0921 | Trust, Ownership, and the Machines That Run the Work

Mia: Hey, welcome back to the briefing. I'm Mia.

Milo: And I'm Milo. Today's a fun one because there's a thread running through almost everything we saw in the last day: software that makes promises and keeps them — in writing, sometimes literally.

Mia: Yeah, a bunch of these launches exist because some big platform changed the deal on people. Prices doubled, features got removed, data got harder to leave with. And instead of just complaining, a few solo developers rebuilt the thing they lost, on honest terms.

Milo: And then the flip side of that — what happens when the software isn't a tool you keep, but almost a worker you hire. We've got AI that runs client work, AI that runs a little store, and some interesting questions about how much autonomy any of that should really have.

Mia: Plus how AI work survives when you switch tools, how developer tools get found in the first place, and a couple of small craft projects that just exist because someone cared. Let's start with the ownership story, because it's the most emotional one.

Milo: Okay, Harbor. This is a notes app — a "second brain" kind of app — but the story behind it is basically an Evernote horror story. The founder, Spicer, from Cloudmanic Labs, used Evernote for more than a decade. Scanned mail, kids' report cards, meeting recordings, fifteen years of notes.

Mia: And then Evernote got sold, the price roughly doubled, the free plan shrank to fifty notes, and exporting your own data got harder. He's actually fair to Evernote about it — he says the app is fine. And he backed that up with data: they counted the 500 most recent US App Store reviews, and of the negative ones, only 9% mention bugs or crashes. 54% are about price.

Milo: That's the interesting finding. The product got better while the deal got worse. So Harbor is built to fix the deal, not just the features.

Mia: So what does the deal actually say? One plan, $8.25 a month on annual billing, no ads, no subscription tiers, no credit card to start. And the promises are explicit: the price is locked for three years, then increases are capped at 10% a year. That's written down on the site. And if you stop paying, your account doesn't vanish — it goes read-only and you can still export everything.

Milo: Which is the "never held hostage" promise. Full export on every plan — HTML, Markdown, ENEX — plus an open API, a CLI, and an MCP server. The phrasing is great: they want you to stay because Harbor is good, not because you're stuck.

Mia: On the product side: native apps on Mac, Windows, iPhone, iPad, Android, web, and a CLI — no Electron, works offline. Search does OCR, so it reads words inside photos and scanned PDFs including handwriting, and recordings become searchable transcripts.

Milo: One more thing worth naming: bring your own AI. There's no built-in assistant. You connect the ChatGPT or Claude you already pay for and let it search and organize your notes. Plus optional zero-knowledge encryption per note or notebook — your keys, not theirs. And one-step Evernote import for notebooks, tags, checklists, attachments, web clips.

Mia: And the community heard that. One commenter said Evernote has made it impossible to export, and hoped Harbor fixed that — which is exactly the pain point the import feature targets.

Milo: Now, honest framing here — all of these are the maker's claims. The price lock is a promise, not a track record. Spicer says they've been building and running software for twenty years and Harbor is built to run for the next twenty, but the guarantee is only as good as the company behind it. Still, the structure of the promise is real: export on every plan, a written price cap.

Mia: Same instinct at the operating system level, though, with Launchie. Apple removed Launchpad in macOS 26 Tahoe, and kept it out of macOS 27. Their replacement is the Apps view in Spotlight, which sorts everything for you and doesn't let you make folders.

Milo: And Nick, the solo developer, lost the way he'd opened apps for years — a grid he arranged himself, folders he could find without reading a single name. So he rebuilt it.

Mia: And this one has genuinely strong third-party evidence. It's 4.6 out of 5 across 141 ratings on the Mac App Store, and it was reviewed by iDownloadBlog — Ankur Thakur called it the closest you can get to an actual Launchpad on a Mac running Tahoe. Also mentioned in Mac & i and MacGadget.

Milo: What does it do beyond the old Launchpad? Your own layout with drag and drop, unlimited folders — all free. Spaces, which are separate pages of apps for work, games, or design, switched with a swipe. Instant search and a Quick Access section for recent and most-used apps.

Mia: Pro is a one-time purchase — no subscription — and adds your wallpaper blurred behind the grid like the original, hot corners, hidden apps, and a pinch gesture that opens Launchie instead of Apple's Apps view. It's on the Mac App Store, direct download, and Homebrew.

Milo: The reviews are worth quoting because they're attributed, not invented. One reviewer said they tried all the replacements they could find and this is by a wide margin the best, even on the free tier. Another said it's better in every regard than the old launcher — sorting tools, additional folders, hiding apps. And one called it "the perfect un-Sherlock'ing of Launchpad."

