From Product Hunt
Local-first AI ran the table — on-device Mac autocomplete, a terminal that thinks offline, a ~285ms computer-use agent, and a Claude Code context-hygiene hook — plus Valve's new Steam Machine for the living room.
Cotypist — local AI autocomplete in your voice, anywhere on your Mac
Smart autocomplete for every app you type in (Mail, Slack, Notes, docs, even AI prompts): press Tab to accept suggestions that update in real time. Runs fully on-device on Apple Silicon (default model Gemma 4 E2B, ~1GB active) — no cloud, no API — and learns your writing voice across apps. Freemium; Plus $6/mo, Pro $9/mo.
Steam Machine — Valve's tiny, powerful living-room PC is back
A roughly 6-inch cube running SteamOS on a semi-custom AMD Zen 4 6-core CPU + RDNA 3 GPU (28 CUs), targeting 4K/60 via FSR. A fully open, Linux-based console-PC you can also use as a desktop. Launches June 30 from $1,049 (512GB) / $1,349 (2TB); reservations through June 25 to curb scalpers.
NeuralAgent 3.0 — open-source computer-use agent that drives your desktop
Lives on your desktop and operates it like a human — seeing the screen and clicking, typing, scrolling, navigating apps and the browser. 3.0 adds a self-supervising model executing UI actions in ~285ms, saveable Fast Model Replays, and smart routing. Runs locally (no VM/cloud) on macOS/Windows; works with Claude, GPT, Gemini, Ollama, Bedrock. Electron + React + FastAPI, fully open-source.
Sipcode — keep Claude Code's context clean for sharper answers
A free, MIT-licensed Claude Code PreToolUse hook that caps verbose tool output (git log, npm install, grep) and dedupes same-session file re-reads so the model sees signal over noise. Exposes 15 MCP tools for context-hygiene stats, makes zero network calls, and claims a 62.6% median tool-output reduction across a 20-task benchmark.
jebi — a supercharged Mac terminal with built-in local AI
Free, open-source macOS terminal with on-device models (Qwen, Phi-3, Gemma) — no API key, subscription, or cloud. It explains failed commands in plain English, suggests next steps, and answers via
/ask, intentionally only speaking up after errors. Ships split panes, tabs, themes, and slash commands; installs via Homebrew on Apple Silicon.
Conduit — an SSH/RDP/VNC manager that lets your agent drive your infra
Cross-platform connection manager with an AES-256 credential vault and a built-in native MCP server (50-60+ tools) so Claude Code or Codex can run SSH commands, screenshot RDP/VNC desktops, and fill web forms — with approval, and without ever seeing your keys or proxying your AI calls. Scopes tools to your existing connections to fight MCP tool-list bloat. Apache 2.0, free for personal use.
Hush — open-source noise suppression built for voice AI agents
The first speech-enhancement model aimed specifically at voice AI, adding real-time background-speaker suppression (not just speech-vs-noise) for phone agents, call bots, and transcription pipelines. A tiny 8MB / 1.8M-param model that runs fully on CPU in real time (<1ms per 10ms of audio), language-agnostic, built on DeepFilterNet3. Apache 2.0 on GitHub + Hugging Face.
Latitude — open-source observability & evals for AI agents
Captures production traces (multi-turn sessions, tool calls, prompts, latency, tokens, cost) and builds an eval-driven reliability loop: domain experts annotate failures, evals auto-generate and run continuously, and an alignment score tracks drift from human judgment. No proprietary format — use the SDK or point an existing OpenTelemetry pipeline at it. MIT, 3.7k+ stars.
From Reddit
An IPO-shadowed conflict probe into Sam Altman, the first lawsuit over Anthropic's Fable 5 export ban, and a wave of slipping release dates (GPT-5.6, Gemini 3.5 Pro) — plus two local-inference field reports and humanoids on a real factory line.
A court filing (surfaced in the Musk trial) shows Sam Altman holds no direct OpenAI equity yet has stakes exceeding $2B in companies that have done or discussed deals with OpenAI — most prominently a ~$1.7B position in fusion startup Helion, where he reportedly pushed OpenAI toward a ~$500M commitment, plus Stoke Space (data centers in space) and BCI startup Merge Labs.
The House Oversight Committee and six Republican state AGs are pressing the SEC to scrutinize OpenAI's governance ahead of its confidential S-1. The structural oddity: Altman profits when portfolio companies win OpenAI contracts but gains nothing from OpenAI's own success. He left Helion's board in March and says he recused himself from the dealings.
