0920 | Tools With a Point of View

Mia: Welcome back to the show. I'm Mia, and across from me, as always, is Milo.

Milo: Hey everyone. Today we're doing our daily Product Hunt briefing, and I'll be honest, the lineup this time is a nice mix. There's meeting software and data modeling, an app blocker that makes you pay yourself, open-source coding models, a couple of meditation apps with very different philosophies, and a bunch of small utilities for saving things you'll lose otherwise.

Mia: And if there's one thread running through all of it, it's trust. Who has the agreed record of what happened? Who decides what counts as your real Instagram usage or your real task list? Everything we're covering today is some tool saying, "here's the version of the truth you can actually rely on." So let's start at the top, with the meeting problem, because it's the most relatable story here.

Milo: Yeah, so the product is VoiceCap, and the maker story is genuinely good. A guy named Rokas runs an AI software agency in Lithuania, and the pattern in his client meetings was: a call in Lithuanian, three people taking their own notes, and two weeks later nobody agrees on what was actually decided.

Mia: Which is a universal experience, I think, regardless of language. But the language part is what makes this one interesting. He tried every AI notetaker, and they were all built English-first. His summaries came back — his words — like a bad translation of a meeting he hadn't had.

Milo: That's a great line. And so VoiceCap transcribes in over a hundred languages, and critically, it writes the summary, the action items, and the decisions in that same language. Not translated into English first. He says the majority of smaller languages are covered, which matters a lot for places like Lithuania where mainstream tools treat your language as an afterthought.

Mia: So who is this for? Teams and agencies that work in non-English languages and need a shared record. And how it works: you capture meetings three ways — you record in the room from your phone or laptop with native iOS and Android apps, you can send a bot into Zoom, Meet, Teams, or Webex, and it can auto-record from your calendar, or you upload an audio or video file.

Milo: And the part he actually wants feedback on — and this tells you what he thinks is differentiated — is the decision log. His argument is that every notetaker does summaries and action items, but almost nobody treats decisions as separate records. So VoiceCap pulls out every decision, links it to the exact second it was said, to the meeting and context, and keeps a searchable log across all your meetings.

Mia: So if you've ever had to reconstruct who agreed to what three months after a meeting — which, everyone listening has — that's the feature. There's a demo on their site, a website redesign kick-off meeting, where the decision "phase-two budget approved by end of September" is a record with context: Tomas owns sign-off, Rūta sends the updated proposal.

Milo: And it connects to Claude and ChatGPT over MCP, so you can ask something like "what did we promise this client?" without opening the app. In their example, it reads three meetings and comes back with two commitments still open. On privacy — and this is a real pitch point for an EU product — they're an EU company, recordings stored in Frankfurt, never used to train models, encrypted in transit and at rest.

Mia: Pricing is concrete: free plan with 300 minutes and no card, Pro at 29 euros per capture seat per month for a thousand minutes, Business at 49 for unlimited recording. Viewers are free on every plan, unlimited, which is a sensible structure — you pay for people who record, not people who read.

Milo: Now, treating maker descriptions as claims: the "industry-leading accuracy across European languages" line is their claim, we have no independent verification. And the open question is adoption of the decision log itself. A decision log only works if decisions actually get logged and reviewed — it's a discipline, not just a feature. He's explicitly asking the community how they handle this today, which suggests he knows it's the risky bet.

Mia: Right. And it's a useful framing for the next product, because Basedash Models attacks essentially the same problem — the "one trusted definition" problem — but from the data side instead of the meetings side.

Milo: So here's the setup from Max, the founder and CEO of Basedash. Every company has core concepts — customers, orders, active accounts — and every team quietly redefines them in their own SQL. Someone asks for active enterprise customers, and depending on which Slack thread the question came from, you get three different counts.

Mia: And Basedash Models is their answer: a semantic workspace of reusable, governed SQL. You define "customers" once, with measures — like active customers, MRR, LTV — segments, like enterprise or trialing, relationships to other models, synonyms so people can ask for it in plain language, and usage guidance. Then you query it like an actual table: select star from models dot customers.

