1005 | The AI Coding Wave Meets the Mac Desktop

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

Milo: And I'm Milo. Today we've got a pile of new launches to get through, and there's a thread running through most of them that I find genuinely interesting.

Mia: Yeah — coding agents are no longer just writing code. They're producing videos in your own voice, and entire documents. And on the other side, people are building the guardrails that make it safe to let agents act at all.

Milo: So we'll start there, with the agents themselves, then move through the infrastructure, docs, some self-hosted stuff, a lot of Mac tools, and end on a lighter note. Let's get into it.

Mia: Alright, first up is Sente, and it's a coding agent you talk to. Not type at — talk to. You speak, it answers.

Milo: And here's the detail that caught my eye: it responds in your own voice. So it's not some synthetic assistant voice, it's cloning or synthesizing your voice for the replies.

Mia: Which makes it feel less like a tool and more like pair programming with someone sitting next to you. You describe a problem out loud, it talks back, you iterate. That's a very different interaction model from typing prompts into a terminal.

Milo: Two more things worth knowing. One account gets you access to 404 models — so you're not locked into a single vendor's model, you can pick per task. And the whole thing is MIT open-source, so you can self-host it or audit what it's doing.

Mia: I think the open-source part matters a lot here, because a voice agent that listens to your environment is a trust-heavy product. If you can read the code, that changes the calculus.

Milo: The open question for me is practical: how good is the voice round-trip actually in a coding session? Voice is great for describing intent, but for anything precise — a variable name, an API path — people usually end up typing anyway. The makers present the voice interface as the interface; whether it holds up during real debugging is something we can't verify from the description.

Mia: Right, and we should be clear: everything we're saying about these products comes from how the makers describe them. We haven't tested them. Okay, second one in this theme, and it's a bigger leap: LaunchReel.

Milo: LaunchReel is a set of Claude Code plugins that generate videos from your repository — as code. So you point it at a repo and it produces product videos, even talking-head videos, and the video is treated like a code artifact.

Mia: Which means version control, editing, regeneration — all the things you get with code. And then there's a Studio component where you can make pin-point fixes and export up to 4K.

Milo: The idea here is that a lot of teams need product demos and launch videos, and those are usually hand-made, expensive, and instantly stale the moment the product changes. If the video is generated from the repo, in theory it stays in sync with the product.

Mia: That's the compelling part. The skeptical part is: talking-head videos generated from code — how convincing are those? A polished product demo is plausible; a talking head that doesn't feel uncanny is a much harder generation problem. The maker describes the capability, but video quality is exactly the kind of claim that needs eyes-on verification.

Milo: Agreed. But the direction is what's interesting. Between Sente and LaunchReel, the pattern is the same: agents aren't just emitting code anymore, they're emitting the artifacts around the code — voice interactions, videos, documentation.

Mia: And once agents are producing things that ship, you immediately hit the next question, which is: how do you trust any of it? Which brings us to CoreSpeed.

Milo: CoreSpeed is infrastructure for agents — it's an MCP platform that sits between your agents and the apps they need to touch. It handles app connections, multiple accounts, memory, budgets, and approval controls, all in one place.

Mia: The single most important design decision, though, is this: credentials never pass to the agent. The agent never actually holds your API keys or logins. CoreSpeed brokers the access.

Milo: And that targets the biggest trust gap in agent tooling right now. If you give an agent your credentials, you're trusting it fully — and everything downstream of a leak. If the agent instead asks a broker for access, and the broker enforces budgets and approvals, you've got a control point.

Mia: So instead of "the agent has the keys to everything," you get "the agent can request actions, and the platform decides what's allowed and how much it costs." Budgets are a nice touch too — agents can burn through API spend fast, and having a hard cap is the kind of boring feature that prevents real disasters.

Milo: Approval flows mean you can require a human sign-off before an agent does anything consequential. That's the pattern that keeps showing up in this space: autonomy, but with a human checkpoint where it matters.