Mia: And Nick says most of the new features since the 2025 launch came from user requests — one reviewer even noted their suggestions got quickly implemented. That's the continuity story: Apple retracts a feature, an independent developer keeps it alive and grows it.

Milo: There's a smaller macOS story in the same family. SmartPause. It fixes something tiny but daily: on a Mac, the play/pause key goes to whatever macOS last remembers as "Now Playing." So you're watching YouTube, you press pause, and the video keeps going while Apple Music opens.

Mia: Which is maddening. SmartPause catches the key, asks Core Audio which process is actually outputting sound right now, and sends the command to that app. If nothing is playing, the last thing you paused resumes. Made by Yasin Özmen, free, MIT licensed, no analytics, no accounts, one permission — Accessibility.

Milo: There's a clever touch: if two apps are playing, a single press switches — the playing one pauses, the other starts. Double press plays or pauses the selected app. And a little widget shows what happened and what the next press will do. Apps without an adapter get the key handed back to the system untouched, so it's never a dead key.

Mia: It's open source, written in Swift, you can read every line. He's asking for coffee donations toward the Apple Developer ID so the app can be notarized. It's a small honest tool, but it's the same philosophy: fix the mismatch between what the user means and what the system does, transparently.

Milo: Let's bring in the support piece here — Scrapboard Cloud 4. It shows Apple-ecosystem continuity done a different way. It's the mobile companion to Scrapboard 4 for Mac — a digital bulletin board that's been running for twelve years, recreating the feel of pinning things to a family refrigerator door.

Mia: It syncs boards across iPad, iPhone, and Mac via iCloud and CloudKit, with push notifications when edits sync from another device. Six layer types: text with full typography, pictures, SF Symbols, appointments with countdowns and reminders, live embedded web content, and maps. Multi-page boards, multiple projects, push and fetch.

Milo: And notably, the developer says no data is collected at all. Free app, made by Tomoyuki Okawa. It's not reacting to a retraction like Launchie — it's more like the quiet counterexample. Someone maintaining their own little ecosystem product for over a decade, on Apple's platform, without drama.

Mia: Okay. So we've talked about tools you keep. Now the shift: what happens when the software isn't a tool at all, but something closer to a hire.

Milo: This is where Mycel comes in. It's for service businesses — the people whose actual product is client work: reports, updates, decks, invoices. And the pitch starts with a brutal bottleneck: you are the last pair of eyes on everything, and that caps how many clients you can carry. It gets worse as you grow.

Mia: The founder lays out the three ways out, with prices attached. Hire at $45k plus six weeks of ramp. Keep doing it yourself every evening. Or turn the work down. And he's dismissive of the obvious alternative — a chat window drafts fine, but it forgets what you corrected the moment the tab closes. Month six costs you what week one did.

Milo: So how does Mycel work differently? Every job gets its own machine — a sandbox that boots, does the work, and is destroyed. The agent never holds your keys or credentials. Then checks run on the output — and this is the detail people caught onto: if a summary says four findings but the table lists five, the draft is held and made to say why. You never see that draft.

Mia: It writes against what that client received last month, so an open item doesn't vanish between deliveries. And here's the learning loop: you edit before you approve, and both versions are kept. The diff is what it learns from.

Milo: It also runs on its own clock — five timers are live in the demo. Chasing what's owed, resuming stalled work, keeping the business findable. The founder's line is good: "Nobody starts them in the morning." And nothing reaches a client without your sign-off.

Mia: Pricing: from $299 a month, or self-host free. And he ended with a question that got real responses: what's the piece of client work you'd never let a machine touch?

Milo: The community zeroed in on exactly the right things. One commenter loved the approval layer and learning from past edits — AI that remembers how you actually deliver, versus another chat window. Another asked the practical question: the correction loop — if I fix a draft for one client whose reports look nothing like another's, does that fix change how the others get drafted?

Mia: That's an open question, and worth flagging as such. Nobody answered it in what we have. Same with someone asking how much of a draft is usually right first time — how much is reading and fixing versus just signing off. Those are the real diligence questions and they're still open.

Milo: Also someone asked whether deliverables have to be files at the end, or whether it can publish something — and noticed what looked like ads in the video. Unanswered in our material. And one person was honest that they had to read the pitch a couple of times to understand what it does — but once they got it, they knew exactly who to recommend it to. That's a fair criticism of the messaging, and a decent signal of the value.

