From Product Hunt

Today's launches were almost all agents and the plumbing around them — local coding workspaces, on-device dictation, an agent firewall, and a hardware-idea inspector.

OpenCode Superapp — Codex power with local, self-hosted models and voice A native Mac AI workspace on the OpenCode harness that runs cloud, local, or self-hosted models with full model choice. Agents touch files, Git, and terminals, respond to voice, and drive Mac apps via supervised Computer Use. Free tier plus a $49 one-time "Superpowers" unlock.
▲ 131 · opencodesuper.app
Megaphone — Open-source Mac dictation that's 100% on-device Hold Fn, speak, and Megaphone types clean text into any Mac app using Apple's SpeechAnalyzer plus on-device Foundation Models to strip fillers and run inline voice commands. No account, API key, or server — free, MIT-licensed, native Swift for Apple silicon on macOS 26.
▲ 101 · megaphone.kuber.studio
Quaso — AI automation agent across your apps or the browser A prompt-driven automation agent from the Notte team. Connects 3,000+ integrations (Gmail, Slack, Linear, Stripe) and drives a real cloud browser when no API exists. Runs on schedules and triggers, and any thread can be saved as a reusable skill.
▲ 134 · quaso.ai
canitbebuilt — Your hardware idea, inspected: verdict, BOM, 3D model Describe a hardware idea in a sentence and get a feasibility verdict, a bill of materials costed at 100 / 1,000 / 10,000 units, subsystem risk ratings, certification gates, and a 3D concept model — in about two minutes, no signup. Verdict logic encodes 23 years of hardware building; runs on GPT-5.6 Terra.
▲ 114 · canitbebuilt.ai
Cosyra 2.0 — A cloud workspace for coding agents from your phone A persistent Ubuntu cloud workspace for coding agents, accessed from your phone, with Claude Code, Codex CLI, OpenCode, and Gemini CLI pre-installed. Repos, branches, terminals, and agent work stay together — close the app and resume later without keeping a laptop awake. $29.99/mo, free 10-hour trial.
▲ 115 · cosyra.com
HOL Guard — The first firewall for AI agents An open-source security layer that sits between AI agents and your systems, blocking high-risk actions before they run — deleting production data, exfiltrating secrets, installing malicious packages, prompt injection. Protects Codex, Claude Code, Cursor, Gemini, and OpenCode, plus plugins, skills, and MCP servers. MIT-style open source with 400K+ downloads.
▲ 92 · github.com
AgentLoop — A fresh Codex worker and critic every cycle An open-source, zero-dependency Node.js tool that automates the Codex plan/build/test relay. Each cycle spawns a fresh worker plus a critic that tests output against your GUIDELINES.md rubric and writes fix notes for the next clean context, so project files carry the memory instead of a bloated chat. Runs stay local, sandboxed, and cancellable.
▲ 101 · agentloop.tools
From Reddit

The frontier labs kept hoovering up mathematicians while open-weight releases and clever consumer-GPU hacks did the heavy lifting on the local side.

Ling-3.0-flash goes live on OpenRouter, free through Aug 3 InclusionAI's Ling-3.0-flash — a hybrid-reasoning MoE built for production-scale agents, from the same company behind Qwen — landed on OpenRouter and is free through August 3. SGLang is already working on day-zero support, a strong signal that open weights are coming. Benchmarks were posted alongside the launch.
Kimi K3 scores well below the frontier on cyber, per UK AISI + CAISI A joint UK AI Security Institute and CAISI preliminary evaluation finds Moonshot's Kimi K3 trails leading US closed-weight models on cyber capability — directly relevant to the open-weights export debate, with the weights due to drop July 27. Top community read: this removes the safety rationale for blocking next week's release.
DeepSeek V4 Flash hits ~105 t/s on dual 4090D by porting Blackwell kernels to Ada A user re-implemented DeepSeek's Blackwell-only kernels (DeepGEMM, FlashInfer sparse-MLA, block-scaled FP8) in Triton for sm89, so V4 Flash runs on two 4090D 48GB cards in vLLM at ~105 t/s — 2–3× faster than llama.cpp for parallel agentic work. Includes a p2p patch and a compression step that fits the model into ~96GB VRAM.
audio.cpp 0.4 adds Higgs Audio v3 TTS and first-class GGUF across 35 model families A llama.cpp-style C++/GGML engine for local audio shipped 0.4, adding Higgs Audio v3 TTS 4B, Fish Audio S2 Pro, and Voxtral realtime ASR, with GGUF now a first-class citizen everywhere. On an RTX 5090 with Q8, Higgs TTS runs ~8.8–10× realtime and Voxtral ASR ~15.7× — local voice and audio with no Python stack.

Quick hits from the local scene.

Open benchmark compares local models and quants on SWE-bench Verified A 100-task SWE-bench Verified subset across many local models, quants, and serving configs (llama.cpp vs SGLang, BF16 vs FP8), with an interactive results site — directional given it's single-run.
OpenAI moves into legal AI directly, hiring Ironclad's Jason Boehmig Rather than leaving the vertical to Harvey and Legora, OpenAI is entering legal AI itself — the latest sign of its push into office and enterprise workflows.
From Reddit

Krea 2 tooling dominated the day, with early Flux 3 motion tests and an AMD-on-Windows training guide rounding it out.

