1025 | Agents, Pixels and Notches: This Week's Releases

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

Milo: And I'm Milo. So, Mia, we've got a batch of launches today that, at first glance, look totally unrelated — an AI password vault, an X analytics server, a 3D CAD generator, some Mac utilities, a CD changer nostalgia app.

Mia: Right, but there are threads running through all of them. A bunch of these are about where AI actually does work now — not just chatting, but holding credentials, pulling real data, even designing parts. And then there's a strong counter-current: local-first, no-cloud, one-time-purchase tools that want nothing to do with AI at all.

Milo: Which is honestly the more interesting tension. Let's dig into the AI side first, because there's one product there that I think says a lot about where things are heading: 1Password for Agents.

Mia: So the pitch is this — AI agents can sign in to services on your behalf without ever seeing your passwords. You confirm the sign-in on your own device, and it's included in the existing Individual and Family plans. No separate enterprise tier, at least as announced.

Milo: That confirmation detail is the whole trust model, isn't it? Because the obvious worry with agentic software is: I've given an agent access to my accounts, and now what does it actually know?

Mia: Exactly. The agent doesn't learn the password itself. It gets to act, but the secret stays in the vault, and a human taps approve on their own hardware. That's a real answer to a real question, and it matters because right now, a lot of people are wiring agents into their accounts by just pasting credentials or API keys into prompts, which is, you know, terrifying.

Milo: And the fact that it ships inside Individual and Family plans is notable. It means this isn't a "contact sales" feature. A regular subscriber could turn it on. That's how a trust layer becomes normal instead of an enterprise checkbox.

Mia: Though I want to be careful here — everything we know about this comes from the makers' own description. We don't have independent evidence of how well it works in practice, which services it covers, what an agent can and can't do once authenticated. Those are open questions.

Milo: Fair. Now let me put this next to the second one, because together they sketch a picture. LinkIntel is an MCP server — that's the protocol that lets AI assistants connect to external tools — and it feeds your assistant two things: your own X-post results, and X's open ranking weights.

Mia: With what they're calling GitHub-commit proof, which I read as: they show you the actual code changes behind the weights they're serving you, so you can verify where the numbers came from.

Milo: Right. And that's interesting because ranking on X has always been this black box. You post something, it flops or it flies, and you have no idea why. X has apparently opened up some of those ranking weights, and LinkIntel is packaging them for your AI assistant to reason over.

Mia: So instead of "the algorithm hates me," you could theoretically ask your assistant, "here are my last twenty posts and here's how the ranking weights treat this kind of content — what's going on?" Whether the assistant can actually give you useful answers from that is the open question, but the data plumbing is real, at least as described.

Milo: It's 39 dollars a month, by the way. Which puts it in "professional tool" territory, not toy territory. And notice what these two products have in common with 1Password for Agents: they're all about what an agent is allowed to know and touch.

Mia: Yes — 1Password governs credentials, LinkIntel governs data access with verifiable provenance. Both are essentially building the scaffolding that lets an agent do real work instead of just producing text. The unknown, which I keep coming back to, is adoption. These trust layers only matter if platforms and users actually pick them up. One password manager doing it is a start; it's not yet a standard.

Milo: Right, and there's a third piece to this puzzle that makes it concrete: design work. zCAD AI takes text or an image and generates editable parametric 3D models — with actual dimensions, and assembly mates, which are the constraints that let parts fit together and move correctly in an assembly.

Mia: And the export is STEP, which is the interchange format real CAD and manufacturing workflows use. That's the detail that tells you they're not aiming at "make a fun 3D print doodle." STEP means this could flow into professional CAD tools and machine shops.

Milo: "Editable" and "parametric" are the load-bearing words there, though. A lot of AI 3D generation gives you a mesh — a dumb blob of triangles you can't meaningfully change. If zCAD really gives you a model where you can change a dimension and the geometry updates, that's a different category of tool.

Mia: Again — claim from the maker, not something we've verified. But if it holds up, the workflow it implies is: describe the part, get a real, editable model with proper dimensions, export it. That's hours of CAD work compressed, at least for simpler components. The open questions are accuracy and how complex a design it can actually handle.

Milo: So the throughline of this whole first stretch: agents moving out of the chat window. One product gives them authenticated access without leaking secrets. One feeds them verifiable data about how content performs. One lets them produce engineering-grade output. Chat was phase one; this looks like phase two.

Mia: And there's even a fourth data point, a small one, that fits. AgentValley shows every Claude Code agent running on your machine — including sub-agents — in the MacBook notch, as little pixel characters, with a "needs you" alarm when one is waiting for your input.

