0810 | Argos AI Agent, Soup CLI Fine-Tuning, Grok Imagine 2.0 & DuckDisk

Mia: Hey everyone, welcome back to ProductHunt Daily. I'm Mia, and I'm here with my co-host Milo, for a rundown of what's shaking up the Product Hunt universe today.

Milo: And we've got a packed lineup for you. We're talking about an AI agent that actually acts on your behalf, a way to fine-tune a large language model on a modest laptop, and a seriously upgraded image generator.

Mia: Plus some clever macOS tools to keep your workspace humming, an open-source take on AI team members, a gamified app aimed at a really specific habit, and a couple of others worth your time.

Milo: So grab a coffee and settle in, because we are diving into all of it right here on ProductHunt Daily.

Mia: Let's start with Argos, which is described as the only launch on today's list that actually acts on your behalf instead of just giving you answers. The pitch is that most AI tools will hand you a great answer and then leave the hard work — the copying between tabs, the form-filling, the "now go run this yourself" — to you. Argos is meant to close that gap right inside your browser, using your own logged-in accounts to click, type, fill in forms, and finish real tasks. You can ask it in the sidebar or send it a message on Telegram or WhatsApp, and it runs live while you watch, or in the background, and then hands you the finished result. It's claimed to be native in Gmail, Docs, and Sheets, and to connect to GitHub, Slack, Notion, and more — and it's free to start.

Milo: But these are self-reported claims, not verified results, and there are a couple of things worth knowing before you jump in. On trust — which the maker calls the hardest part — the stated design is that everything runs locally, data never leaves your device, and anything destructive stops to ask you first. However, the team also disclosed that Argos is actually a modified, rebranded version of a past project — same team, same underlying product, new name and refreshed direction. That old name, Lyto, still shows up in a few places like the Chrome Web Store listing while the rebrand finishes rolling out.

Mia: And the most substantive question from the community is about verification. One commenter flagged the failure mode they keep hitting with browser agents: a click that reports success when the action never actually landed, often because a late-loading element shifted the page underneath it. Their question to the maker is direct — how does Argos verify that a click or form fill actually landed, and not just assume it worked because the tool says so? That's a fair thing to push on, because for an agent that finishes tasks on your behalf, the trust hinge isn't just whether the data stays local — it's whether you can believe the task was actually done. There's no answer from the team in the supplied discussion, so that one stays open.

Milo: Next up is Soup CLI, and this one has a delightfully specific origin story. The maker built it because they own a laptop with just 4 gigabytes of VRAM, and they wanted to fine-tune models that don't fit on it. So the tool fine-tunes an 8-billion-parameter model on that 4-gigabyte laptop GPU. The trick is that during training, the base model is frozen, so it sits in system memory and gets streamed into the GPU one decoder layer at a time. That means the peak graphics memory needed is just one layer instead of the whole model. Measured on an RTX 3050 laptop with 4 gigabytes of memory, a Llama 3.1 8-billion model trained at about 120 tokens per second in a peak of roughly 3.3 gigabytes.

Mia: And it's all wrapped up in one YAML file and one command, covering several training methods — SFT, DPO, GRPO, and KTO — plus evaluation, gating, and export. It's open source under the Apache 2.0 license. But here's what makes this one interesting beyond the memory trick: the maker says the hard part wasn't speed, it was proving correctness. Streaming can fail silently. If the automatic differentiation path gets cut, the loss curve still goes down, because the upper layers keep learning — so you'd never know anything was wrong.

Milo: So every release compares a streamed run against a fully resident run and requires the logits to match exactly. And that protocol actually paid off. Last week, someone lent the maker eight H100 GPUs for three days, and the check found a real bug in their own released code — above a certain layer size, the gradients are silently wrong while the loss curve looks perfectly healthy. The maker published it with a reproducible example. One commenter said that correctness protocol is what won them over, and called publishing the bug rarer than the optimization itself. There is one open question in the discussion though: a commenter asked whether the correctness check holds up against a base model that's already been quantized to 4-bit precision, or whether it's only validated against full precision so far. No answer appears in what we were given.

Mia: Let's turn to images. Grok Imagine 2.0 is a next-generation image generator launching today, and the big headline feature is something the launch calls segmentation editing — the ability to select and edit just one region of an image rather than regenerating the whole thing. The maker claims several upgrades in this version: significantly better instruction following, sharper text rendering and typography, complex coherent layouts, and that precise region editing, controlled through a Magic Wand tool.

Milo: Now, a caveat before we go further. The launch gives no pricing, no model specifications, and no API plans, and the linked product website returned only a "Just a moment..." challenge page when it was captured, so its content couldn't be verified. The community reaction was mixed but substantive. On the positive side, one commenter called the upgrade major and said they're excited to see what creators build with it. Another praised the sharper, finer rendering and said Grok has made real quality gains in its last few updates. But a different commenter said the model isn't as powerful or close to hyper-realism as the others, though still a good addition to the X suite.