A 281-style gallery and ComfyUI node built for Krea 2 Turbo's long context An open-source ComfyUI custom node that injects 281 dense style descriptions (400–1400 chars each) into prompts, exploiting Krea 2's long-context encoder to bypass CLIP's 77-token limit. Ships an interactive browser gallery with a same-seed split-slider and a headless batch pipeline to render one subject through all 281 styles.
Community "style pointer" wildcards for Krea 2 (the day's top GenAI post) A widely-upvoted wildcard collection of style references — painters, illustrators, anime/manga, comics — framed as "pointers to model capabilities" rather than fixed presets, with guidance on splitting prompts into separate Style and Subject blocks. A same-subject grid shows the model's stylistic range.

Quick hits.

Early Flux 3 impressions: better motion than LTX-2.3, needs storyboards Hands-on testing of Flux 3 for complex motion — visuals are strong and motion beats LTX-2.3 but trails Seedance 2; coherent action needs keyframe storyboards, and it may run on 16GB VRAM.
Guide: training a 32B FLUX.2 LoRA on a 24GB AMD 7900 XTX with native ROCm A rare AMD-on-Windows walkthrough for training a 32B FLUX.2 LoRA on consumer hardware, including the patches needed to make native ROCm cooperate.
From Reddit

A strong day for Mac utilities and small dev tools — Dock widgets, a MacUpdater successor, a local photo app, and a git TUI in Zig.

docktor — Widgets and live previews in the empty space beside the Dock A free macOS Tahoe app that fills the blank space on either side of the Dock with music, calendar, and weather widgets, plus live window-preview hover and a window-based Cmd+Tab. Built on Liquid Glass; currently a free public pre-release.
Version Tracker — A MacUpdater successor watching 47 update sources A menu bar app that keeps every app current by checking 47 sources (Homebrew, Mac App Store, npm, pip, Cargo, Docker, Adobe CC, Sparkle) against what's installed. Built by an Intego co-founder as a replacement for MacUpdater, which shuts down Dec 31, 2026.
Iris — A local Mac photo app with on-device face recognition A native photo manager from indie dev Tyler Hall that indexes photos where they already live — folders, external drives, or an existing Apple Photos library — instead of importing them. On-device face recognition, OCR/speech search, and a Photo Trails map. No cloud, no account; $19 launch price.
Scap 1.4 — Native macOS screenshot, annotation, and whiteboard tool A CleanShot X / Shottr-style capture tool leaning into creative editing: freeform drawing, a dedicated board, image layers, watermarks, QR codes, badges, custom backgrounds, and screenshot OCR. v1.4 adds image import and drag-and-drop. Mostly free, with a $19.99 lifetime unlock.
Claude Codex Battery — Menu bar usage meter for Claude Code and Codex A tiny open-source macOS menu bar app that shows your Claude Code (including Fable) and OpenAI Codex usage limits as battery icons, so remaining usage is always visible at a glance. Native, free, open source.
Voxa — Have your AI coding agent actually phone you when it's done An open-source Mac + iPhone tool: pair your phone to a running Claude Code session via QR, talk to it hands-free, and get a real incoming call (VoIP/CallKit, no carrier cost) when the agent finishes or needs input. Early alpha.
Ziggity — A fast git TUI written in Zig that fixes lazygit's annoyances A free, open-source keyboard-driven git TUI written from scratch in Zig (no libgit2). Follows lazygit's workflow but non-blocking — fetch/pull/push run off the UI thread — with in-app HTTPS credential prompts, real mouse support with OSC 52 copy, and smarter remote-branch checkout.
drawDB — Open-source, self-hosted database diagrams with SQL import/export A browser-based, account-free schema designer: visually build tables, fields, indexes, and relationships, then import/export SQL or DBML for PostgreSQL, MySQL, MariaDB, SQLite, SQL Server, and Oracle. Local persistence via IndexedDB, 50+ UI languages, and it never touches your live database.
Dawarich adds Poster Studio — printable custom maps from your location history Dawarich, the self-hosted Google Location History alternative, shipped a Poster Studio that turns tracked trips into printable custom maps with live preview, theme and layout presets, text controls, and PNG/PDF export. Posters pre-load from any trip's route and save to a server-rendered gallery.

Quick hit.

SubTunnel — A one-binary self-hosted ngrok alternative on your own domain A single ~2MB Rust binary that exposes local ports at https://myapp.your-domain.com via wildcard DNS and cert — traffic stays on your box, TLS via Let's Encrypt, MIT licensed.
From Twitter

On Twitter today, the dominant threads were Kimi K3's record-fast open-weights adoption, AI agents finding and weaponizing real 0-days, a US policy fight over Chinese open models, Black Forest Labs' unified FLUX 3, agent-driven AI video pipelines, and Codex Computer Use going mainstream.

Also on Twitter — two standalone items worth a look.