Mia: Now take the same bottleneck to the opposite scale. Minicart. Everyone has something worth selling — crochet animals, trading cards, resold sneakers — but running the store is the hard part. So Minicart gives you a store plus an AI team.

Milo: Three named agents. Sloane starts the business: creates the storefront, turns photos into product listings, sets up inventory and payments. So you literally upload a photo of something you made, say "I want to sell this for $15," and she creates the listing and imagery. Milo grows it: promotional discounts, social posts, marketing campaigns, publishing to Instagram. Logan operates: orders, shipping and logistics, drafted customer support replies.

Mia: The claim from co-founder Chris: 4,000-plus stores already built, spanning handmade goods, collectibles, apparel, and services. Treat that as the maker's number. And the differentiation argument is that traditional store builders give you tools to configure, while Minicart's agents work inside one connected platform using your products and business context to take action.

Milo: The CTO's comment adds a technical claim: an AI harness purpose-built for commerce. And they offered the Product Hunt community $1 a month for the first three months with code PH2026.

Mia: The community asked the exact same question here as they did of Mycel — how much autonomy is real? One commenter directly asked: can Milo and Logan actually execute end to end, or are they still mostly suggesting and waiting for approval?

Milo: Others asked about the review process — can you approve or edit at each phase or only at the end? Whether Sloane handles discount codes and seasonal changes or that's manual. Whether Logan sounds like a real person when replying to customers or it's obviously a bot. Whether the agents are customizable. Whether it works internationally or just in the US. Whether there's a mobile app.

Mia: And notice — none of those are answered in what we have. That's the honest state of it: the concept drew genuine interest, several people specifically liked that the agents are separate for different functions rather than one general AI, and one commenter made a sharp observation that agents sharing business context across marketing, inventory, and support could avoid unnecessary communication and save time. But the autonomy questions are live.

Milo: Worth noting there's a third product in this neighborhood. ManyPI — it's more of a sales agent. Describe your ideal customer in a sentence, it finds matching companies and the people who sign, verifies every email before you send, and runs multi-step cold email campaigns from your own inboxes with warmup and replies in one place. Plus workflow automation — a lead replies, it tags it and starts the next step. CRM, unified inbox, AI agent, scrapers, analysis in one subscription.

Mia: It just shipped an MCP server, so you can ask Claude, ChatGPT, or Gemini for leads and the list lands in your table, not in the transcript. Community questions there were also about plumbing: how long does lead research take, how does it find leads — raw web search or partners like Apollo — and how are the contacts verified and kept updated. Again, good questions, open ones.

Milo: So that's the worker-hiring theme. And it sets up the next thing perfectly, because if AI is doing work for you, the question becomes: what does it need to remember when the human steps away?

Mia: Epismo. The founder's frustration is precise: he was tired of being the integration layer between AI tools. We start in ChatGPT, continue in Claude, finish in Cursor, hand it to a teammate. The models are getting very good at the job, but the work doesn't move with them. A draft can move. The decision behind it does not.

Milo: And he has a story from preparing this very launch that proves the failure mode — and he's generous in telling it. They had already decided not to say that Epismo automatically captures every conversation. A draft moved from one AI to another, but that decision didn't travel with it — and the next version implied exactly the thing they'd decided not to claim. Nothing was wrong with the model. They'd copied the text, not the work around it.

Mia: So Epismo's answer is the Case. A Case holds the current result, the decisions behind it, the reviews, and the next step — outside any one chat. His framing: a shared doc holds a draft, a task tracker holds an owner, an AI chat holds a session. Epismo holds the active state of the work.

Milo: Then there's auto review. Once the work is gathered, Epismo gives the saved result a fresh second look, flags issues, and leaves the original unchanged — so the next person or AI continues with the context in view. Their example: a renewal brief drafted from last quarter, and auto review flags "usage isn't sourced, the churn claim is still an assumption." Then the next turn in Cursor fixes it, with the notes in view.

Mia: And here's a nice bit of self-awareness: they used Epismo to prepare their own launch — positioning, launch video, sales pitch, the Product Hunt launch itself — and published the actual trail as a link. Also, the workflow pattern is work-first: save the work, hand it off, and only then, if it worked, does it become a playbook. Steps can even name the specific Skill, MCP, or CLI to use.

Milo: One skeptical comment is worth mentioning. Someone pointed out this is the founder's seventh launch on Product Hunt and asked what's actually new versus a repackaged pitch — what specifically changed in the product since the last round. That's a fair challenge, and it's unanswered in our material. There's also a jokey one — "copy and paste at least makes me feel as if I'm still needed."