Milo: Which is charming, but it's also answering a genuine problem: once you run multiple agents, you lose track of what's doing what and when one is blocked waiting on you. A pixel guy in your notch blinking at you is a silly solution to a real orchestration problem.

Mia: And it's free and open source, which makes it an easy thing to try. But it also quietly confirms the premise: people are running enough agents in parallel now that they need a status display for them.

Milo: Okay, so from agents that do work for you — let's talk about tools that help you make things yourself, but with AI doing the heavy lifting. First one: myPip.

Mia: So myPip is vibe-coding for games, in the browser. You describe what you want, it builds it, and then — and this is the part that stands out — it goes beyond the usual single-player demo. Multiplayer is included. There's "Hive-coding," which means building together with friends in real time. Controller support. And monetization, so you can actually charge for what you make.

Milo: That's a stack that used to be a studio's worth of infrastructure. Netcode alone is famously hard. If multiplayer is genuinely a checkbox in a browser-based game builder, that's a meaningful claim — and also the one I'd most want to see tested, because "supports multiplayer" can mean a lot of things.

Mia: But the framing I find interesting is that creation here is explicitly social. Hive-coding with friends in real time — that's not a solo person prompting an AI at 2 a.m., that's a group activity, like the way people used to build things together in Minecraft or Roblox, except the building itself is AI-assisted.

Milo: And monetization built in from day one means they're imagining an economy around this, not just toys. Though, same caveat as always: the maker description tells us what features exist, not whether the games are any good or the netcode holds up with real players.

Mia: Right. Now the second one in this group is from a much bigger player: Canva Craft. This is Canva's first in-house image model, which is notable on its own — until now they've presumably relied on third-party models, and now they're building their own.

Milo: The architecture is two-stage: plan, then render. So the model apparently first works out a plan for the image and then renders it, rather than generating pixels in one blind pass.

Mia: The specs they're claiming: 2K output, more than 30 formats, and text rendering in 55 languages. That last one matters — getting legible text inside generated images is a classic weakness of image models, and if it really handles 55 languages, that's aimed squarely at Canva's global user base making posters and social graphics with words on them.

Milo: But the feature that matters most for professionals is editable layers. Most image generators hand you a flat picture — take it or leave it. If Craft gives you something you can take apart and edit layer by layer, that's not just generation, that's the start of a design workflow.

Mia: Which makes sense for Canva specifically. Their whole product is editing. A flat image that you can't touch is a dead end for them; layered output feeds directly into what their users do next.

Milo: And notice the contrast with myPip. Both lower the barrier to making things, but they sit at different ends: myPip is about games and social creation, zero-to-something fast. Craft is about giving people who already design a better generation tool that fits into editing. One expands who can make things; the other upgrades how existing creators work.

Mia: The unknown for Craft is quality over time — one announcement with spec sheets doesn't tell us how it actually compares to the established models, or whether the plan-then-render approach produces consistently better results. That'll only show up in sustained use.

Milo: Okay, so from AI-assisted making, let's pivot — and this is where the counter-current I mentioned comes in. Because the next group of products seems almost deliberately allergic to everything we've been talking about.

Mia: Right. How It Went takes your GPS track, your photos, your video clips, and turns them into a short film. Rendering happens locally — on iPhone, Android, or Mac — and you get an MP4 out, in vertical 9:16 or horizontal 16:9.

Milo: Local rendering is the headline for me. Video rendering in the cloud means uploading your footage, waiting, maybe paying a subscription. This does the whole thing on your device, and outputs a standard MP4 either way. That's a privacy-and-convenience position.

Mia: And think about what the product actually is: it takes the raw residue of a day — where you went, what you saw — and gives it a narrative form. A short film of your hike, your trip, your weekend. That's an emotional product more than a technical one.

Milo: Next in this group is River Writer, and this one wears its philosophy on its sleeve. It's a free, open-source Mac app, MIT license, for writing screenplays, plays, and novels. Files are .fountain or .markdown, versioned with Git.

Mia: And the two-word summary on their positioning is: no cloud, no AI. For a writing tool, in 2025, that's a statement. Fountain is a plain-text format for screenplays, markdown is universal, Git means your revision history lives in a system that will outlive any company.

Milo: Which is the exact opposite bet from Canva Craft, isn't it? Craft says: let the model help you generate. River Writer says: here's a quiet room and files you'll still be able to open in thirty years. Both can be right, for different people.

Mia: Exactly — and River Writer is for the writer who fears the lock-in more than the blank page. The open-source MIT license also means, worst case, someone else can maintain it forever. There's no company that can take it away.

Milo: And then the third one brings the fun back: Penta turns your iPhone or iPad into a 90s five-CD changer for Apple Music.