Mia: Okay, next one stays on the theme of "verify before you trust": Aperture. It's an open-source AI code editor, and its interesting piece is something called Composer.

Milo: Composer stages diffs — so when the AI proposes changes, you don't just get a wall of code. The diff is pre-checked before it reaches you: it verifies parsing, imports, types, and it runs previews and tests, all sandboxed in a browser tab.

Mia: That's a big deal because the weakest link in AI coding today is review. Humans rubber-stamp AI-generated diffs because reviewing them properly is tedious. If the mechanical checks — does it parse, do the imports resolve, do the types hold, do the tests pass — happen automatically before review, then the human review is about intent, not syntax.

Milo: And being sandboxed in a browser tab means a broken or malicious change can't touch your actual machine. Plus it's open-source, so the checking logic itself is inspectable.

Mia: Between CoreSpeed and Aperture, you can see the same philosophy from two angles: one constrains what the agent can do in the world, the other verifies what the agent produces before a human ever looks at it.

Milo: Third one in this group: Octri.dev. This one starts from an OpenAPI spec and generates documentation, SDKs in ten languages, and an MCP server.

Mia: And there's a clever feedback loop built in: the generated SDKs self-report production errors, and those errors are tied back to the specific release. So when your API consumers hit a problem, you learn which SDK version and which release it correlates with.

Milo: That's the kind of thing that usually requires instrumentation you never get around to building. Whether the self-reporting is low-noise in practice — that's an open question. Error telemetry can drown you if it's not filtered well.

Mia: Right. There are also a couple of adjacent tools worth mentioning that fit this same conversation. Thinking Orbs is a React component library for showing AI agent states — thinking, working, done, that kind of thing — eight states across fifteen variants, zero dependencies, about 3.8KB, MIT licensed.

Milo: Tiny, but it speaks to a real UX problem: when an agent is running, users need to see what it's doing, and everyone keeps rebuilding those little animations. Also in the same orbit is Rival Workshop — it displays your repo's skills as books on a shelf, and lets you manage and rewrite skills for Claude Code, Codex, and Cursor. It ships 150 skills and is priced at $79 until October 31, 2026.

Mia: So the picture across this whole stretch: agents that act, agents that produce, and a growing layer of tools to make their actions controlled, their output verified, and their states visible. Which leads naturally to the next problem — if agents are going to maintain documents, the documents themselves need to change.

Milo: Let's talk about docs. First one: WikiFix. It uses AI to scan a Confluence wiki for contradictions, duplicates, and orphaned pages.

Mia: And the headline evidence here is a real run: it found 57 conflicts across NASA's 732 documents. That's a concrete, checkable claim — a specific number against a specific corpus — which is more than most launches give you.

Milo: Now, I'd still treat that as a maker-reported result, not an independent audit. But even taking it at face value, it tells you something: a well-run org with hundreds of docs has dozens of internal contradictions nobody noticed. That's the actual problem WikiFix is aimed at.

Mia: And importantly, the workflow keeps humans in the loop — the AI finds the problems, proposes fixes, and the user approves what gets applied. It's not auto-editing your wiki.

Milo: The second one, DocsAlot MCP, attacks docs from a different direction. It's an MCP server, so Claude or Cursor can create, update, and publish a help center directly from your coding environment.

Mia: So writing docs becomes an agent task. And there's a second capability that's arguably more interesting: it analyzes reader questions to find documentation gaps. So instead of guessing what to document next, you look at what people are actually asking.

Milo: That closes a loop: readers ask questions, the tool spots where the docs failed to answer them, the agent drafts the missing section, it gets published. Docs stop being a static artifact and become a maintained system.

Mia: The obvious caveat: agent-maintained docs are only as good as the agent's understanding of the product. If the agent misreads the codebase, it'll confidently publish wrong documentation. The human approval step in WikiFix is the right instinct; you'd want the same discipline with DocsAlot.

