0914 | Private by Design: This Week in Local-First AI

Mia: Welcome back to the show. I'm Mia.

Milo: And I'm Milo. Today's briefing is built entirely from launches and discussion posted in the last twenty-four hours, and there's a thread running through all of them: where your data lives. Some of these tools keep everything on your machine. Some split the work between your machine and the cloud. And one of them is about making the local side of that equation genuinely cheaper and smarter.

Mia: Yeah, and we'll end on how products get seen at all in this new landscape — demos ranked by audiences, and a couple of small tools that respect you. Let's start with the fully local camp, because three launches today make the same bet in very different ways.

Milo: The first one is Resurf. It's a personal context library — the maker, a design engineer named Deep, describes it as an app for things you like, care about, and work on. You save notes, links, images, PDFs, videos, voice memos. Then you organize later with spaces, tags, and stacks. It's fully native on Mac, iPhone, and iPad.

Mia: And the interesting part isn't just the capture. It's the handoff to AI. You can ask questions of your library inside Resurf, or pass that context to your other AI tools through MCP and CLI. So it's not a notes app that ends at notes — it becomes the memory your AI assistant draws on.

Milo: The origin story is relatable. Deep said he was saving inspiration, articles, screenshots, random ideas, and they ended up scattered across apps — and he rarely came back to them. So he built Resurf for himself, shared earlier versions on Reddit, and the feedback shaped the app. That's a real iteration loop, though obviously it's the maker describing his own product.

Mia: On pricing: free to try on Mac, free on iPhone and iPad, and a one-time purchase for unlimited Mac use. No subscription, which fits the whole ethos here. Your library lives locally, with optional private iCloud sync.

Milo: One community question worth flagging, because it's a genuinely open issue: someone asked whether you can limit MCP access to specific spaces or collections. That's the crux of handing your library to AI — if an AI tool can read everything, can you scope it down? As far as the discussion shows, that's still an open question.

Mia: Now, Resurf is about capturing everything. The second product is about capturing one specific, hard thing: meetings. This is Epilude Notetaker, and the pitch is meeting notes that are one hundred percent private.

Milo: The mechanics matter here. Epilude records the meeting on your Mac itself — no bot joins the call. So there's no "AI Notetaker" sitting in the participant list, which changes how people behave on the call. If your Mac can hear the meeting, it can take the notes, whether you're in Zoom, Google Meet, Teams, a Slack huddle, or just talking across a table.

Mia: And the transcription and summarization run entirely on-device, using an in-house trained model they call EP-1. Nothing leaves your Mac. The maker, Gustav, said the product started as a way to write with his voice after he hurt his elbow — Notetaker is the next step of that.

Milo: The reasoning is pretty honest: existing AI notetakers solve the same problem but send your data to cloud providers, where your meeting could become part of their training data. Gustav framed the underlying problem well — either you focus on the conversation and forget half of what was agreed, or you take detailed notes and miss the conversation in front of you.

Mia: What I like is the output isn't generic. Epilude classifies the meeting type and adapts the summary — a standup, a sales call, and a planning session don't all come back in the same format. A standup summary surfaces decisions, blockers, commitments. Action items get owners and due dates, and here's the part worth emphasizing: it extracts only what was agreed out loud. It never invents an owner or guesses a deadline.

Milo: And every key point links back to the transcript, so you can check what was actually said. Each meeting saves as a Markdown file in a folder you choose — Documents by default — so you can open it in Obsidian, paste it into Notion, or export a PDF. Your notes are files you own, not rows in someone's database.

Mia: There's practical detail too. It detects when a call starts and offers to record with one click, and if your Mac restarts mid-meeting, the recording is recovered. Free trial covers a thousand words and ten meeting notes, no credit card.

Milo: A couple of open questions from the community. One commenter asked how it handles meetings where people switch between languages — that's untested territory as far as the discussion shows. Another raised a sharper concern: the pay-once option only includes one year of app and model updates, and they asked directly whether that means buying the app again every year, especially if it breaks when a new macOS drops. Those are fair questions and they weren't answered in what we have.

