1109 | Mac Notches, AI Agents, and the Tools That Power Them

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

Milo: And I'm Milo. Today we've got a really nice spread — a bunch of small tools, a couple of big model stories, and honestly a thread that runs through almost all of it: making the computer work the way you want it to, whether that's the dead space around your MacBook notch, or an AI agent that gets its own machine to run on.

Mia: Right. And the way we're going to do this is kind of like a daily briefing. We're looking at a batch of recent launches, and we're not just going to list them off — we picked a handful where the problem is clear, the solution is actually differentiated, and there's real substance: concrete capabilities, pricing, and things you can actually check. And where things are murky or unverified, we'll say so, because a lot of what we know here comes from the makers describing their own products.

Mia: Those are claims, not verified results.

Milo: Which is the honest way to do it. Okay, so let's start somewhere fun — the MacBook notch.

Mia: Yes! The notch. That little black cutout at the top of the screen that Apple gave us whether we wanted it or not. For years it was just… there. Dead space. And now there's a whole little ecosystem of apps competing to turn it into something useful.

Milo: So the baseline here is NotchIsland. And the thing that stands out about it is that it's free and open source — MIT license — written in pure Swift. It turns the notch into a Dynamic Island, like the iPhone thing. You hover over it and get a readout, you can expand it into a shade, and it has a launcher that you summon with Option-Space.

Mia: That's the floor, basically. If you just want the notch to do something without paying anything, that's your option. But then there are two paid takes, and they represent genuinely different bets about what the notch is for.

Milo: NotchX is the first one. It's a 14-day trial, then $9.99 one-time. And their pitch is that the notch should be a live hub. So music, messages, calls, meetings, weather — it's ambient information flowing past that corner of your screen.

Mia: And then Archie, which is also $9.99 one-time but with a 3-day trial, takes a completely different angle. Archie turns the notch into a screenshot shelf. So you take screenshots and they stack up there, and there's a quick editor built in — including text blur, which is honestly a feature I wish every screenshot tool had, because blurring out someone's email or an API key is such a common chore.

Milo: And it does local OCR too, so text in those screenshots is presumably searchable or extractable, and it happens on-device.

Mia: Right. So think about the scope difference. NotchIsland is a free island — it's about the notch as a status surface and a launcher. NotchX says no, the notch is a hub for everything happening on your machine. Archie says the notch is a workspace, specifically for screenshots. Those are three different answers to the same question, and they're not really in conflict — you'd pick based on what you actually want up there.

Milo: One thing worth saying: all three live in exactly the same physical space, so I don't think you'd run them together. It's a one-slot battle.

Mia: A one-slot battle. I love that. Okay, let's move from the top of the screen to the rest of the machine, because there's a cluster of Mac-native essentials here that all make the same argument in different ways: the built-in tools are fine, but there's room to do better.

Milo: The biggest swing of the group is Tilde. It's a native Finder replacement — not an overlay, a replacement — and it's $29.99 one-time. The headline features are split views, custom file properties, and saved workspaces.

Mia: Saved workspaces is the one that gets me. If you've ever had that moment where you're working across five folders and you close the window and lose your whole arrangement — that's the pain it's going after. And custom file properties means you can organize by things Finder doesn't natively know about.

Milo: Now, replacing Finder is a big claim. Finder is deeply wired into macOS, and any replacement has to earn daily-driver status. That's the open question: does it hold up as your only file manager, or is it a second tool you reach for? The pricing at $29.99 one-time suggests they're positioning it as a serious utility, not a toy.

Mia: And next to Tilde there's Zigzag Browser, which makes an argument with numbers. It's an open-source, native Mac browser built on WebKit, and it's 18 megabytes. They compare that to Arc at 871 megabytes.

Milo: That's not a rounding error. That's roughly 48 times smaller.

Mia: Right, and the point isn't just smallness for its own sake — being native and lean is the argument. But Zigzag isn't a minimal toy. It has spaces, split view, it supports Chrome extensions, and it can import from Arc.

Milo: The Arc import is smart. Arc's situation has been… let's say unsettled, and there are a lot of people who invested real effort into their browser setup. Giving them a migration path is a real feature, not a checkbox.

Mia: And being on WebKit means it's not another Chromium fork, which matters if you care about browser diversity on the Mac — Safari's engine is right there, and Zigzag is building on it rather than shipping another copy of Chrome.

Milo: Then we get to two smaller utilities. Textly is $14.99 one-time, and it solves a genuinely annoying problem: text you can't select. Images, PDFs, whatever — you can see the text, you can't copy it. Textly uses Apple Vision to capture that text, including structured tables, and exports to CSV, Markdown, or TSV.