Milo: Both of these are small-team products, so longevity and depth of the analysis are fair questions. But the direction — docs as a living, agent-maintained surface — seems likely to stick.

Mia: Now, a lot of the tooling we've talked about — MCP servers, self-hosted agents, open-source editors — implies people want to run things themselves. So let's talk about self-hosting proper. First: opens.cc.

Milo: opens.cc is a self-hosted email delivery platform. It's free under Apache-2.0, it's API-compatible with Resend — meaning if you've built against Resend, you can swap it in — and it sends through AWS SES.

Mia: That's a smart compatibility choice. Resend is popular, so matching its API lowers the switching cost to basically changing a base URL. And riding on AWS SES means you inherit SES's deliverability infrastructure instead of building your own.

Milo: The trade-off, as always with self-hosted email, is that you own the operational burden — SES setup, bounce handling, monitoring. The maker's bet is that control plus no recurring platform fee beats convenience, and for teams already on AWS, that bet is pretty reasonable.

Mia: Next: Quven, a self-hosted media server. The distinguishing choice here is native-app-first direct playback — so instead of everything getting transcoded on the fly, native apps play files directly when the format is supported.

Milo: Direct playback matters because transcoding is the expensive part of running a media server. If your devices can play the file as-is, your server just serves bytes, which means it runs fine on modest hardware. Transcoding is still supported for when it's needed, and it's free for personal use.

Mia: This is clearly positioned against the established players in that space, and the differentiator is that architecture decision. Whether the native app coverage is good enough across platforms — that's what would decide it in practice, and that's hard to judge from a description.

Milo: One more in the self-hosted, low-ops family — and this one is charming. Calnio syncs Notion and Apple Calendar bidirectionally, and it writes real iCloud events, not some parallel calendar thing. It's free, made by a solo developer, and runs on a Raspberry Pi.

Mia: I love that as a spec sheet. A real sync between Notion databases and Apple Calendar, actual iCloud events so they show up everywhere natively, running on a $50 computer. It's the "small tool, exactly one job, costs nothing" school of software.

Milo: The known limits are the flip side: it's one person's project on a Raspberry Pi, so reliability under heavy load and long-term maintenance are open questions. But for personal use, that's often completely fine.

Mia: Staying with open source and low-friction for a moment — Blume 2.0 is a zero-config documentation framework. You drop Markdown files into a structure and you get a production-grade site.

Milo: The "zero-config" claim is the whole pitch: most docs frameworks ask you to write config files, pick themes, wire up builds. Blume's bet is that the file layout itself should be the configuration. And there's a CLI migration command — one command to move over from an existing setup.

Mia: That one-command migration is a recognition that the real cost of switching tools isn't the new tool, it's leaving the old one. Lower the exit cost from the competitor and switching becomes rational.

Milo: Blume and Calnio together sketch a philosophy: open source, minimal configuration, run it on whatever hardware you have. No subscriptions, no dashboards to log into.

Mia: Which is a nice segue to the Mac itself, because the next group is all about reducing friction on the desktop. First: Snapset.

Milo: Snapset saves your entire workspace — which apps are open, which windows, which tabs, which folders — and restores it, per display. So if you switch between, say, a coding setup and a writing setup, or you dock and undock a laptop, you get the whole context back, not just the app list.

Mia: Pricing: 14-day trial, then Core at $12 a year or Max at $30 a year. That's subscription, but at a low annual number — this is a utility you'd evaluate on whether the restore is truly reliable, because a workspace restore tool that half-restores is worse than nothing.

Milo: Exactly — correctness is everything here. Windows management on macOS is fiddly, and edge cases are where these tools live or die.

Mia: Second Mac tool: Smooth Recorder. Screen recording for Mac, and the features are aimed at people recording tutorials and demos: automatic zoom on clicks, cursor smoothing, and captions generated on-device.

Milo: On-device captions is a nice detail — no upload, no per-minute transcription fees, and privacy stays intact. And it's a one-time $9 in early access, which is a very low bar to try it.