Mia: There was also praise worth noting — one commenter said the lifetime option with no cloud dependency "feels like the good old days," and that app purchases are back thanks to local AI models. That's one person's reaction, but it captures why this category is interesting right now.

Milo: Third product in the local camp, and it's a different world entirely: Kirokune. This is an iPhone app for keeping work incident notes — the kind of documentation people always say you should do. Notes, recordings, and photos in a timeline, no account needed.

Mia: The use case is heavier than a notes app. The team behind it, AMA CLASS, describes the starting point as a small task: write down what happened in a work conversation while the details are still fresh. One tap starts a recording even when you can't bring yourself to type. A photo or a quick note works too.

Milo: Structurally, it separates the event date from the writing date, and observations from personal reactions. It keeps everything in time order — dates, times, exact words — and nothing you already wrote gets changed. Later, when you're ready, you can pull it into a summary you can walk someone through, like HR, a lawyer, or a support service.

Mia: On privacy: record content stays on the iPhone unless you choose to share it. There's no automatic sync or cloud backup at all — and they're explicit that restore-purchases doesn't recover records, so you need to keep your own copies before deleting the app. That's a real limitation, stated plainly.

Milo: Pricing: unlimited records in one case is free. Extra cases, detailed reviews, PDF export, and ad removal are paid via Kirokune Plus — monthly, yearly, or a one-time purchase. There's also a free worksheet with a fictional example you can use without even installing the app.

Mia: And I want to credit the makers for their restraint here. They're explicit that Kirokune helps organize records — it does not determine harassment, provide legal or medical advice, or guarantee evidentiary value. It doesn't diagnose or decide what happened. They even warn about recording laws varying by place and situation, and ask users to share fictional examples rather than personal records in feedback.

Milo: Practical availability: iPhone with iOS 17 or later, currently in the United States and India, with English, Japanese, Traditional Chinese, Korean, and Spanish interfaces. You can also import chat export files from Apple Messages, Messenger, or Instagram without connecting your chat accounts.

Mia: So that's the fully local camp: Resurf as your context library, Epilude as your meeting memory, Kirokune as your incident record. All three keep everything on your device.

Milo: But local-only has a ceiling. What happens when you need cloud-level power and local privacy at the same time? That's where the next two products live — the hybrid approach.

Mia: The big one is Perplexity Hybrid Compute, now live on Perplexity Computer, their Mac app. The idea: one task runs in two places. Research and reasoning happen in the cloud with their strongest AI. Anything touching your private files stays on your Mac with a local AI, so it never gets uploaded.

Milo: The example they walk through is an LBO model rebuild. The cloud side researches comparable deals — multiples, recent transactions, return profiles. The Mac side reads the meeting notes and rebuilds the model with three cases. The results land back in the same task. You can even trigger tasks remotely from your iPhone while the sensitive steps still run on your Mac.

Mia: The privacy mechanism is the interesting engineering. There's an on-device PII classifier — they call it Privacy Gate — that reads each task before it's sent. It's trained in collaboration with Perplexity's Secure Intelligence Institute, and they've open-sourced it. When it finds personal information — names, email addresses, phone numbers — it can mask them with stand-ins, block the upload, or ask you first. Then it restores the real values when the answer comes back.

Milo: So in that legal brief example, the case files get reviewed locally, and only anonymized questions go to the cloud for case law research. The draft edits happen in place on the Mac. You also get a choice at the moment: process locally, or upload as-is.

Mia: Requirements and availability: Apple silicon, twenty-four gigabytes of RAM minimum, and it's live now for Pro, Max, and Enterprise plans. There are three local models to choose from, one-click setup, no Ollama or API key needed. Enterprise gets admin controls and audit logs for what leaves a device.

Milo: And that's the honest trade-off this category is negotiating: local AI keeps things private but is weaker; cloud AI is smarter but means uploading sensitive stuff. Hybrid Compute is Perplexity's answer to not choosing. Of course, all of that is the maker's own description — how well the split actually works in practice is something we can't verify from here.

