0804 | Podcast

Mia: Hey everyone, and welcome back to ProductHunt Daily. I'm Mia, and I'm joined as always by my co-host Milo.

Milo: Hey Mia! Today we've got a packed rundown of the freshest launches on Product Hunt — from a major new AI model to an open-source mobile agent, and some seriously clever gizmos in between.

Mia: We're talking a next-generation Qwen model, AI coding agents that live in the cloud instead of your laptop, a sign language interpreter built into smart glasses, and yes — even a Pokémon-style game for the idle moments while you wait on Claude.

Milo: Plus a Mac app cloner, a faceless video generator, a no-code interactive 3D platform, and a webcam instrument for playing live electronic music with your hands. Stick with us — we're about to dive in.

Mia: Big news from the Qwen team this week. They've launched Qwen3.8-Max, which they're calling their most capable model yet — and their strongest one specifically for coding and development work alongside a human. It's a huge mixture-of-experts model, with a context window large enough to hold roughly a million tokens of conversation at once.

Milo: And I want to talk about what someone actually did with it. One tester set the model loose for sixteen days straight. It started with an empty repository, built up its own coding harness, and kept working as new issues, feedback, and test results came in. By the end, that repo held 265 commits, 127 pull requests, and 151 issues.

Mia: That's a convincing long-horizon test — exactly the kind of sustained work these agent models are supposed to handle. And there's a bigger community angle here. The page notes that the API is live now on Qwen's cloud, and open weights are promised for next week, which would be the largest open-weight release in Qwen's history.

Milo: That's the part people are most excited about. One community comment summed it up: most labs gate everything behind paid APIs these days, and seeing Qwen drop full weights on GitHub feels like a genuine gift to the community. The fact that these models keep getting open-sourced at this scale, the comment said, is honestly wild. So the things to watch: whether that open release actually lands next week, and whether the delivered open version matches what the launch page is promising.

Mia: Next up is Open Minis, a free, open-source AI agent for iPhones and Android, and its creator's pitch is that it gives the model a real computer to work with. Instead of just chatting, the agent gets a sandboxed Linux shell running entirely on the device itself — it can install packages, run Python, edit actual files, and drive a browser.

Milo: So on a phone, a model can genuinely touch things, not just talk about them. And it extends well beyond the shell — the agent has more than thirty bridges that expose Apple's frameworks to the model as ordinary command-line tools. That means it can reach your health data, your calendar, your photos, HomeKit devices, and other system areas the same way it reaches a basic shell command.

Mia: What drove the developer to build it? They said every AI assistant on their phone could talk about data but never actually touch it, while the apps that could act were locked to one vendor's model and one company's cloud. Open Minis flips that with three principles: it's local-first, so keys and data stay on your device with no account, no telemetry, and no developer-controlled server in the path. Then it's bring-your-own-model — Claude, GPT, Gemini, DeepSeek, Kimi, Grok, or any OpenAI-compatible endpoint.

Milo: And it's fully open source on both iOS and Android, free with no in-app purchases. The supported model list is broad — the Claude family up through Opus 4.6, the GPT-4o and o-series, and the newer Gemini 2.5 and 3 models — and the roadmap, the developer notes, is mostly built out of the issues users file.

Mia: Two launches this cycle are pushing AI coding agents off the laptop and into the cloud. First is Murmell, a shared cloud canvas where a whole team and AI agents work together in the browser on the same repository, with a live preview — meaning closing your laptop doesn't stop the work.

Milo: The clever part is how it avoids agents tripping over each other. Before writing anything, every agent claims the file it's about to touch — one holder per path, and everyone else waits — so people and agents can move at once without overwriting each other, and the work lands back in git. Today it runs several coding agents together, with more on the way, and phone and tablet simulators listed as coming next.

Mia: The creator says the idea came from a hackathon with three people, four AI agents, and one laptop nobody dared close. He says he built most of Murmell with Claude Code, and now builds Murmell with Murmell. The launch deal is a free first week plus sixty dollars in credits, and pricing starts at 39 dollars a month for the solo plan.

Milo: Community commenters, though, pressed on the weak spots. One is what happens to a file claim when the agent holding it dies — leaving a silently stale lock on a cloud machine. Another is coordination failures, where two actors finish the same work without ever colliding. And the second cloud launch here, AgentSky, is a separate managed agent-as-a-service — taking that same idea of offloading work to the cloud and packaging it as a managed offering.

Mia: Finally, two independent products built on a similar instinct: keeping your own audio and your own conversations. yapyap is a local-first voice and meeting recorder. It records and transcribes conversations, names the speakers, and produces summaries and action items entirely on your machine — the page says no cloud, no account, no gadget needed, and everything stays on the computer.