Mia: And there's a key open question about that Magic Wand — and it's a good one if you work with product or ad images. One commenter explains the usual failure mode: you edit one region and everything else subtly shifts — text warps, the layout drifts. So they asked whether version 2.0 keeps the untouched regions pixel-stable, or whether it just performs a full regenerate guided by the mask. And separately, whether segmentation editing is exposed in the API at all, or if it's app-only for now. The supplied evidence doesn't include an answer to either question. Also worth noting, one commenter made a broader point — there are so many good AI image models now that choosing just one is hard, and they're hoping AI video tools improve next.

Milo: Finally, let's look at DuckDisk, a free, open-source storage analyzer for macOS. The philosophy here is table-first: instead of the usual visual daisy-map of folders, it keeps directory size, allocated space, percentage of parent, item counts, and file-type totals visible while you drill down. It scans local disks, OneDrive, Google Drive, and SSH paths — all from one app. And importantly, cloud scans are metadata-only: your file contents are never downloaded.

Mia: The cleanup is designed to be safe and staged. Deletions on OneDrive and Google Drive go to the provider's Trash, so they're recoverable. Permanent deletions on local disks or over SSH need explicit confirmation. And the privacy story is spelled out carefully: full-disk scans require granting Full Disk Access, tokens are stored in the macOS Keychain, and the app is licensed under AGPL-3.0, with the code open on GitHub. There's a recent release detail too — version 0.5.9 added Apple notarization for the direct download and a sandboxed Mac App Store edition. The Store build covers local disks, OneDrive, and SSH, while the direct download adds Google Drive and uses your system SSH configuration.

Milo: One privacy nuance worth flagging: Google Drive scans request full drive access, not because the app wants to read your content, but because Google doesn't offer narrower metadata-only permissions that would still let the app move selected items to your Trash. The app says it uses that access only to read metadata and perform user-initiated trash actions. That's a clear and honest tradeoff to state openly. And in the discussion, the maker explained he built it because existing tools didn't fit what he needed — a scanner that keeps the numbers visible while you drill in, the way WizTree does on Windows. It's a sensible pitch for anyone whose drive keeps filling up and who wants the numbers, not the pretty map.

Mia: Let's start with Workflo, a macOS tool that wants to kill the endless fiddling at the start of a call or a work session. It's a menu-bar app built around the idea that your workspace shouldn't have to be rebuilt by hand every time your context changes.

Milo: And the pitch here is interesting, because it's not another app that makes arranging windows faster. The maker, Chirag, frames it as the difference between tools that wait to be asked and one that just acts. Workflo watches the clock, your calendar, and your display setup, and stages the right windows before a meeting starts or a focus block begins. So on a docked desk with dual monitors versus sitting on your laptop, it restores the layout you want for that setup automatically.

Mia: The core idea is something they call Desks — Workflo learns a layout for each display setup you use. On top of that you have Scenes, named presets per Desk, like writing, review, or monitoring. A Scene can open apps that aren't running, hide everything else, or close what you don't need. There's a canvas for dragging and resizing with dual-monitor support, and you can trigger a Scene at a set time, or in the minutes before a calendar event that has a video link.

Milo: And there are small touches aimed at the everyday annoyance. A hotkey stashes every window so your screen's clean, then brings back the Desk exactly as it was. Newly launched apps land at remembered positions, and you can set per-app rules to ignore an app entirely or always restore its size. The alternatives listed are familiar names — Rectangle, Magnet, and Moom — so you know exactly the category it's competing in.

Mia: Privacy is a real part of the sell. Workflo runs on Accessibility permission alone and never asks for Screen Recording, which the maker says means it structurally can't read your window content. No account, no cloud — your Desks, Scenes, and rules live as plain JSON in a folder you can open, back up, or delete yourself.

Milo: Next up is AgentConnect, an open-source platform for running AI agents not as personal assistants in a terminal, but as shared team members that everyone can see and work with.

Mia: The backstory matters here. The maker says they started like most teams — everyone running Claude Code or Codex in their own terminal. Agents became genuinely useful teammates, but no one else could see a session, take it over, or reuse its context. So they wrote their own glue, and found existing tools wanting: personal tools that had no team model, team tools that forced a migration into a new chat app, and closed-source tools locked to a single provider.

Milo: AgentConnect answers that by being provider-neutral and fully self-hosted. It connects Claude Code, Codex, Grok Build, DeepSeek, Pi, or any ACP-compatible runtime, and works inside Slack, Telegram, Discord, and GitHub. Each agent gets a role plus configurable model, workspace, memory, tools, skills, and permissions, and can be kicked off from a conversation, a pull request, an issue, a webhook, or a schedule.

Mia: The bigger shift is that agents can call one another, while teammates follow only the work they're allowed to see from a single console. The self-hosting claim is central — agents run on your machines, so the maker never sees your code or messages.