Mia: GhostWriter by MyHandler takes a related but different angle. This is Windows software — the maker, Tery, has been shipping for forty years and jokes that MyHandler is "local-first AI, for the rest of us who don't have a Mac Mini in a closet."

Milo: Here's the interaction: you put your cursor in any text field, in any app, and double-tap Left Ctrl. GhostWriter reads what's on screen around that cursor — the thread above it, the question beside it, the half-finished sentence in it — works out what kind of writing this is, and types a draft at your cursor.

Mia: It draws on context indexed in an encrypted vault on your own PC — your screen, files, mail, dictations, calendar, and past meetings. Not their servers. If your draft proposes a time, it cross-checks against the next two weeks of your calendar first, so it won't double-book you.

Milo: And it tries to get the situation right before writing: a reply owed in a thread, a sentence to continue, a blank box under a question. It maps each message to its sender and works out which handle is yours, so it answers the other party rather than summarizing your own words back at you. It also matches how you write to that person, down to the sign-off.

Mia: The key design line: it types into the box and never sends. There is no send path in the product. You read it, edit it, hit send yourself. Now, one transparency note from their own site: the assembled context is sent to their cloud model — with zero data retention, they claim — to write the draft. So it's local indexing, cloud drafting. That's a hybrid too, just with the boundary drawn differently than Perplexity's.

Milo: Availability: Windows 10 and 11, free to start with three hundred credits a month, no card, and a Product Hunt code takes thirty percent off. macOS and Linux are coming soon, per the maker.

Mia: Community questions worth carrying: one commenter asked what it actually sees when it reads your screen, and another asked how it knows things about you — whether you have to connect something like Supermemory. Those are the right questions for a screen-reading product, and the answers aren't in what we have.

Milo: So we've gone from fully local to hybrid. But the hybrid approach still needs local AI to be good, and local AI has traditionally meant weaker AI. The next launch attacks exactly that problem — from the cost side.

Mia: Cognition's SWE-2. This is their new coding model, and the headline claim is frontier-adjacent performance at a fraction of frontier pricing. On FrontierCode 1.1 Main it scores 50.0% — within one point of Fable 5.1 while being 64% cheaper, and within a few points of GPT-6 Astra at a quarter of the cost.

Milo: The training trick is what makes this worth talking about. SWE-2 is post-trained from Kimi K3 — a 2.8 trillion parameter model that had already been through extensive agentic coding RL. Cognition's addition is an RL algorithm that puts the actual dollar cost of a run into the training reward.

Mia: And they train all three effort levels — medium, high, max — in a single run instead of stitching together separate models. They say this advances the whole cost-performance frontier, not just the top end. The penalty per effort level is tuned to the slope of the base model's Pareto frontier. That's first-principles-sounding, but it is their own framing.

Milo: The efficiency numbers against their own previous model, SWE-1.7, are concrete: 58% fewer turns, 81% lower cost, and a higher score. On their step counts, SWE-2 medium makes its first real code edit after a median of 18 steps, compared with 48 for SWE-1.7. They acknowledge users said SWE-1.7 over-explored simple tasks, and SWE-2's gains come from focused exploration — judging which parts of the codebase actually matter.

Mia: Behaviorally, they report three things: better end-to-end test writing that catches regressions and edge cases; resourcefulness within your boundaries — in one case an MCP integration was unavailable, so it reconstructed the data from Slack channel history it already had; and verification discipline — when you push back, it re-derives conclusions instead of just agreeing with you.

Milo: On Terminal-Bench 2.1 it scores 92.8%, which tops their comparison table. It's available now in Devin Desktop and CLI, rolling out to Devin Web and Fusion. The scaling note is also notable — this is the first time they've scaled RL to the multi-trillion-parameter regime.

Mia: What's unknown, honestly, is real-world reliability at volume. These are their own benchmark runs — three runs per task in their methodology, but still their numbers. If you're running coding agents at volume and per-task cost is a line item, this is the model aimed at you. Whether the cost savings hold up on your workloads is an open question.

