVIDIA shipped day-1 RTX support.
OpenAI rebuilds ChatGPT memory with "Dreaming" — factual recall jumps to 83% New memory architecture: factual recall up from 67.9% to 82.8%, preference adherence from 55.3% to 71.3%. Memories auto-update over time. 2x capacity for Plus/Pro, 5x compute reduction enables free-tier rollout.
China preparing $295B for nationwide AI data center network Beijing plans ~2 trillion yuan over 5 years for interconnected AI compute hubs. Mandates 80% domestic tech (Huawei chips), squeezing out NVIDIA and AMD.