Milo: And the AI assistant reads that same context when it writes SQL, so it doesn't guess which filter you meant — it reaches for the measures and segments the team already agreed on. Their own dogfood story: they ran their Northstar metrics on it, and questions that used to spawn three competing counts now hit one model.

Mia: A couple of details worth noting for anyone using Basedash already. Models live as a top-level workspace under Data. Measures and segments are first-class parts of the model, not comments buried in SQL. Legacy definitions — their old feature — keep working and migrate automatically. And admins can turn Models off per organization if a workspace isn't ready. So it's live, and it's an evolution of something they already had.

Milo: What's the comparison here? The core argument, on both VoiceCap and Basedash, is that the value isn't capture — anyone can transcribe a meeting or run a query. The value is agreement. Shared record, shared definition. The difference is the artifact: one is human decisions, the other is business metrics.

Mia: And the open question for Basedash is similar in kind: does that shared definition actually hold across teams, and does the assistant reliably use it? They claim it does; whether the governance stays honest as more people and questions pile in is something only usage will show. Also worth saying: Basedash is an AI-native BI platform more broadly — 750-plus data sources, dashboards, MCP server — so Models sits in a bigger product.

Mia: They offered the Product Hunt community an extra week on the trial this week.

Milo: Okay. So that's trust in what was decided and trust in what the data means. Next, let's shift to a very personal kind of trust: trusting yourself not to open the app you always open. This is Squirrel, an iPhone app by a solo developer named Brian.

Mia: And the origin story is very relatable. He was talking with friends after a night out over Thanksgiving break, and two things came up and everyone agreed on both. One: social media has hijacked our brains — you open Instagram on a knee-jerk reaction, then 30 minutes go by and you have no idea how you got there. Two: everyone felt behind on investing, knew they should be doing more of it, but weren't.

Milo: So he connected the two problems. Squirrel is an app blocker where you choose the apps you open too much and set your own toll. When you tap Instagram, it's actually blocked — a hard block, not a screen you can wave away. If you really want in, you open Squirrel, pick how long you want it unlocked, and set aside a few dollars for yourself.

Mia: And this is the part people always ask about, so let's be clear: no money moves through the app. Squirrel never charges you, never connects to your card, bank, or brokerage. It just records what you set aside, like a swear jar. When you're ready, you invest it yourself in your own brokerage, tell Squirrel what you bought, and it tracks your holdings with live prices.

Milo: And his personal numbers — these are his claims about his own usage, but they're specific. He had a 45-minute daily Instagram limit for months and still averaged over 10 hours a week. First week with Squirrel: 1 hour 45 minutes. Week after: 56 minutes. His own diagnosis of why it works is interesting: it's not the money, it's the few seconds where you actually have to decide. Nothing was ever a conscious choice to open Instagram in the first place.

Mia: There's a comment in the discussion that nails the competitor problem. Someone asked about the "just this once" override button — the thing that kills every screen-time app — and whether Squirrel's block is genuinely hard once it's on. That's exactly the right question, because Brian's own story proves the soft version fails: the built-in limit had "ignore limit for today" one tap away, and he spent 12 hours and 48 minutes on Instagram that same week.

Milo: Which, honestly, doubles as the strongest source-supported evidence here — a documented before-and-after from the maker himself, with the caveat that it's one person, n of one, and he says as much. Another commenter suggested flipping the currency entirely — if you spend an hour scrolling instead of creating, you owe your craft an hour — and someone else mentioned digging out a Gameboy Advance to substitute for scrolling. So the discussion is genuinely about the mechanism, not just applause.

Mia: On monetization, it's clean: blocking, setting money aside, and savings history are free forever. The optional Portfolio upgrade — holdings, growth charts, projections — is 19.99 once, or 8.99 a year. No account, no login, no ads, no tracking, data stays on your phone and iCloud. And it's iPhone only. Not a financial advisor, you make every investment yourself.