Mia: With real recorded device sounds — the actual clunk and whir of a CD changer, which is the detail that tells you this is a labor of love, not a gimmick slapped together in a weekend. Someone recorded a real machine.

Milo: It's free with an optional Pro unlock. So the nostalgia is the product, essentially — the ritual of loading discs, the mechanical sounds, the slowness that felt like intention.

Mia: But step back and look at the pattern across all three: How It Went processes locally, River Writer is cloud-free and open-source, Penta is free with an optional unlock. Local-first, no lock-in, one-time-or-free economics. It's a coherent worldview, and it's aimed at a very real fatigue with subscriptions and everything-phoning-home.

Milo: There's even a deeper point in River Writer's choice of Git specifically. Git versioning is opaque to most writers — but if you're a developer who also writes, it's home. It's a tool made by someone who understands a specific audience deeply, rather than a mass-market compromise.

Mia: Agreed. So now let's land on the last group, the Mac utilities, because they share a different theme: taking back the parts of your Mac that the default setup wastes — the notch, the menu bar — and doing it with honest, one-time pricing.

Milo: Start with Shedi, because "honest Mac cleaner" is almost a contradiction in terms after decades of scareware cleaners. Shedi targets developers specifically — it knows about Xcode, Docker, node_modules, that family of things that quietly eat tens of gigabytes.

Mia: And the trust feature is that for every item it wants to delete, it shows you the risk and how you'd get back what you lose. "Delete this cache — low risk, it regenerates." That transparency is the anti-scareware move. Plus there's a paper-to-trash animation, which is a bit of delight.

Milo: One-time purchase, 9.99 dollars. Not a subscription. Again.

Mia: Again! Then there's sideOS, which lives in the notch itself. You hold Control and Option, speak, and it controls your apps. It does dictation in 62 languages, handles meetings, and — key detail — it runs locally or with your own API key. One-time purchase.

Milo: So even the AI product in this group gives you the local/no-cloud escape hatch. You can bring your own key, meaning you know exactly what service you're talking to and you're paying that provider directly, not an intermediary subscription.

Mia: Which puts it in interesting contrast to the agent stuff we opened with. 1Password for Agents is a trust layer on top of cloud services; sideOS is a voice layer that tries to keep as much as possible on your machine. Two different answers to the same question: how do I use this power without giving everything away?

Milo: Then Many Displays, which solves a problem so specific and so annoying that anyone who's had it will buy instantly: windows scattered after your monitors go to sleep, or you unplug and replug a dock. It restores your windows automatically after sleep, unlock, or replug.

Mia: And the genuinely clever bit: it can distinguish between identical monitors, using their UUID and port. So if you have two of the same model, it knows which is which. That's the kind of detail you only build if you've personally suffered the problem.

Milo: Nineteen dollars, one-time. And finally Byter, which is free: a menu bar app that tracks your data volume per network — Wi-Fi, hotspot, cable — with live speed and speedtests, all processed locally, in eleven languages.

Mia: Per-network tracking is the feature there. If you're on a limited hotspot plan, you want to know which network burned your data, not just a global total. And free, with the data staying local, fits the group's ethos.

Milo: So themes across the whole utility stretch: one-time pricing everywhere, local data everywhere, and the physical geography of the Mac — notch, menu bar — being reclaimed as workspace.

Mia: And one open question ties them together: how deep can notch apps like sideOS actually go without operating-system support? The notch is a piece of hardware Apple designed for camera housing, not an app platform. Utilities are building on it anyway, but that's ground that could shift with any macOS update.

Milo: Which is the risk with all of these, honestly — they're building in the gaps Apple leaves. Usually the gaps hold. But it's a structural dependency you don't have with, say, River Writer's plain text files.

Mia: So let's zoom out, because the shape of today is genuinely interesting. On one side: agents getting real credentials, real data, real design output — with early trust scaffolding from 1Password, verifiable provenance from LinkIntel, and even a notch-based status display from AgentValley.

Milo: On the other side: local rendering, Git-versioned plain-text writing, a fake CD changer, and a shelf of one-time-purchase utilities that promise your data never leaves your machine.

Mia: And maybe the most honest read is that both currents are responding to the same thing — people figuring out what they're willing to hand over, and to whom. The agent tools are building trust infrastructure because trust is now the bottleneck. The local tools are building products where the trust question never comes up.

Milo: That's the frame I'd leave everyone with. None of the claims we've covered are independently verified — these are maker descriptions — but the directions they're pointing are real, and worth watching. Especially the open ones: whether agent trust layers get adopted, whether Craft's quality holds up, whether notch apps survive OS updates.

Mia: Exactly. Thanks for listening, everyone — find us wherever you get your podcasts, and we'll see you next time.

Milo: Take care.