Mia: The structured tables part is the differentiator. Plenty of tools do OCR-to-text. Getting a table out of an image as actual CSV that your spreadsheet can eat — that's the hard, valuable version of the problem.

Milo: And the last one in this cluster is free and open source, and it's called Keypause. The scenario: you want to clean your keyboard while the machine is running. Keypause locks all the keys so you can wipe things down without triggering chaos, but your mouse and trackpad keep working so you can still see what's on screen, and it's timed.

Mia: It's a tiny idea, but it's exactly the kind of thing where the built-in options are terrible. Lock screen? Then you can't see anything. Keypause is the obvious-in-hindsight tool. And notice the pattern across this whole group — Tilde, Zigzag, Textly, Keypause — they're all one-time purchases or free, all native, all going after a specific friction point in the day-to-day Mac experience rather than trying to be platforms.

Milo: Which is a nice bridge, actually, because the next group is about the opposite end of the spectrum: agents. And the interesting thing there is that agents are getting the same treatment — bodies, hardware, workflows.

Mia: So Agent Clan is a free 3D app that lets you control Claude Code and Codex agents using hand gestures and voice.

Milo: Hand gestures, in 3D. And here's the detail that makes it concrete rather than gimmicky: permission approvals. When an agent wants to do something that needs your sign-off, you pull right or left to approve or reject. That's a gesture-based permission flow.

Mia: That's actually clever, because permission approval is the bottleneck in agentic coding. The agent stops, asks, you click a dialog. If that becomes a gesture — pull right, it goes — the interaction cost drops. Whether that's actually nicer in practice than a keyboard shortcut, I don't know. That's an open question. But the concept is real: they're mapping a workflow, not just making a demo.

Milo: Then NoMac takes a totally different approach to the same problem. Instead of giving agents a new interface, they give agents the machine itself. It's an on-demand cloud Mac for AI agents to build iOS apps — $0.90 an hour, billed per minute — and here's the detail that matters: the Mac is deleted after each session.

Mia: Ephemeral hardware. That's a security story as much as a pricing story. If you're letting an agent loose to build and run iOS code, doing it on a disposable machine that gets wiped means your actual Mac isn't in the blast radius. And per-minute billing means a short build session doesn't cost you a full hour.

Milo: $0.90 an hour — a two-hour build session is under two bucks. Compare that to the cost of a developer's time babysitting a local setup. The claim from the maker is that this is for AI agents specifically, so the whole point is that you don't have to own or maintain the build machine.

Mia: And then the third piece of this trio is AXL, which zooms out to the workflow level. It's an AI-native sales and marketing platform, and the thing that makes it interesting is its MCP server — Model Context Protocol — which exposes over 1,200 actions to Claude, ChatGPT, and Cursor.

Milo: Twelve hundred actions. So instead of an agent that can do one narrow thing, the idea is that your sales and marketing stack becomes a giant tool surface the agent can operate. There's a free plan with 500 contacts to get started.

Mia: So look at the shape of this group together. Agent Clan is the body — gestures and voice, the physical interface. NoMac is the hardware — a real machine, rented, disposable. AXL is the workflow — a thousand-plus actions in a business domain. Agents went from chat windows to having hands, machines, and job descriptions.

Milo: That's a good framing. And it connects naturally to the big model stories, because agents are only as capable as the models underneath them. And there are two model stories here that are quite different from each other.

Mia: The first one is Thinking Machines — that's the company — with a model called Inkling 2 Small. The specs: it's a mixture-of-experts model, 594 billion parameters total, but only 17 billion active at any time. That MoE architecture is the whole point — you get the capacity of a huge model with the running cost closer to a small one.

Milo: And what's it for? Coding, vision, and agents — so a multimodal, agentic model. And the really notable part: they're planning to release open weights under Apache 2.0. Soon, they say — which is a promise, not a fact yet. But if it lands, that's a permissive license, meaning commercial use is on the table.

Mia: Right, and I want to be careful here — "soon" is doing a lot of work in that sentence. We know the architecture, we know the intended use cases, and we know the licensing plan. We don't have benchmarks in front of us, we don't have independent evaluation. Treat it as a claim.

Milo: Now the second model story is more immediately experienceable. World Labs has a world model called Atlas, and there's a browser demo called Frameboy.

Mia: And the way Frameboy works: there are curated scenes, and you move the camera through them. The views get streamed to you at up to 15 frames per second.

Milo: So this is not a video you're watching — it's a scene the model is generating, and you're choosing where to look. That's the difference between watching a recording and being in a space. The 15 fps ceiling tells you where the technology is right now: it works, but it's not fluid enough to feel like a game engine yet.