Milo: Now here's the thread I want to pull, because it leads to the next story. Squirrel works by making the default harder. The next product makes the default cheaper — and it's asking you to trade something for it. This is Bolt Forge.

Mia: Bolt Forge is a new agent inside Bolt.new, the AI app builder, and it runs entirely on open-source models. In the agent picker it sits next to the Standard and Max agents, and it launches on GLM 5.3 Flash with GLM 5.3 alongside it, plus Kimi K3 and DeepSeek v4 Pro as experimental picks.

Milo: The headline: every individual Pro plan gets up to 50X more usage on Forge at no extra cost, through October 14, 2026. It's a research preview launched September 14. There's one monthly usage bar, no daily limits, and a hard stop at 100 percent — when you hit it, it auto-switches to Standard, no surprise pause mid-project, no overage charges.

Mia: And here's the trade, which is the whole story: to get that allocation, you opt in. Your build sessions — prompts, code, fix traces — get de-identified, with secrets and personal information stripped, and they go into datasets that Bolt licenses to AI developers, starting with Arcee AI, a US open-model lab. Switch back to Standard or Max and Forge stops collecting new sessions.

Milo: So the framing from Bolt is that your consented sessions are the payment. The rationale they give is that inference costs have collapsed — they cite a 280x drop in 18 months per the Stanford HAI AI Index — and that open models improve when they see how real software gets built, which is the one thing you can't scrape.

Mia: Now, the benchmark claim, and this deserves scrutiny. Forge's models scored 92.2 versus 101 for Bolt's top paid model on the Bolt Build Index — their internal benchmark — so about 91 percent of the top score. And a commenter pushed back on exactly this: internal benchmarks tend to be picked to flatter the new thing. Is there a public eval outsiders can rerun, or is 91 percent something we take on faith? That's a fair criticism, and as of what's in the source, there's no public rerunnable eval.

Milo: Right. Treat it as a claim. There are also honest limitations from Bolt themselves: these models are experimental inside Bolt, duplicate your project before switching a serious build into Forge, keep production work in Standard or Max, Kimi and DeepSeek burn through usage faster than the GLM pair so start on the default, and Forge can't take PDF uploads yet.

Mia: Who's it for? Builders who burn through credits in the brainstorm and draft phase — the phase where usage caps punish exactly the wandering you need to do — and want room to experiment without touching production usage. And one commenter commended it as, as far as they know, the first vibe-coding solution to really go all in on open-source models.

Milo: So, common thread with Squirrel: both are about making the default easier or harder in a deliberate direction. Squirrel makes the bad default expensive. Bolt makes the cheap default expensive in a different currency — your data. And both put a decision point in the middle.

Mia: Let's move to the kitchen and the cushion, because there's a trio of products here that are all pushing back on the same bloated-wellness — or bloated-cooking — market. First, Mise, from the Robot Recipes team.

Milo: The problem is deceptively simple. Recipes help you cook one dish. But meals have multiple dishes, and getting everything ready at the same time is genuinely hard. Most people — the maker's words — have just expected that things won't be ready together.

Mia: Mise schedules every dish backwards from the minute you want to eat. You pick your dishes — there are thousands of recipes on Robot Recipes, or you can generate new ones on demand — choose your meal time and servings, and you get one start time, one running order, a merged shopping list, and a cooking mode.

Milo: The clever part is step three, if we go through how it works. The robots read each recipe's steps to work out how long each dish takes and which parts need your hands. Then dishes are nudged earlier so two hands-on steps never collide. And cooking mode shows only what to do right now, keeps the screen awake, and beeps when a step is due. It works offline once the page loads.

Mia: Availability is generous: free, no login, no app, no ads in the way. And the maker is candid about limitations — he says it's not always 100 percent perfect, but better than guessing or paper notes. That's a refreshingly honest claim from the person selling it.

Milo: Now the wellness side of this trio, and it's two very different answers to the same complaint: meditation apps are bloated, push guided programs, streaks, social features, and subscriptions. First, Lull, by Evan — three years of nights and weekends.