Milo: And cheaper local-class AI loops back to everything we've talked about — it's what makes hybrid products like Perplexity's viable. But there's another shift happening that makes AI capability matter in a different way: buyer discovery is moving into AI answers. Next up: how you win visibility there.

Mia: Visiby is a tool for tracking your brand's visibility across AI search — ChatGPT, Claude, Perplexity, Gemini, and Google AI Overviews. The pitch: traditional SEO tools don't show you what AI assistants are saying about your brand. Visiby samples prompts, parses model answers, and tells you where you're mentioned, where competitors are recommended instead, and which citations you're missing.

Milo: The workspace has real depth on paper: per-engine citation share, a prompt explorer with over fifty thousand tracked prompts weekly in their own numbers, a citations explorer ranking every URL the engines cite, competitor intelligence with per-competitor recipe diffs, and priced action briefs. They also track "adjective ownership" — which words like "trustworthy" or "enterprise" attach to your brand versus competitors.

Mia: Now, their marketing makes strong claims — one in three B2B buyers starting vendor research in an AI assistant, 62% of enterprise marketers already optimizing for AI engines, all attributed to their own research or buyer panel. Treat those as Visiby's numbers, not independently verified.

Mia: And the community caught real inconsistencies: one commenter asked about the Starter tier's refresh frequency — the description says daily tracking but the feature list says weekly refresh — and noted the listing mentions Gemini while the pricing page lists a different engine lineup. Those are exactly the questions to ask before paying.

Milo: Another commenter asked how it differs from trakkr, a tool they use today — so there are existing players here, and differentiation isn't established from our sources.

Mia: Neopress is the supply-side complement. It's a workspace where you build and grow your website by chatting with AI — website creation, CMS, SEO, and analytics in one loop. The specific bet: publish server-rendered, AI-readable pages that search engines and AI crawlers can actually access, then ask AI to analyze your metrics and turn them into improvements. Build, publish, measure, learn, improve — all through conversation.

Milo: One of the makers, Joon, put the target question to the community directly: what takes the most work after launching a website? There's a live demo at demo.neopress.app. Compared to stitching together a site builder, an analytics tool, and an SEO consultant, Neopress wants to be one continuous conversational loop. Whether that loop is actually good at each stage is the open question — we only have the maker's description.

Mia: But together they frame a new discipline. If buyers start research inside AI assistants, brands need to both measure what models say about them — that's Visiby — and publish content those models can read — that's Neopress. And that connects straight to our next pair: if AI answers increasingly decide what gets discovered, a good demo of your product matters more than ever.

Milo: DemoTV is a genuinely odd and interesting launch: a 24/7 channel of product demos, ranked by the people watching. You watch two demos head-to-head in a Channel Battle, pick the one you'd actually try — no sign-up — and an Elo-style audience score from thousands of votes decides who sits on Channels 1 through 3.

Mia: The maker's framing: every product has a slick demo video, and none of them tell you whether the thing is actually good. The polish belongs to the product, not a reviewer. So DemoTV removes the reviewer and lets the audience decide.

Milo: And the line they keep drawing — and it's the whole point, in their words — is that money can buy exposure but never rank. Submitting a demo is free, reviewed before it goes live. Businesses can buy clearly labelled airtime: real views, completed views, clicks. But airtime never touches the Channel rank.

Milo: Their station numbers — around 109 live demos, 542 audience backings, 53,700 views — are their own counts, and the founder explicitly asked whether "channel" works for demos the way "leaderboard" works for launches. That's an open format question, and one commenter suggested it could be cool to stream the best commercials of all time, which is a fun extension but just one person's idea.

Mia: Now, if your product wins attention on a channel like that, you need a good demo — and making one is traditionally expensive. ScreenCursor attacks that cost. It's a Chrome extension screen recorder where every click, drag, and keystroke becomes an automatic camera zoom. The zoom arrives a beat before the action, holds steady, then follows your cursor. When you stop recording, the zooms are already placed — you get a finished video, not a project to edit.