Mia: Context.dev attacks the same plumbing from the other side. Yahia, the founder, describes the familiar pain: a simple question becomes an engineering project. Find the right pages, scrape them, feed them into a model, get the output into something your code can use. Their answer is /answers — one API call. You give it a research task and an example of the JSON you want back. It searches the web, reads relevant pages, and returns structured data with source URLs.

Milo: Examples they give: find this company's pricing plans and what's included. Does this product support self-hosting? Compare these three tools on pricing and integrations. There's a Fast mode for focused lookups and Ultra for deeper research. You choose the output fields so the answer fits directly into your code.

Mia: The theme across both is the same: context, not model quality, is the bottleneck. And one early user comment supports the claim in a limited way — someone said they integrated it into their agent loop for product research and outreach and the output quality is great. One person's experience, not a benchmark.

Milo: And there's the demand-side story for all of this: developer tools are being built with these agents, and then those developers have to find customers. Which brings us to Morsa Signals.

Mia: Sergei and Stan built it from their own pain — they're a dev team who built it for their own needs first. The problem: strong developer products often struggle to reach the right developers, explain their relevance, and become discoverable through search and AI systems.

Milo: Four workflows, and the key design idea is that each one starts from your actual product, audience, and problem — producing a focused result instead of a generic report or database. Contact Search finds relevant developers for scaled outreach, explains why each person fits, and suggests a personalized opening angle. So not just a name — "saw your issue about this exact thing."

Mia: First Users is for solo and early-stage founders without a sales team — paste your product, get a small hand-picked list of developers already dealing with the problem you solve, matched from bio, repo activity, and public technical signals, sized for a first outreach round. The site shows example rows: someone who opened an issue asking for exactly this kind of tool, someone whose repo shows a hand-written script solving the same problem.

Milo: The AI Visibility Audit checks whether ChatGPT, Perplexity, Claude, and Google AI Overviews have enough product context to classify, compare, and recommend your tool. It scores technical readiness, positioning, and public footprint, flags gaps — like weak category signals or missing comparison context — and generates a ready-to-use prompt for addressing them. Their sample output shows a 62 out of 100 score.

Mia: And Competitor Tracking gives a first read on comparable dev tools — GitHub activity, releases, pricing changes, launches on Reddit, Hacker News, Product Hunt. Plus a positioning workflow to pressure-test your category and ICP, and a monitoring workflow that watches developer channels for conversations worth joining — though they're explicit they never post or reply on your behalf.

Milo: The open question the audience keeps raising: how do the contact lists stay relevant and updated? One commenter directly asked how they make sure the contacts are relevant and current, given how specific the audience is. That's unanswered in what we have, and it's the crux — the whole product's value rests on the quality of those matches.

Mia: Okay. Let's end on the craft stuff, because there's a nice thread of honesty running through it too. YABAI — which is literally the word the app is named after. It's a Japanese slang dictionary app with 308 real slang, internet, and fandom words.

Milo: The maker's a solo developer from Japan who kept meeting people who'd studied Japanese for years and still couldn't follow a group chat. The vocabulary that actually carries everyday conversation is exactly what textbooks leave out. But he didn't want another study app — no streaks, no tests, no XP, nothing that guilt-trips you for missing a day.

Mia: The format is a snack: one word per card with giant Japanese typography, swipe to discover. Every card has a meaning, an original example sentence, a vibe check — safe to use, friends only, or be careful, plus tone hints like self-deprecating or Kansai dialect — a culture note, and a story about where the word came from.

Milo: And here's the part he says he spent the most time on, and it's genuinely unusual: honesty. Every word is researched and written in Japanese first, then translated into 13 languages — not machine-translated out of English. And where a word's origin is genuinely disputed, the app says "nobody agrees" instead of picking a tidy story. Yabai itself is the example — its origin is disputed, and the app just says so.

Mia: He's explicit that invented etymology is everywhere in slang content and he didn't want to add to it. It's a paid app, no ads, no sign-up, no subscription, works offline, built end to end with AI agents in Claude Code.

Mia: He asked learners which words surprised them and what's missing — and got practical questions: someone asked if there's an iOS version, someone asked about free versus paid, and someone asked about the specific number 308 — is it meant to be a complete "you can follow native content now" set, or a starter pack with more planned. Another open question.