Mia: Lull doesn't play recordings. You talk for a minute about what's actually going on — out loud or typed — and it writes a meditation for that exact moment, read in one of eleven voices. No two sessions are the same, because no two days are.

Milo: And Evan is big on what he calls honesty. Connect an Oura ring and Lull reads your recovery before writing a word, so a rough night changes the tone. Wear an Apple Watch and the session runs on your wrist with guided breathing and haptics, heart rate recorded through it — and afterward it shows how far you settled below your resting baseline, or nothing, if it didn't get a clean reading. He explicitly refused to invent a calm score.

Mia: He also lists what he deliberately left out: no streak pressure, no guilt notifications, no gold stars. Real Apple integration — HealthKit in and out, so heart rate, HRV, and sleep come in, Mindful Minutes go out — a standalone Watch app, Live Activities, widgets. There's even a whisper voice for 3 a.m. It's free to download with a seven-day trial on the subscription, iPhone and Apple Watch only.

Milo: One community comment worth flagging: someone asked about hallucinated responses. That's the right question for a generative meditation app — the source doesn't include his answer, so that's an open question. And the philosophical contrast brings us to Mantra Timer.

Mia: Mantra Timer is the opposite pole. Built by Brent, an indie developer who's practiced Transcendental Meditation daily for over 3,700 consecutive days — the site says 3,710 now. His complaint: the wellness market is saturated with bloated apps. The official TM options required intrusive personal background data, everything else pushed guided programs and subscriptions just to sit in silence.

Milo: So his app is deliberately minimal: it opens directly to the timer, no menus, silent notifications, no tracking, no social features, built specifically for the 20-minute mantra practice. He argues mantra meditation isn't mindfulness or breath-awareness — it requires absolute simplicity, a timer, a mantra, and silence.

Mia: And there's a business-model story here too. He originally planned a standard annual subscription, but his studio has an anti-rental philosophy, and keeping a subscription felt hypocritical — so he completely changed direction. The core app is free with timers, warm-up and cool-down phases, session reflections, 12 months of local history, basic stats. Mantra+ is a one-off 9.99 payment — 5.

Mia: 99 with the launch code LAUNCH2026 — unlocking unlimited custom timers, advanced stats, milestone badges, up to eight daily reminders, iCloud sync, Family Sharing, and Apple Watch support. Buy once, own forever.

Milo: He's building at what he calls a "human pace" while managing a chronic illness, ME, which forced hyper-intentionality. And the community questions are sharp in a good way: one person asked what he decided not to include, tying to the mindfulness theme. Another asked — with zero data collection, how does he even know if people use it day to day, or has he let go of tracking on purpose? That's a genuine open question; no answer in the source.

Mia: So the contrast within this trio: Lull is generative and connected, Mantra is silent and self-tracked — and both refuse streaks and guilt. And Mise, in a weird way, is the same philosophy applied to dinner: remove the stress, give you exactly what you need when you need it.

Milo: Which brings us to the grab-bag of small utilities, and this is the "I know I saved that somewhere" problem. Three products, all small, privacy-leaning. First, BiBimba.

Mia: BiBimba is a Mac clipboard manager by Rihito, a solo developer in Japan. The origin: he takes a lot of screenshots — receipts, error dialogs, slides — and when he needs one number from one of them a week later, he can't search for it. So BiBimba treats screenshots like clipboard entries: it runs OCR on them right away, and everything you copied or captured lands in one search box.

Milo: So you type "invoice" and you get the text you copied, the screenshot with that word in it, and any snippet. It's keyboard-first — control-shift-C opens history, return pastes back into the app you were in. The AI features — translate, summarize, rewrite, your own saved instructions — use Apple's on-device models, so nothing goes to a server. Everything stays on the Mac, you control retention, no analytics.

Mia: Pricing: 9 dollars, one time, for three Macs; buy again with the same email for three more. No subscription. Requirements are a real limitation: Apple silicon and macOS 26 or later, because the OCR and on-device models come from the OS itself, not bundled. Interface in ten languages.