Milo: It can record Zoom, Teams, Meet, phone calls, and in-person conversations. The maker told that story plainly in a comment: he built it as the opposite of a subscription transcription app that charged monthly to upload conversations to servers. Nothing leaves your device — transcription, speaker identification, and AI analysis all run on local hardware, no bot joins your call, and no audio goes anywhere. The phone app records away from the desk and syncs to the desktop over your own network, without the internet.

Mia: Pricing is a seven-day free trial with no card, then a one-time payment of 69 euros, with updates included, for Mac, Windows, and Linux. It supports more than ninety languages. And a feature called lenses turns one recording into different outputs — built-in transcription, summary, and action items, plus standup minutes, interview quotes, decision logs, lecture notes, one-on-one notes, and retro boards — and users can write their own prompts and keep them.

Milo: And alongside it is Inventory, a private local index for AI agent and IDE conversations. The principle is the same: indexing the work and conversations the tools produce, without shipping them anywhere. Pairing voice recordings and agent chat history that both stay local on your own machine — that's a theme worth keeping an eye on as more of these tools fight over who owns your data.

Mia: Two new products from the last day actually turn developer idle time into something fun, and they both sit on the same tools a lot of coders already have open. First up, claudemon — a free, fully local Pokémon-style game built for the idle moments inside Claude Code. MIT-licensed, no account, no backend, made by a developer who goes by zamarrowski. The idea: every prompt you throw at Claude is a step through the grass, and every twenty seconds the agent spends working becomes another step forward. When something jumps out at you, the status line tells you, and you fight it in a second terminal tab. It ships the original 151 Pokémon with their real base stats, types, catch rates, and the classic Red and Blue movesets. It runs genuine battles with the type chart, critical hits, status conditions, and PP, plus catching, levelling, evolution, and a Pokédex. It needs Claude Code, Node.js 18 or newer, and a truecolor terminal. The only download is about 1.2 megabytes of sprites, and encounters come one at a time and wander off after roughly half a minute. The community had some fun with the details — one commenter asked whether waiting on it affects the token wait times, and another suggested tying encounters to the exit code instead of elapsed time, so a failed run would actually cost you the catch.

Milo: And the second one, mpai, takes a different angle — open-source, terminal-based multiplayer for Codex and Claude Code. It's MIT-licensed, in public alpha, and launched on Product Hunt on August third as version point four eight, with a quick point four one three update on day one. A named teammate joins a deliberately shared native Codex or Claude Code session from another Mac over Tailscale. They arrive with the real transcript, prompt with their name attached, and the host Mac keeps control. Standalone Codex sessions are view-only by default, managed-mode Codex prompting already works, and Claude support targets the local Claude Code CLI store. Invites bind to the first Tailscale identity that uses them.

Mia: Next up, Hand Wave — a project that turns sign language into text and speech using the camera on Meta smart glasses, and it also runs cross-platform on iOS and web. It was designed and built by Aadi Sanghvi, and the model behind it is described as lightweight, open-source, and trained on Google's FSBoard dataset. It's built to run locally across devices, though that local-run status is currently marked as a work in progress.

Milo: The intent, per the builder, is to make sign-language conversations more accessible using the camera already on Meta glasses and other devices. One commenter praised running the model locally on-device, saying it keeps latency low and the privacy story clean, and welcomed open-source sign-language work actually shipping to glasses instead of staying in demo hell. Another said building on an open-source, local-first model matters because accessibility tools should never depend on a paywall or a connection. But there's a real open question from a third commenter: the FSBoard dataset is a fingerspelling dataset — it covers individual letters, not full signs. So is Hand Wave reading fingerspelled words letter by letter, or does it also handle actual ASL signs and grammar? The product copy says sign language into text and speech, but the evidence doesn't resolve that distinction. Useful signals going forward would be builder clarification on model coverage, updates on the local-run status, and progress on the cross-platform iOS and web version.

Mia: Then there's MacDupl, a Mac utility that clones an already-installed Mac app into a fully isolated instance. Each clone gets its own login, its own data, its own Dock icon, and a color-tinted, badged identity — nine tints available. Maker Sigit says he built it after repeatedly having to log out of a personal Slack to log into a client's, and found browser profiles, separate Mac users, and extra browsers all felt like workarounds.

Milo: Technically, it uses APFS copy-on-write, so cloning a 700-megabyte app costs roughly nothing extra on disk, and clones build in about two seconds. Each clone gets unique bundle identifiers and separate data directories, with a mock keychain so credentials stay isolated from the shared macOS keychain. Container and VM-based isolation were considered but dropped to keep the app fast and light. Each clone is a genuine, independently launchable Mac app, and the original stays untouched. There are limits though — sandboxed apps from the App Store, like WhatsApp and Microsoft Office, can't be cloned because their data lives in shared containers, so MacDupl hides them. Electron apps, Chromium browsers, and the Firefox family are fully supported, and native apps are in beta via an isolated home directory. It's a nineteen-dollar one-time payment with a seven-day full-featured trial. After the trial it's free forever with one clone, and Pro unlocks unlimited clones. It needs macOS 13 or later, works on Apple Silicon and Intel, and has no telemetry or analytics — the only network calls are license checks and app updates. Clones also stay current with the original app's updates.

