From Product Hunt
AI-native tooling keeps pouring in: a comms layer that drops agents into Slack and WhatsApp, ads inside the Claude Code spinner, a Claude-powered Premiere editor, and an open agent-observability stack — plus a couple of beautifully un-AI side projects.
Novu Connect — ship agents where your users already work
Open-source communication infrastructure that gives any AI agent two-way conversations across Slack, Teams, WhatsApp, Telegram and email — you bring the agent logic (Claude agents, AI SDK, LangGraph or custom), Novu handles delivery, identity resolution, threading and per-channel formatting. Runs on SOC 2 / HIPAA / ISO infra; prompt it to add an agent and it's live in under a minute.
Kickbacks.ai — get paid to wait for Claude Code to finish
Sells the little "thinking…" line inside the Claude Code and Codex spinners as a tiny, tasteful sponsored slot, auctioned to advertisers — and pays the developer whose machine showed it. The newer kickbacks.dev rebuild bumps the split to 70% of net ad revenue, with self-serve Stripe payouts over a $10 threshold. A genuinely strange new market: monetizing agent dead time.
AutoEdit — a Claude AI video editor inside Premiere Pro
A Premiere Pro plugin that transcribes your footage with ElevenLabs, sends the transcript to Claude, and builds a rough cut by meaning rather than waveform — auto-removing silences, filler words, bad takes and restarts, then generating captions and structuring the timeline. Editors keep every creative call; the plugin just kills the boring cleanup.
PandaProbe Cloud — agent engineering, fully managed
The managed version of the open-source agent-observability platform: full-stack tracing (sessions, traces, spans across LLMs, tools, sub-agents), agent-specific evals, and scheduled monitoring to catch quality regressions in production. Integrates with LangGraph, CrewAI and the Claude Agent SDK; self-host the OSS core or run it in the cloud.
MockPilot — turn live websites into editable mockups
A desktop app that pulls any live webpage onto an editable canvas, then lets you restyle it with natural-language commands — it extracts the site's design system automatically and exports clean HTML. Useful for fast redesigns, A/B mockups, or lifting a layout you like into something you can actually edit.
Synopsule — on-device private meeting transcripts
Records and transcribes meetings entirely on your Mac or iPhone with Whisper — no accounts, no uploads, with speaker labels, replayable audio and open Markdown files. Templated summaries run on-device via Apple Intelligence (or your own key), and a built-in MCP server lets Claude, Codex and ChatGPT read your transcripts locally. $1.99 one-time, covers Mac + iPhone.
Reignat — privacy-friendly web analytics for makers
A cookie-free, GA4-alternative analytics platform aimed at founders and indie makers: realtime visitors, funnels, visitor profiles and Core Web Vitals without the surveillance or the complexity. A cleaner Plausible-style option if you want clear traffic insight without consent banners.
MiMo Code — a coding agent with explicit long-term memory
Xiaomi's open-source (MIT) terminal coding agent, built on OpenCode for long-horizon tasks. Persists knowledge across sessions in plain files — project memory, auto-written checkpoints, scratch notes and per-task progress — with context reconstruction when the window fills, and a weekly
/dream agent that compresses old memory. Xiaomi claims a 65%+ win rate over Claude Code past 200 steps.
Tinfoil Pigeons — the aircraft overhead, on a retro radar scope
Enter your postcode and watch the flights above you on a CRT-style phosphor radar — the sweep, the glow, blips fading between passes — then tap one to pull its type, operator, registration, altitude, speed and heading. A non-commercial side project built on Astro + Cloudflare, fed by the community ADS-B network. Pure romance over yet another Google map.
Verol — real-time AI hallucination & fact checker
Connects to ChatGPT, Claude or Gemini and fact-checks responses as they appear: it extracts each factual claim, cross-references trusted sources, and shows inline badges for verified, unverified or false statements — so you catch fabrications before you ship or share AI output.
From Reddit
The Fable 5 suspension keeps reverberating through open-source land — a takedown-resilient model archive and a guerrilla Fable distill both landed this week. Plus a landmark AI-liability ruling and the opening shots of a price war.
With hostility toward local and decensored models rising — Heretic itself has already drawn a legal notice from Meta — the project launched a local-first backup system designed to keep its models reachable no matter what happens to GitHub or Hugging Face.
The new official site at heretic-project.org acts as a permanent pointer to wherever each resource currently lives, with redundant install sources, a full tutorial and searchable docs for every config parameter. It's a direct response to the single point of failure in centralized model hosting — and a sign of where the open-weight community's head is at after the Fable suspension.
"Stop using Ollama"
A widely-shared critique arguing Ollama built its popularity on llama.cpp while obscuring that dependence, muddying which model you're actually running, and steering toward proprietary cloud services. Its prescription: go straight to llama.cpp (with llama-swap), or use honest wrappers like LM Studio, Jan or koboldcpp. The comments push back — Ollama still wins on UX — making this a useful read on the real tradeoffs.
German court holds Google liable for its AI Overviews
In what looks like the first ruling to hold an AI firm liable for AI-generated speech, a German court ordered Google to stop spreading false claims its AI Overviews made about two publishers (calling them scams). The court's reasoning — "nobody needs AI to search the internet" — denies AI summaries the legal protections traditional search enjoys, since they're an optional add-on. Potentially global implications.