Milo: And the community questions show what people care about. One user makes help guides in five languages and asked whether the OCR handles Italian, German, and French accents, and whether OCR runs at capture time or only on search — because they need the text hours later when the screenshot is buried. Another heavy screenshotter asked whether OCR keeps up without lag. Those are exactly the right questions; the source doesn't include answers.

Mia: Next, Lumiko, a Chrome screen recorder. The maker lost an hour to keyframes after every five-minute recording — the slick zoom-and-pan in tutorial videos usually comes from someone manually placing markers over every click. Lumiko watches your cursor and clicks while you record, then generates the zoom and pan automatically when you stop. No keyframes.

Milo: Plus a full browser-based timeline editor — trim, split, reorder — Smart Blur to hide API keys, emails, passwords, revenue dashboards with one drag, a draggable webcam bubble, custom backgrounds including Unsplash images or your own brand assets, click effects, and multiple aspect ratios: 16:9 for YouTube, 9:16 for Shorts and TikTok, 1:1 for LinkedIn.

Mia: The privacy and performance story is strong: everything renders locally in your browser, no account, no upload queue, no server. Free tier is genuinely generous — unlimited recording, auto-zoom, Smart Blur, webcam, exports up to 1440p, with a "Made with Lumiko" watermark. Pro is 9 dollars — one time, lifetime, three devices — removing the watermark and unlocking 4K. Note the site has a banner showing Pro at 29 crossed down to 9, so the 9 is the current launch price.

Milo: One community comment is a real quality probe: cursor position is right most of the time, but wrong on takes where you move the mouse while talking — can you override a zoom it placed? That's the key open question for any auto-editing tool: what's the escape hatch when the automation guesses wrong? The source doesn't show an answer. And there's a practical bug report too: someone bought Pro and couldn't find where to paste the license code.

Mia: Third in this trio, Punch. The origin story: one of the co-founders was scrolling endlessly through a group chat for an out-of-state friend's address to send a birthday gift — an annual tradition, he says. He gathered the other co-founders and it turned out everyone had the same issue.

Milo: Punch is a place to store bits of info for yourself or a group — the Airbnb address on vacation, a friend's gate code, an itinerary, a grocery list, important photos. It's described as a cross between shared notes and a visual board. Their positioning versus alternatives: Notes app fills up, gets clunky, you can't share individual pieces, and it breaks across OSes. Drive and Dropbox have storage and sharing but are geared toward storage, not quick lookup.

Mia: It's cross-platform — App Store and Google Play — and the build story is fun: they flip-flopped between React Native and Flutter, had big design disagreements, and the real killer was App Store review — three weeks with multiple rejections, two of which claimed missing things they actually had, while the Play store approved in 24 hours. Anyone who's shipped an iOS app just nodded.

Milo: So that's the utilities trio: BiBimba for your clipboard and screenshots, Lumiko for recordings, Punch for group-chat essentials. All small, all cheap or free, all leaning local or private. Now let's zoom out to the infrastructure layer, because there's a fascinating open-standards story here.

Mia: This is Ruby UTCP, by Kamil Mościszko. UTCP is the Universal Tool Calling Protocol — an open standard, an alternative to MCP, for letting AI agents call tools. The core idea: instead of proxying every call through a new server, after discovery the agent speaks directly to the tool's native endpoint — HTTP, gRPC, WebSocket, CLI — eliminating what they call the "wrapper tax," reducing latency, and keeping your existing auth, billing, and security in place.

Milo: Their stated principles: no wrapper tax, no security tax — the AI gets the same security as a human calling the API — scalable, and simple. Ruby UTCP brings version 1.1 to Ruby with 12 transports including HTTP, CLI, GraphQL, and MCP itself, plus streaming, authentication, OpenAPI discovery, and CodeMode for orchestrating multi-tool workflows with compact Ruby code. It's open source, MIT licensed.