Milo: Then the one that made me smile: The 101 Plays Itself. Jay Judah turned a live Caltrans traffic camera on the southbound US-101 in Studio City into a musical instrument. In your browser, right now — every car that crosses the line plays a note.

Mia: And the "how it works" section is genuinely detailed. Nothing is pre-recorded — the timestamp burned into the top of the picture is the real one. Come back at three in the morning and you'll hear an almost empty freeway. Ten sampling patches are laid on the road surface, two per lane, and they're not rectangles — each is a slice of its own lane, leaning and narrowing with distance exactly as the lane does.

Milo: The browser reads those patches fifteen times a second and watches for the road to stop looking like the road. That's a car — and that's a note. The two patches per lane are 13.4 metres apart, so the gap between triggers gives the car's speed — it's literally a speed trap. And how long a vehicle covers a patch gives its length, which is how the page tells a truck from a car and gives it a different voice.

Mia: Detection and synthesis run entirely in the browser — no server, no samples, no recordings. If the camera's down, the page tells you Caltrans' whole network is unreachable, and offers three recorded minutes of the same freeway instead. Lane geometry was solved offline from an earlier recording.

Milo: It got noticed, too — Jason Kottke featured it as "five lanes of LA highway turned into an instrument," and NBC4 Los Angeles did a segment, framing it exactly right: a local freeway, a public camera, and real traffic composing the score in real time. The Drive called it "strangely beautiful." Video used under Caltrans' conditions of use.

Mia: Which loops us right back to where we started. A public camera, honest about being live, honest about when it's down. A slang dictionary that says "nobody agrees" about etymology. An SSH client with keys stored encrypted on your phone. A notes app that prints its price cap.

Milo: Oh — Termphin, actually, we should do that one properly before we close, because it fits the craft theme and it's genuinely practical. It's an SSH client where sessions never drop. The founder kept losing work on his phone — a long-running job, an editor session, or an AI coding agent would die the moment he locked the screen or switched from wifi to mobile data.

Mia: The usual fix is tmux, but that means installing and configuring it on every server. So he added a small Rust helper that keeps the remote shell alive even when the connection drops. The app uploads it over SSH, and a foreground service keeps the socket warm so the system doesn't freeze it.

Milo: And this matters more now because of the coding agents — he specifically calls out that it works well with Claude Code, Codex, OpenCode, and Kilo Code. Their tasks run long, and you don't lose them when the connection drops or the app process is killed. It reattaches you to the same session.

Mia: The details are thoughtful: the terminal is drawn natively, not in a web view, so htop, vim, and full-screen TUIs render like they do on a desktop — diffs and progress bars draw correctly with no mangled box characters. SFTP browser in the same tab as the shell, snippets you run with one tap, port forwarding, jump hosts like ProxyJump where the bastion carries traffic but can't read it.

Milo: Security posture: keys stored encrypted on the phone behind a PIN or fingerprint, nothing of theirs in the path between your phone and your server, server fingerprints checked every connection. And the terminal renderer and the server helper are open source, so you can read what runs on your machine. It's free, no ads, no in-app purchases, on Google Play now — iOS and desktop in progress, no release date.

Mia: One sharp community question: is the server-side agent something you install yourself, or does Termphin push it automatically on first connect? The site says the app uploads it over SSH, so it's pushed — but the question shows people are thinking about what runs on their servers. Also worth noting honestly: a server reboot still kills the shell — Termphin reattaches you to a fresh one in the same directory, previous state gone. They say that outright in their FAQ.

Milo: So let's pull it together. Today's pattern: the platforms changed the deal, and individuals responded — Evernote doubled prices and Harbor wrote a price cap; Apple removed Launchpad and one developer rebuilt it better than ever; macOS misroutes a media key and someone fixed it in a free MIT app.

Mia: Then the deal got bigger: AI as a hire. Mycel for client work, Minicart for the little store, ManyPI for the sales pipeline — with the same live question hanging over all of them: how much autonomy, and how much stays human-approved?

Milo: Then the memory question — Epismo keeping the decisions when the tools change, Context.dev making research a single API call — and the demand side, Morsa Signals, helping the developers who build all this actually get found by other developers and by AI search.

Mia: And the craft closes — YABAI refusing to invent etymology, a freeway singing itself, Termphin refusing to drop your session.

Milo: The honest note to end on: everything today is maker claims plus community questions, not verified results. The most useful signal, honestly, is how many sharp open questions each product attracted — because those are the ones to ask next.

Mia: That's the briefing. Thanks for listening — we'll see you tomorrow.