Mia: And finally, PassiveShorts — an AI faceless video generator for TikTok and YouTube, announced by its maker Vijayabhaskar J. The pitch: you pick a topic, a voice, and a schedule, and the tool writes the script, generates the voiceover, creates the visuals, adds captions, and then auto-posts every short to your connected YouTube or TikTok channels with no manual uploads. It comes with five ready-made niches — scary stories, jokes, life pro tips, today I learned, and you should know — or a custom topic. There are nineteen visual styles including photorealistic, cinematic, anime, Kurzgesagt, and claymation, a library of natural AI voices, six caption presets, and publishing controls that let you post public, private, or as a TikTok draft for review before going live. You can run up to ten series at once from one dashboard, and created videos carry full commercial rights.

Milo: Pricing is self-reported on the site: Basic is nineteen dollars a month for one series, three posts a week, twelve videos a month, and 240 slideshow images. Pro is thirty-nine dollars a month, marked most popular, with once-a-day posting, thirty-one videos a month, and 620 slideshow images, plus a seven-day money-back guarantee. But commenters raised real concerns. The fully automated pipeline may clash with YouTube and TikTok tightening enforcement on mass-produced, low-effort AI content. And because nothing in the loop has a human before publish, a confidently false script could go out under your own channel. It's a product that makes hands-off publishing easy — but that ease is exactly where the risk sits.

Milo: Let's dig into the first launch making noise this week. Ctruh Studio is a new AI-powered, no-code platform, and it landed on Product Hunt this cycle. The pitch, from the company's go-to-market director who posted on behalf of the founder and CEO, is that building immersive 3D and augmented reality experiences shouldn't require a game engine, a specialist team, or weeks of work. Their argument is that online shopping is still stuck in two dimensions, and that hesitation and wrong purchases lead to returns. Their bet is that 3D and AR should be as simple as sharing a link.

Mia: Right, and the whole thing runs right in the browser—no plugins, no downloads. You can generate 3D assets with AI, then build product showcases, virtual storefronts, configurators where shoppers pick colors and materials, and augmented reality try-on that lets someone place a product in their own space or even on themselves. The company says a workflow that used to mean hiring developers and stitching together multiple vendors can collapse into minutes for one person, and they're offering free launch credits. One differentiator they push hard is being all-in-one: a single asset library and a single content management system, with instant load across devices.

Milo: And what are people actually saying about it? One early reaction praised the AI asset generator for handling organic shapes well—specifically, plant models that actually look like plants. Another came from a ceramics shop owner who was impressed, which ties right back to that 3D product showcase use case. So the reception suggests the tool resonates most with smaller makers and sellers who want immersive product display without the tech overhead.

Mia: Next up is gesture.live, a webcam-based instrument for playing live electronic music with hand gestures. The homepage describes it as letting you play chords, bass, drums, and effects just by moving your hands. It's played in landscape, with both hands sharing one 16:9 frame. And the current release is marked as version 1.0.

Milo: The interesting backstory here: a commenter on Product Hunt credits an earlier gesture synth that went viral a few weeks ago on Instagram. That one lets you use both hands to perform chords. gesture.live is described as an exploration of pushing that idea further—keeping the same chord language but compressing it to one hand, so the other hand is free to play melodies, add drum and bass patterns, and control effects.

Mia: So what's the verdict? Honestly, the community raised open questions rather than settled results. One commenter noted the demo video looks great in a well-lit room at a desk, but wondered whether hand tracking holds up under stage lighting or dimmer, more colored light, on a small stage or at a house party. They framed that as the real test of whether it escapes its bedroom-producer use case. Someone else, who isn't a musician, called it amazing but admitted they'd probably only use it to play with their kids. And the most technical comment flagged a timing concern: webcam hand tracking runs around thirty frames per second on a good day, which means the audio thread hears your input with noticeable jitter before it lands. So whether the note timing gets cleaned up is the decisive open question.

Milo: So it's a promising instrument that people are charmed by, but the live-performance question is still unanswered. We'll see whether later builds tighten that timing.

Mia: That's our roundup of what's launching on Product Hunt right now. From Qwen's biggest model yet, to AI agents that run on your phone or in the cloud, to a Mac utility that clones apps and a webcam that becomes a music instrument.

Milo: Lots to try if any of those caught your eye. Thanks so much for listening, and we'll be back with more launches soon. Until then, happy hunting!