Mia: And the ecosystem context from the broader UTCP project: there's a UTCP-MCP bridge so MCP users can connect to over 230 tools through one MCP server, SDKs in Python, TypeScript, and Go, and a list of agent frameworks that have integrated it — LangChain adapters, Pydantic AI, that kind of thing.

Milo: Now, the tie-back to Bolt Forge is worth saying out loud. Both are about the infrastructure-versus-economics tension in AI tooling. UTCP asks: why pay a proxy tax on every tool call? Bolt asks: what's the fair price for cheaper compute? And both are opt-in trades — UTCP trades the convenience of a single proxy for directness, Forge trades data for usage.

Mia: And then there's Steam Frame, which is Valve doing open-platform hardware. It's a wireless VR headset that runs SteamOS — the Steam Deck thinking underneath. It can play supported games standalone, and when you want the full PC library, it streams from your gaming PC, Deck, or Machine over an included 6GHz adapter with dual radios for stable streaming.

Milo: The detail the hunters loved: the controllers work as VR wands and a full gamepad, so you don't have to choose between VR input and a real controller. One commenter called that the smart part and asked about weight for long sessions versus Quest. Another noted the elephant in the room — there's a normal Linux PC underneath the VR layer, and they suspected they'd spend as much time in desktop mode as playing VR games. Very Valve, indeed.

Mia: The criticism from the community: pricing. One commenter called it quite expensive next to a Meta Quest 3, which is almost a third of the price. That's a fair framing — you're paying for the PC-library streaming and SteamOS, not just standalone VR. No pricing figure in the source, so we won't invent one.

Milo: Which brings us to our last product, and it loops the whole day back to a question we've been circling: what happens when AI acts for you, and where's the moment you get to check it? Doneit 3.2.

Mia: Doneit is a task app, and this update reimagines Doneit Assist. The headline feature: multimodal planning. You attach a photo — a handwritten list, a whiteboard, a document — and it turns that into tasks. It can also plan around what's visible in a photo.

Milo: Beyond that, Doneit Assist got more planning options — add extra details for accuracy, suggest reminders and attributes, break tasks into subtasks. It can now summarize a task list, highlight key tasks to focus on, and show upcoming tasks.

Milo: And on the platform side: task lists and tasks are accessible to the new Siri AI for onscreen awareness, you can add tasks through Siri AI, there's a richer Shortcuts configuration, and the list widget supports the new extra-large portrait option on latest iOS, iPadOS, macOS, and watchOS.

Mia: And the community asked the exact question this update needed someone to ask: turning a whiteboard photo into tasks is a genuinely useful demo, but handwriting recognition is where these things fall apart — messy shorthand, arrows, crossed-out items. What happens when Assist misreads something? Does it show you what it "saw" before creating tasks, or do you find out only after a wrong task shows up? No answer in the source, so it's an open question.

Milo: And that question connects everything we covered today, honestly. Squirrel puts a decision between you and the app. Bolt Forge puts a consent screen before the data sharing. VoiceCap links every decision to the second it was said so you can check it later. Basedash puts the agreed definition where the AI can read it. Doneit is being asked to add the same thing: verification before action.

Mia: So, quick recap for anyone who just joined: the briefing covered VoiceCap for multilingual meeting records with a decision log, Basedash Models for governed SQL definitions, Squirrel for habit tolls that pay you back, Bolt Forge's open-model trade of usage for session data, Mise for synchronized cooking, Lull and Mantra Timer's two answers to bloated meditation apps, BiBimba, Lumiko, and Punch for finding what you saved, Ruby UTCP and Steam Frame on the open-standards side, and Doneit 3.

Mia: 2 turning paper into tasks.

Milo: And a reminder on how to hear all of this: maker numbers are claims, benchmarks are self-selected until someone independent can rerun them, and community comments are individual experiences and questions, not verdicts. The strongest signal today, in my view, was the pair of founders — Rokas and Max — both betting that agreement, not capture, is the actual product.

Mia: That's the briefing. Thanks for listening — we'll be back tomorrow with the next day's launches.

Milo: See you then.