Mia: And "curated scenes" is an important qualifier too — you're moving through environments the team has prepared, not generating arbitrary worlds from a prompt in the demo. But as a demonstration of what a world model can do — hold a coherent space that responds to camera movement — it's a genuinely interesting thing you can go try right now in a browser.

Milo: Which is a nice contrast with Inkling. One is a massive model headed for open weights that you can't touch yet; the other is a smaller experience you can try in five minutes. Both are claims about where models are going — one toward agentic capability, one toward generating environments.

Mia: Okay, next cluster — and this one is about trust, specifically trust with words and data. Three products, all attacking the same core issue: getting reliable, grounded output out of language systems.

Milo: DocsAlot is a Slack and Discord bot that answers customer questions from your knowledge base. And the two features that define it are exactly the trust features: it admits when something isn't documented, and it cites its sources.

Mia: The admitting part is the hard one. Most bots will happily hallucinate a plausible answer, and the failure is invisible until a customer acts on it. A bot that says "I don't have that documented, here's what I do have" is changing the failure mode from confident wrongness to honest gaps. And citations mean your support team can audit what it said.

Milo: That's a design choice about epistemics, basically. Whether it's accurate in practice — whether it correctly recognizes its own gaps — that's the thing you'd want to test with your own knowledge base before trusting it.

Mia: Then Sage2 is the API version of the same idea. Their APIs extract structured data from natural language — dates, money amounts, durations — and two things come with the output: source links, and streaming decisions.

Milo: Streaming decisions is interesting — it suggests you can watch the extraction reasoning happen as it goes, rather than getting a black-box answer. And source links mean every extracted date or dollar figure can be traced back to where it came from in the input.

Mia: Which is what you need if you're piping this into anything serious. Structured extraction without provenance is a liability; with provenance, it's auditable. So DocsAlot and Sage2 are the same philosophy — grounded output, cited output, honest output — applied at two different layers: the customer-facing conversation layer, and the data-processing layer.

Milo: And the third one in this group is very different in flavor: Memorias. It turns spoken life stories into memoir books. You talk, and it produces actual books — EPUB, PDF, DOCX.

Mia: The business model stands out: pay-per-use credits, no subscription. For something you'd realistically use once in a life — you're not going to write your memoir monthly — a subscription would be exactly the wrong model. Credits match the actual shape of the usage.

Milo: And it's the same trust question in a softer form, right? You're handing a system your family's stories and trusting it to shape them into a book. That's a different kind of stakes than customer support, but it's still "words you have to be able to rely on."

Mia: Alright, last cluster, and it's about measuring things — which, honestly, loops back to how we opened this show. Reliability and language, measured.

Milo: First is outages.live. It's free, it aggregates official provider status pages into a live feed, refreshed every few minutes. No login needed. And it has an MCP server, so agents can consume it too.

Mia: The MCP part fits the theme from earlier — if your agents are doing work that depends on third-party services, an agent that can check the status feed before diagnosing a failure is genuinely useful. And the no-login, no-cost thing makes it zero-friction.

Milo: One thing to note: it's aggregating official status pages. So it's only as good as the providers' own reporting — it's a convenience layer over self-reported status, not an independent probe network. That's a fair limitation to keep in mind.

Mia: And the second one is stranger and more fun: Deduplication.tech. It's a site on HackerNoon that scores 719 duplicated tech terms across 14 fields, using live Wikipedia pageviews as the signal.

Milo: So the underlying observation is that tech language is full of duplicates — different names for the same or nearly the same thing — and they're scoring which duplicates are actually being used, based on how much people are viewing the relevant Wikipedia pages.

Mia: It's a lens on how messy tech vocabulary is. And it connects back to everything else we talked about, in a way. Every product today lives or dies on whether people can find it and understand it — and when your category is called three different things, discoverability suffers.

Milo: So let's zoom out and close. What's the through-line of this batch?

Mia: I think it's ownership of your environment. The notch apps, the Finder replacement, the browser, the keyboard lock — that's people rebuilding their desktop to fit them. The agent tools — gestures, rented machines, giant action surfaces — that's people rebuilding how work gets done. And the trust layer — citations, provenance, honest gaps — that's people rebuilding what they'll accept from AI output.

Milo: And the honest caveat across the board: almost everything we covered today is described by its maker. The features are documented, the pricing is concrete, and the availability is stated — but none of it is independently verified here. If any of these solve a problem you actually have, the trials exist for a reason: NotchX gives you 14 days, Archie gives you 3, and half the list is free or open source anyway.

Mia: That's the show. Thanks for listening, everyone — we'll see you next time.

Milo: Bye, everybody.