Mia: Both Snapset and Smooth Recorder are about the same thing from different ends: Smooth Recorder reduces the friction of producing work you show to others; Snapset reduces the friction of resuming work you stopped. Both are one-time or cheap-annual pricing, aimed at individual Mac users rather than teams.

Milo: And that brings us to the most visually interesting group: the desktop as a living surface. Three products here, and they're all treating your wallpaper and your notch as things that can be alive.

Mia: Start with Pixel Soup. It's a Mac app that generates dynamic wallpapers live on the GPU — the wallpaper isn't a static image, it's being rendered continuously. And the wallpapers are described as day-shortened, meaning they shift with the time of day.

Milo: GPU rendering for a wallpaper sounds extravagant, but it's actually the right tool — shaders are cheap for this kind of generative visuals and the GPU is otherwise idle. Pricing is a buyout model that's moving weekly: it's been $4.99, going to $9.99, with a 3-day free trial. That escalating price is a launch tactic to reward early buyers.

Mia: Next: NotchMac. It puts a character called Notchy inside the MacBook's notch, and that character surfaces your notifications, music controls, calendar — the notch becomes a little living status area.

Milo: This is part of a small genre of notch-utilization apps, and NotchMac's angle is personality — a character rather than a passive panel. Pricing is $14.99 one-time or $3.99 a month, so they're testing both models. Which one survives will tell you something about how people value it.

Mia: And third: Art4, which turns your desktop into an art museum. It ships 3,804 artworks, each with a story attached, as wallpapers. $14 one-time, 7-day trial, requires macOS 14 or later.

Milo: The stories are the differentiator — it's not just a wallpaper pack, it's curated context. If you're going to stare at your desktop all day, having a painting with a paragraph about why it matters is a genuinely different experience from a stock photo.

Mia: And there are two smaller pieces that round out this desktop-canvas theme. TinyFolder makes Dock folders for Mac where the folder icon is auto-generated from the apps inside — a tiny bit of visual sense-making. Note there's an App Store version limited to one folder, and a Pro version with a buyout license.

Milo: And Control My Mac — which actually bridges out of the desktop — turns an iPhone or iPad into a trackpad, keyboard, and shortcut deck for your Mac. Pairing is by QR code, and there's no server involved, so it works locally.

Mia: The no-server detail matters for both privacy and for it just working anywhere, including offline networks.

Milo: So the through-line of this whole desktop section: the Mac's static surfaces — wallpaper, notch, Dock — are being turned into dynamic, personalized, sometimes playful ones. And Control My Mac extends the Mac beyond its own hardware.

Mia: Which is a good pivot, because the next group is about working across devices and shared surfaces — and it includes two things with an AI angle. We've mentioned Control My Mac; let's give the next two their due.

Milo: Sellio is AI customer support built around a shared inbox. It unifies live chat, WhatsApp, Instagram, and other channels into one place, with an AI agent handling conversations.

Mia: Pricing structure is worth noting: the AI agent has a free tier, plus a credit system beyond it. That's a sensible pattern for support AI, because message volume varies wildly — a flat fee either overcharges quiet months or underprices busy ones.

Milo: The real evaluation question for any shared-inbox AI is handoff quality: when the AI can't handle something, does the conversation move to a human with full context, or does the customer repeat themselves? That's not answered by the description.

Mia: Then ChatGPT Space — this is a team shared canvas. You gather files and links into one place, and you can tag ChatGPT or Codex directly into the page to co-edit with the team.

Milo: So the AI isn't a separate chat window — it's a participant on the same canvas your team is working on. Tag it, it contributes, the humans see and shape the result. That's the "AI moves into shared everyday surfaces" pattern: not a destination you visit, but a collaborator already where the work happens.

Mia: Between Sellio and ChatGPT Space, the model is consistent: AI embedded into surfaces teams already use — the support inbox, the shared doc — rather than another tab to switch to.