Milo: Two community reactions stood out. One commenter said that if the console genuinely shows what each agent is allowed to see, it solves a real headache around visibility, approvals, and who approved what. Another welcomed the runtime-agnostic approach, calling reliance on one model provider risky and liking the ability to mix DeepSeek and Claude Code. Several questions around the details were still unanswered in the discussion.

Mia: SoloUno is a very different launch — a gamified self-help app for body-focused repetitive behaviors, specifically hair pulling, skin picking, and nail biting. And the design philosophy is deliberately counterintuitive.

Milo: Maker Omer Bialer has dealt with trichotillomania for many years, and he built SoloUno around small daily wins instead of an all-or-nothing goal. The logic is that for many people, simply deciding to just stop isn't enough, and apps built around never doing the behavior again make every slip feel like failure. So the app is built around habit-free challenges and streaks, quick habit logging, trigger analysis, urge-acceptance sessions, and a daily wrap-up. It's positioned as a self-help tool that can also be used for therapy homework, with a two-week free trial.

Mia: The inspiration comes from established methods — Habit Reversal Training, CBT, and ACT — but softened by that streak-based approach. The site reports more than six thousand users, over a hundred thousand habit-free hours, and twenty-five thousand habit reports. Worth noting those are self-reported site figures, not independently verified.

Milo: The community comments are individual and attributed. One person says it helped avoid nail biting for months. Another says it helped almost get rid of decades of nail biting. Someone's been using it since the early versions and no longer needs it after breaking a thirty-plus-year habit. Another says at thirty-six it didn't make the habit vanish but made it controllable and easier to minimize or stop. One credits it with stopping plucked beard hairs. And one person names the most valuable part — not feeling defeated or like a failure after pulling hair.

Milo: Finally, there's Macrobite, a macro-tracking app that's attacking what it calls the fastest way to actually get macros right — accurate calorie, protein, carb, and fat tracking without the friction of traditional apps.

Mia: The problem it targets, echoed by a commenter, is the split personality of existing trackers. Photo logging gets close but leaves wrong estimates to be fixed through clunky menus. Manual entry is accurate but slow. Neither holds up for someone chasing a daily protein target. Macrobite's answer: snap a photo of a meal and get an instant breakdown of calories, protein, carbs, and fat, then fix any estimate in seconds without scrolling through huge food databases. You can also log by voice, including through Siri, or scan a barcode.

Milo: There are quality-of-life features on top. Re-used meals can be saved for instant access, and your daily protein, calorie, carb, and fat intake is glanceable from an iPhone widget or an Apple Watch without opening the app. Developer Alex says the goal was to make macro tracking fast without sacrificing accuracy, and that care went into making every meal easy to log, review, and correct.

Mia: But some features are still in the works, not shipped — personalized meal and product recommendations based on remaining macros, with the ability to order through Instacart. Alex calls this launch only the beginning. The discussion also raised open questions and limitations about the approach, so it's a launch with clear ambition and a clear gap between what's here today and what's promised next.

Mia: ScrapeOps has launched a benchmarking tool called Proxy Tester, and the premise is that there is no single best proxy provider. Ian, who is behind ScrapeOps, makes the point that a provider can perform great on one target site and poorly on another, so teams end up buying credits, running manual tests, comparing logs, and losing days to the decision.

Milo: So instead of relying on generic rankings or vendor claims, you give it your own target URL, and it benchmarks twenty-plus provider configurations — residential, datacenter, mobile, proxy API, and unblocker setups — specifically against that exact site.

Mia: Right. The website describes a three-step process that's free and needs no credit card. You enter the target URL, the tester sends requests through all those providers using the configurations they support — including residential routing, JavaScript rendering, and anti-bot options — and then the ranked report arrives by email.

Milo: And what makes the ranking trustworthy is what it actually measures. Each report checks that a page counted as successful is genuinely a valid response and not a ban page, and it tracks how those false positives distort success rates and pricing.

Mia: Beyond that, it reports the real success rate of valid responses — not bans, CAPTCHAs, or errors — plus average success latency, the bandwidth and API credits actually charged, and the cheapest plan that would cover your volume.

Milo: And the headline number is the cost to scrape a million pages, the CPM, so you get a side-by-side comparison before you commit to anything.

Mia: That cost-per-million figure is really the point — it turns a multi-day evaluation into a free report on your exact workload, so you can pick a provider based on your target, not on marketing.

Mia: That wraps up this roundup of Product Hunt launches — from Argos, the browser-based AI agent that works on your behalf, to Soup CLI for fine-tuning models on modest hardware, and the new Grok Imagine 2.0 image generator. We also covered DuckDisk and Workflo for macOS, plus AgentConnect for running AI agents as shared team members.'

Milo: And don't forget SoloUno for body-focused repetitive behaviors, Macrobite for faster macro tracking, and Proxy Tester by ScrapeOps for benchmarking proxies at scale.

Mia: A pretty varied week across AI agents, dev tools, and everyday utilities. Thanks for listening — we'll be back with another launch roundup soon. Until then, take care.