Qwable-v1 — an open distill of the now-suspended Claude Fable 5
An open-weights Qwen3.6-35B-A3B distilled from the ~4,659 cleartext agentic-coding traces that leaked during Fable 5's brief public window, before Anthropic's anti-distillation classifier and the government suspension shut it down. Notably, Fable's tool surface bled into the weights — it emits Claude-flavored
<tool_use> XML calling tools like str_replace_editor. GGUFs and the SFT dataset are public; the community is skeptical it's more than a curiosity.
OpenAI weighs drastic price cuts to win users from Anthropic
Per the WSJ, OpenAI is considering significant token price cuts if Anthropic moves first — early deliberations, but a sign of a looming price war as both head toward IPOs and Anthropic's valuation pulls ahead.
From Reddit
Now that Ideogram 4 is open-weight, a whole ecosystem of JSON-prompt tooling is forming around it — and LTX teases the next generation of its open video model.
Ideogram 4 Autoprompter — AI writes the structured JSON for you
A ComfyUI custom node that turns a plain idea (or a reference image) into a complete Ideogram 4 JSON prompt — numbered regions, bounding boxes, style, aesthetics, lighting and medium all filled in — which you then edit by hand, Kijai-style. Ideogram 4's prompt adherence is excellent but only with proper structure; this removes the tedious part. Local, API or hosted engine options.
Okims JSON Builder V2 — a visual Ideogram prompt editor in ComfyUI
A complementary take: a ComfyUI node that opens a fullscreen HTML builder with a drag-and-resize bounding-box canvas, an impact-grid layout guide, presets, auto-saved state and dark/light themes, then outputs a clean ComfyUI string. Built for hand-crafting structured Ideogram-style JSON prompts and bbox layouts; a standalone HTML version is included.
LTX teases the next LTX-2 — dense + MoE variants
In a "what's next" post, Lightricks' CEO laid out the next LTX-2 release: two architectural flavors (a dense model and a mixture-of-experts that activates only the parts it needs, scaling quality without linear compute), plus a significantly stronger text encoder for complex multi-shot prompts. LTX-2 is the open-source, synchronized audio+video model with native 4K.
Tested 50+ gen-AI video tools — match the model to the job
A practical breakdown after a year of testing: paid social / UGC ads → Zeely or Creatify; music videos → Neural Frames (stem-synced visuals); cinematic film & shorts → Kling 3.0 or Veo 3.1. The takeaway: a model that's great for a film usually flops at a TikTok ad.
From Reddit
A strong week for native Mac tooling and open-source self-hosting — from a reverse-engineered $7 smart ring to an infinite-canvas IDE.
Tired of $300 wearables with $10/mo subscriptions piping heart-rate and sleep data to a vendor cloud, the maker bought a generic Chinese BLE ring off Temu, sniffed its Bluetooth packets, reverse-engineered the protocol, and built an open-source iOS app around it.
PulseLoop decodes heart rate, SpO₂, steps, sleep stages and raw packets into native SwiftUI dashboards backed by SwiftData — entirely local, no account. On top sits an AI coach (bring your own API key) with tools to query your data, run on-device analysis, draw charts and remember context. The thesis: a $20 ring plus an LLM can replace a subscription wearable, with the intelligence on your phone, not behind a paywall.
Cate — an infinite-canvas IDE for agentic coding
An open-source desktop IDE that spreads editors, terminals, browsers, docs and AI agents across a freeform zoomable canvas instead of stacked tabs — float them, dock them into splits, or detach to OS windows, and it restores the whole layout per folder. Built-in coding agent (Pi) with per-chat model memory; connect Anthropic, Codex, Copilot, Gemini, OpenRouter, Groq, Mistral or DeepSeek via OAuth or API key.
Murmur — local AI text-to-speech studio for Apple Silicon
An on-device macOS TTS studio built on Apple's MLX, bundling four models (Kokoro, Qwen3-TTS, Fish S2 Pro, Chatterbox), 860+ voices and voice cloning — 100% offline, no account or per-character meter. Built for the iterate-on-a-draft stage: rewrite a sentence, regenerate a paragraph, audition voices, then export WAV/M4A. One-time pricing; keep private/client scripts off the cloud.
RevPDF 4.5.0 — a fully offline PDF editor for Mac
Its biggest release yet, and everything happens on-device (free, no login, no cloud): auto-redaction that permanently strips sensitive content, a tabbed multi-PDF viewer, document-wide find & replace, automatic form-field detection for documents that were never built as AcroForms, a metadata editor, better font matching, and split/merge. A real local alternative for contracts and invoices you don't want on someone's server.
AltTab fork — app bindings, appearance controls, no license gate
A fork of the popular Windows-style window switcher for macOS that adds per-app keybindings and appearance controls, and removes the license gate. Worth a look if you want true Alt-Tab window switching with deeper customization than the original exposes.
Claude Cap — track your Claude.ai usage from the menu bar
A lightweight macOS menu bar app that surfaces your Claude.ai usage against your limits at a glance — handy if you're on Max and want to keep an eye on where you stand without digging through settings.
Lunarr — an open-source self-hosted Plex/Jellyfin alternative
A new media-server project written in Go, positioned as a lightweight open-source alternative to Plex and Jellyfin. Early but worth tracking if you're always testing what's next for your home server stack.
OpenLogi — an open-source alternative to Logitech Options+
Configure your Logitech mouse and keyboard buttons, gestures and scrolling without the bloated official app. Open source, for people who'd rather not run Options+.