First lawsuit over the Fable 5 export ban: a legal-AI startup sues the US government
Following the ongoing Mythos/Fable 5 saga, Legion Legalist filed suit against the administration in DC federal court, arguing Commerce lacked authority under IEEPA to force Anthropic to pull Claude Fable 5 (released June 9, taken down June 12) — IEEPA doesn't cover access to a hosted model or its text outputs, and no national emergency was declared. Legion says losing Fable 5 caused "immediate, irreparable and existential" harm, and that the move contradicts Trump's own June 2 EO. The first legal challenge to the directive.
June slips: GPT-5.6 pushed to mid-July, Gemini 3.5 Pro held, OpenAI "Bidi" voice incoming
Update on the GPT-5.6 final-testing chatter: a leaker (corroborated by Polymarket and prior accurate scoops) says GPT-5.6 won't ship this week and now targets ~mid-July, DeepMind is unhappy with Gemini 3.5 Pro and won't launch it this month, and OpenAI is prepping "Bidi" (GPT-Bidi-1) — a bidirectional voice model that listens and speaks simultaneously, rides over interruptions, and replaces the ~2-year-old Advanced Voice Mode. Also notes Claude Sonnet 5 is now in limited enterprise Early Access as a stop-gap while Mythos/Fable 5 stay blocked.
Why SWA-hybrid models win on long context: MiMo 2.5 vs MiniMax M3 & DeepSeek V4 (local)
A hands-on long-context decode benchmark on 2× RTX PRO 6000 (consumer Blackwell): MiMo 2.5 (IQ4_XS) holds ~105 tok/s at 8K and stays usable (~62 tok/s) out to ~190K because, like Gemma 3, it uses a 5:1 local/global sliding-window attention pattern (~7× smaller KV cache). MiniMax M3 and DeepSeek V4 fall off a cliff — their custom kernels target datacenter Blackwell (B200/SM100), so M3 silently drops to dense attention and V4 ops fall back to CPU (~14 tok/s). Practical takeaway for prosumer rigs chasing a "local Sonnet."
MiniMax 2.7 agentic running local at ~47 tok/s gen, 1200 tok/s prefill
A build report running MiniMax 2.7 (REAP Q4) on a 96GB-VRAM + 192GB DDR5 rig, praised for instruction-following and tool-calling. Notable for an unusual multi-agent setup: three CPU-bound "dreamer" agents with 20-40K-token system prompts in a round-robin loop, plus a dense 12B watchdog asynchronously flagging errors.
AGIBOT livestreams G2 humanoids working a real tablet production line (Jun 23-28)
A six-day global livestream of multiple G2 humanoids running a quality-inspection section inside Longcheer's live tablet line in Nanchang — picking devices, loading test fixtures, sorting good/defective units alongside humans. AGIBOT claims ~310 units/hour and a 99%+ success rate over a 140-hour trial. A rare raw, long-duration look rather than a staged demo.
From Reddit
Krea 2's open weights got their official Hugging Face release a day after the magnet-link drop, and the ComfyUI community shipped day-one control nodes and quants — plus a 1.8-second video generator and an Ideogram 4 storyboard tool.
Update: following yesterday's magnet-link/FP8 leak, Krea officially open-sourced Krea 2 — live on Hugging Face (krea/Krea-2-Raw, krea/Krea-2-Turbo) and Comfy-Org, under a new "Krea 2 Community License." It's a 12B-parameter Diffusion Transformer in two variants: Raw is the undistilled base (~52 steps, CFG 3.5, up to 1024px; meant for LoRA training/fine-tuning) and Turbo is an 8-step distilled checkpoint (CFG 0, up to 2048px) for everyday fast generation.
The intended workflow is train LoRAs on Raw, run them on Turbo (LoRAs transfer between both). Tuned for natural language; wrap words in quotes for in-image text. Krea claims #1 text-to-image from an independent lab on Artificial Analysis. The Comfy-Org build uses a Qwen3-VL-4B text encoder and Qwen-Image VAE.