Milo: The origin: the maker saw every good product demo made with Screen Studio, which is Mac-only and runs about $108 a year. He's on Windows, so he built ScreenCursor. It runs anywhere Chrome runs — Windows, Mac, Linux — no installer, no admin rights.

Mia: Privacy pattern again: recording, editing, and export all happen inside your browser. Nothing is uploaded, no account, no server, no watermark. And the zooms aren't baked in — you can retime them, push depth in or out, switch between snappy and gentle motion, or delete them entirely. Exports MP4 at up to 1080p.

Milo: It's $49 once, lifetime, updates included, and works on up to three of your own computers. There's a Product Hunt discount of fifty percent for the next seven days — $24.50. The maker specifically asked to hear what breaks on Windows and Linux, since that's who he built it for and where he has the fewest testers — a candid limitation. And honest comparison from their own FAQ: Screen Studio does more — webcam overlay, motion blur, iPhone recording over USB.

Milo: Community questions also probed whether it captures the whole desktop or just the browser tab, and asked about zoom behavior outside the browser, which the maker indicated doesn't work perfectly yet.

Mia: So DemoTV is where demos compete, ScreenCursor is how you make one cheaply. And both echo the theme: one-time purchase, nothing uploaded.

Milo: Last segment — two smaller items that close the loop on respect for the user. First, Clipwise. It's a Notion web clipper from a solo developer, and it's completely free with no account.

Mia: The maker's grievance is specific: every Notion web clipper he tried wanted an account and a login — and those logins kept breaking or syncing slowly, right when he just wanted to save something. Clipwise removes that layer. You connect it once with your own Notion integration — a single paste — and after that it just works. No account with the developer, no server; nothing leaves your computer except the page you choose to save, which goes straight to your own Notion.

Milo: The craft details are what make it credible. It saves the whole article, not half of it — including code blocks and content some clippers miss. You fill in your Notion properties — tags, status, source — before saving, so the page lands finished instead of as an empty row. It remembers your last database, so the second clip is basically one click. And if a page comes out empty, it tells you instead of silently saving nothing.

Mia: One community suggestion worth passing on: a commenter loved the properties-before-saving touch and asked whether Clipwise can detect when an article is already in the database. That's a natural next feature, and it's an open question, not a shipped one.

Milo: Last item, and it's a different kind of launch — a design artifact rather than a tool. SHIUI is a UI kit in the spirit of Hinomaru Japanese posters. Flat 2D, neat 1.5 pixel lines, zero shadows, and only three colors: paper, a warm off-white; sumi ink, a near-black; and shu, a vermilion accent spent sparingly — sun, seal, shout, as they put it.

Mia: It's not just a mockup — the site shows 111 live elements and 24 blocks, all interactive on the page: buttons, form inputs, toggles, navigation, and you can copy the import or full source straight from each card, plus copy the design tokens. Buttons follow a 44-pixel base, primary buttons use the vermilion fill with the rule of one per view, and there are charming touches like hanko seal components and a tilted vermilion price stamp.

Milo: One community comment captured it well: simple without looking boring, with the vermilion touches adding a lot of character. It's English interfaces despite the Japanese aesthetic, and it's really a demonstration of the craft side of building polished interfaces — the same care that shows up in the products we covered all episode.

Mia: So that's the briefing. Local-first tools that keep your library, meetings, and incident records on your own devices. The hybrid middle ground where Perplexity and GhostWriter split work between cloud power and local privacy. Cheaper, sharper local-class models like SWE-2 pushing the cost frontier. The new visibility game with Visiby and Neopress. And the demo layer — DemoTV and ScreenCursor — plus the small tools, Clipwise and SHIUI, that treat restraint as a feature.

Milo: Everything we said about capabilities and results comes from the makers and the discussion — treat the claims as claims, and the open questions we flagged as the ones to ask before you commit. Thanks for listening — we'll see you at the next briefing.

Mia: Bye everyone.