From Product Hunt
Claude Code ecosystem tools dominate today's launches, alongside a solid screenshot app and an open-source health wearable API.
KarmaBox — Run AI Agents from Your Phone
Multi-agent orchestrator that turns your devices into a private compute pool. Send one message from your phone and hundreds of AI agents (Claude, Codex, Gemini) take over in parallel. No infra, no lock-in.
Open Wearables — Open Health Wearable API
Self-hosted, MIT-licensed health intelligence layer. One API for Oura, Garmin, Whoop, Apple Health, and more. Open scoring algorithms, MCP server for LLM reasoning, $0 per user at any scale.
CodeHealth MCP — Keep AI Code Maintainable
MCP server that guides coding agents to write production-ready code. Deterministic code health feedback catches technical debt before it ships. Research shows AI increases defect risk by 60% in unhealthy codebases.
noirdoc — PII Guard for Claude Code
Open-source PreToolUse hook that redacts PII before Claude Code reads it. Names, emails, IBANs become placeholders. Presidio + Flair + GLiNER ensemble runs locally. Also ships as CLI and hosted proxy.
Plannotator — Annotate Anything for AI Agents
Open-source tool to annotate docs, URLs, folders, and agent messages, then send feedback directly into your integrated agent. Works as a Claude Code skill. Runs locally and stays private.
Snapr — All-in-One Screen Capture for Mac
Screenshots, 60fps 4K recording with system audio, scroll capture, annotation, and video editing in one app. Record iPhone/iPad from Mac. One-time purchase, free version with no watermarks.
From Reddit
Mistral drops a 128B dense model with open weights, Google upgrades its TPU line, and Figure AI hits 1 robot per hour.
Mistral releases their largest open-weight model: a dense 128B with 256k context window. Scores 77.6% SWE-Bench Verified and 91.4% on agentic benchmarks. Replaces Magistral and Devstral 2. Vision support included.
API pricing at $1.5/$7.5 per M tokens. Self-hostable on 4 GPUs. Modified MIT license (commercial use requires license). Open weights on HuggingFace. Unsloth quants already available.
IBM Granite 4.1 Family — 3B, 8B, and 30B
IBM launches new open-source foundation models in three sizes. Apache 2.0 licensed. Also released Granite Speech 4.1 for audio understanding. Community reception is positive.
Ling-2.6-1T — Trillion-Parameter Model Goes Open
InclusionAI (Ant Group) releases their latest 1T model with hybrid linear attention, optimized for coding and agentic workflows. Improved inference efficiency and reduced token overhead over Ling-2.5.
Qwen Introduces FlashQLA — 2-3x Speedup for Linear Attention
High-performance linear attention kernels built on TileLang. 2-3x forward speedup, 2x backward. Designed for agentic AI on personal devices. Uses gate-driven automatic intra-card context parallelism.
Google TPU 8t/8i — Massive Compute Leap
Training cost-performance up 170-180% (2.7-2.8x), inference up 80%. Training power efficiency +124%, inference +117%. Network bandwidth 4x (100 to 400 Gb/s). Inference latency cut 56%.
Google Deep Research Max — Autonomous Research Agent
Built on Gemini 3.1 Pro. Autonomously searches the web and private data via MCP, reasons over sources, and produces fully cited reports with native charts. Two modes: fast for real-time apps, Max for comprehensive analysis.
Figure AI Hits 24x Production Scale — 1 Robot Per Hour
Figure AI dramatically scales manufacturing, teasing its growing fleet of humanoid robots. Japan Airlines will deploy humanoids at Haneda Airport next month in response to labor shortages.
OpenAI Expands to AWS — Models + Codex on Amazon Bedrock
Limited preview. Follows last week's Microsoft partnership restructuring that made OpenAI's models non-exclusive to Azure.
DeepSeek Begins Grayscale Testing for Vision
Multimodal capabilities rolling out to select users. Follows V4 launch last week.
llama.cpp Gets Native NVFP4 on Blackwell
Build b8967 adds native NVFP4 support for RTX 50-series. Benchmarks show significant speedup for Qwen3.6-27B on RTX 5090.
From Reddit
New anime model on Z-Image Base, and a slick Blender-to-AI workflow with 1:1 camera tracking.
Z-Anime — Full Anime Fine-Tune on Z-Image Base
6B parameter S3-DiT model, not a LoRA merge but a fully trained anime model. Inherits Z-Image's controllability, negative prompt support, and high fine-tuning ceiling. Distilled 8-step variant available.
Blender Layout to AI Render — 1:1 Camera Tracking
Workflow using proxy geometry in Blender with hand-keyframed F-curves as spatial reference for AI video generation. Camera trajectory, timing, and proportions all locked before AI processes frames. Practical VFX pipeline.
ComfyUI SuperNodes for LTX 2.3 — Less Spaghetti
Wrappers that compress full LTX pipeline stages into clean node interfaces. Same power, way less chaos.
From Reddit
Dock management, multi-instance apps, a Plex/Jellyfin alternative with P2P streaming, and local transcription.
Parall 2.2.2 — Run Multiple Instances of Any Mac App
Create independent shortcuts for browsers, WhatsApp, Telegram, and more. Each instance runs with separate data, separate Dock icon, and separate name. Now supports WhatsApp. Lifetime purchase.
DockPops 3.0 — iPhone-Style App Folders for Mac Dock
Organize your Dock with popup folders. v3.0 adds live Dock icon previews and individual Dock icons for each Pop. Cleaner than macOS Stacks without being tied to real folders on disk.
Hound — Self-Hosted Media Server with P2P Streaming
Open-source Plex/Jellyfin alternative that also streams via torrent or HTTP/Debrid without downloading first. Full local media library management plus instant streaming. Best of both worlds.
Lapser Studio — Beautiful Timelapse Screen Recordings
Record coding sessions as timelapses, then customize background, camera position, and foreground size. No massive video files or slow editing software needed. Native macOS app.
TranscribeX 6.0 — Live Transcription, Fully Local
Real-time transcription for meetings and recordings. Runs entirely on-device with no cloud dependency. Handles background noise and multiple speakers. Lifetime purchase.
File Deep Info 3.1.5 — Forensic File Inspector for Mac
Checksum, hex, EXIF, metadata, spoofed format detection, and privacy analysis in one tool. Inspects nested content inside Office/PDF files.