Day-one Krea 2 control for ComfyUI: bypass the built-in safety filter + bbox regional prompts
Two practical workflows. ConditioningKrea2Rebalance is a custom node doing per-layer conditioning reweighting that neutralizes Krea 2's trained (Flux.1-Kontext-style) safety filter — killing false-positive prompt refusals and tightening adherence; a posted comparison runs 12 steps, Euler simple, CFG 1, 2048×1024 via INT8-ConvRot on a 3080 10GB at ~2.5s/step. Separately, Kijai demoed Ideogram-style bounding-box regional prompting on Krea 2 for spatial layout control — handy for composing graphics with placed elements.
FastWan-QAD — 5 seconds of 480p video in 1.8s on a single RTX 5090
Hao AI Lab (UCSD) released a Quantization-Aware Distillation of Wan2.1-T2V-1.3B that generates 5s of 480p video in 1.8s end-to-end on a 5090 — >3× faster than the prior best. Three checkpoints target different cards: QAD-1.3B (NVFP4 + SageAttention3, Blackwell), QAD-1.3B-SA2 (NVFP4 + SA2++, sharper), and FP8-1.3B (for 4090 and older without FP4 tensor cores). I2V distillation and scaling to Wan2.1-14B are next; code lands in the FastVideo repo.
Ideogram Storyboard Builder — an open-source app for Ideogram 4 JSON prompts
An agentic Electron app for building ordered image storyboards through ComfyUI, with first-class support for Ideogram 4's structured JSON prompt format for consistent multi-shot scenes. Stack: Bun + Hono, React 19/Zustand, SQLite/Drizzle. MIT-licensed — practical for character/scene-consistent sequences and concept boards.
From Reddit
A run of tactile and local-first tools — trackpad haptics for links, a CSV editor that won't mangle your data, a self-hosted trip planner with its own MCP server, and browser-native CAD — plus open eReader firmware and a throwaway VPN.
HapticPad — feel links and buttons through your trackpad
A menu-bar utility that repurposes the MacBook trackpad's haptic motor to vibrate whenever your cursor hovers a link or button, so you "feel" page structure instead of visually hunting for it. Genuinely novel, no real competitor. $5 one-time on the Mac App Store, 7-day trial.
SmoothCSV — the CSV editor that never silently reformats your data
A grid-first CSV editor that won't eat leading zeroes or turn text into dates. v3 is a Tauri rewrite (~80MB, opens 100MB files instantly, claimed 12× faster than Excel) with a command palette, SQL queries, and sort/filter. Cross-platform (macOS/Win/Linux), not Electron.
TREK — self-hosted realtime trip planner with a built-in MCP server
A day-by-day itinerary planner (formerly "Nomad," fully rewritten) with live multi-user collaboration on a shared map, interactive maps, budgets, packing lists, SSO and PWA support — and a built-in MCP server so you can point Claude, Cursor, or a local LLM at your instance and plan trips conversationally. All data stays on your box. AGPL-3.0, 3.3k+ stars.
CrossPoint Reader — open-source firmware for cheap Xteink eReaders, with Calibre sync
Community-built, fully hackable firmware for the ESP32-based Xteink X3/X4 e-paper readers. The reader engine does EPUB 2/3 with hyphenation/kerning, footnotes, bookmarks, auto page-turn, and KOReader progress sync; drop EPUBs in from any browser or use the included Calibre device plugin. A fun hardware-hack pick for self-hosted reading.
NASSCAD — a full parametric CAD that runs 100% locally in a single HTML file
Surprisingly serious offline browser CAD: STEP AP242 B-Rep import/export via an OpenCascade WASM pipeline (targets FreeCAD/Fusion/CATIA compatibility), a Manifold CSG engine for watertight boolean ops on multiple web workers, and parametric generators for ISO threads, bolts, nuts, gears, and pulleys — no login, no cloud, exports STL/OBJ/3MF/GLB/PLY.
Koryomi — an all-in-one self-hosted "*arr stack" for manga
A single Docker container that replaces a Suwayomi + Komga + reader-app stack: discovers series, auto-downloads new chapters (MangaDex built in, plus paste-URL sources), and serves a vertical webtoon reader with resume, multi-user accounts, 2FA, and PWA/offline. Open source (MPL-2.0),
docker compose up setup.
Mayfly — a one-click ephemeral VPN you spin up on your own VPS
A WireGuard-based tool to stand up a throwaway VPN on any cheap VPS in seconds and tear it down just as fast — built as a response to looming commercial-VPN restrictions. Early-stage and self-hosted. (A commenter flagged a committed binary in the repo — worth a glance — but the concept is solid.)