Milo: Okay, time for the lighter side of the list. Three fun ones. First: Capybara Court.

Mia: I've been waiting for this one. Capybara Court: a capybara judge arbitrates small disputes between two people. Each person states their side, the capybara rules — and here's the design that makes it clever: the answers are private to each party, so nobody loses face publicly.

Milo: And two more design details that show real thought. Blame is capped at 25 to 75 percent — the judge can't say one person is 100 percent at fault. And the whole thing erases itself after 30 days.

Mia: That cap is quietly brilliant. Most disputes between friends or roommates aren't 100/0, and a system that structurally can't declare a total winner changes how people engage with the verdict.

Milo: The 30-day deletion is the same instinct — the dispute record doesn't become a permanent grudge artifact. It's a toy, but it's a toy with an unusually thoughtful spec.

Mia: Next: Poddle. Browser-based pickleball, and the controller is the fun part — you can use your phone as a motion-sensing paddle, or a single AirPod, via an open-source Mac helper, and it can track movement through your webcam.

Milo: Playing racquet sports by physically moving a phone is the kind of thing that's obvious in hindsight. Using an AirPod as the motion sensor is the surprising part — one earbud's motion data standing in for a swing. And it's free, which removes every barrier to trying it.

Mia: How well one AirPod approximates a paddle swing — that's an empirical question, and the webcam tracking will have its own accuracy limits. But as an experiment in repurposing hardware you already own, it's delightful.

Milo: Last one: FlexChords. It takes a YouTube video or audio file and produces synchronized guitar chords for it, using machine learning plus music-theory smoothing.

Mia: The music-theory smoothing is the smart bit — raw ML chord detection is noisy, flickering between similar chords. Applying theory to snap the output to musically plausible progressions is what makes it usable. There's a timeline editor for corrections, and you get 5 free custom scans a month.

Milo: The 5-per-month limit tells you the custom scans are the costly part computationally — the model limits free use where the cost actually is. For learners trying to play along with songs, that's a real workflow.

Mia: Alright, two closing items to wrap the episode. First: Pass Designer — and this one is from Apple itself.

Milo: Pass Designer is Apple's official visual design tool for Wallet passes, currently in beta. The key feature: live preview uses the same rendering engine as iOS and watchOS, so what you design is what users actually see.

Mia: Anyone who's designed Wallet passes knows the pain — you fiddle with JSON, push to a device, squint. A WYSIWYG tool from the platform owner, with true-to-device rendering, removes a whole category of guesswork. It's beta, so expect rough edges, but the provenance — official Apple tooling — makes this one worth watching.

Milo: And our final item: Eat Train Feel. A free web app for logging lifting, meals, and how you feel day to day. It has a coach mode for managing clients, Garmin sync, and the data is encrypted in the browser.

Mia: "How you feel" being a first-class logged metric alongside sets and meals is a nice design choice — training outcomes aren't just numbers. Browser-side encryption means even the service can't read your logs, which is the right default for health-adjacent data. And coach mode gives it a two-sided use: individuals, plus trainers managing people.

Milo: One pricing observation to close on: across the smaller tools today — Sellio with its free tier plus credits, Snapset at $12 a year, Smooth Recorder at $9 one-time, Eat Train Feel free — there's a consistent pattern of very low-friction entry. These aren't enterprise prices; they're "just try it" prices, and the free tiers do the selling.

Mia: That's actually a fair summary of the whole episode. Voice and video agents making new kinds of artifacts, an infrastructure layer making them trustworthy, docs maintained by machines but approved by humans, and a long tail of small, cheap, opinionated tools for Macs, self-hosting, and fun.

Milo: As always, everything we covered today is based on how the makers describe their products — treat it as claims worth investigating, not proven results. The evidence varies: NASA's 57 conflicts is a concrete data point, but plenty of what we discussed is promise, not measurement.

Mia: Thanks for listening — we'll see you next time.

Milo: Bye everyone.