1019 | Tech Weekly: On-Call AI, Local Tools and Pocket Games

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

Milo: And I'm Milo. Today we've got a bundle of stories that actually connect in a pretty satisfying way — they're all about tools that do impressive things while making some kind of promise about trust. Trust in AI that touches your infrastructure, trust in tools that handle your secrets, trust in apps that handle your words, your meetings, your memory.

Mia: Yeah, that's the thread I'd pull on. There's an AI that does on-call at 3AM but won't fix anything without a human saying yes. There's a command-line tool that masks secrets when coding agents go poking around. And then later in the episode we'll get into creative AI getting dirt cheap, wearables that remember your day, a new game from the Vampire Survivors studio, and a whole suite of local-first Mac tools.

Milo: So let's start at the top, literally at 3AM, with this on-call AI.

Mia: Right. So the pitch here is an AI that's on guard duty for your systems overnight. When something breaks at 3AM, it goes and finds the root cause — and the claim is that it finds a proven root cause, live, not a guess. And then here's the part that I think is the actual product: it can execute a fix, but only with human approval.

Milo: That approval gate is doing so much work in that description. Because the nightmare scenario everyone has with autonomous ops AI is exactly this — it's 3AM, nobody's watching, and the agent decides to restart the database cluster to "fix" things.

Mia: Exactly. So the design answer is: the AI does the diagnosis, which is the tedious part at 3AM when a human engineer is groggy and clicking through dashboards, but the human keeps the trigger. That's a sensible division of labor. The other two pillars of the pitch are deployment-related: it runs on-prem and air-gapped.

Milo: Which tells you exactly who they're selling to. On-prem and air-gapped means regulated industries, companies with serious compliance requirements, places where "send your logs to our cloud" is a non-starter. That's a deliberate positioning choice — they're not competing on convenience, they're competing on "your data never leaves your building."

Mia: And pricing-wise, there's a 90-day free trial. Which is a long evaluation window, and that actually makes sense for this category, because nobody's going to trust an on-call AI in a weekend demo. You need to let it sit through real incidents before you believe its root-cause analysis.

Milo: Right, and that's where I want to plant the flag of skepticism — in a useful way. Everything I just described is the maker's claim. "Proven root cause" is a strong phrase. What does "proven" mean operationally? Is it causal reasoning with evidence attached, or is it a confident-sounding diagnosis? We don't know the false-positive rate in real-world use, and for an on-call tool, that's arguably the single most important number.

Mia: Yeah, because a false positive at 3AM isn't just a wrong answer, it's a human who got woken up unnecessarily. Do that a few times and the tool loses trust permanently. And the flip side is false negatives — it says "all clear" when something's actually wrong.

Milo: So who's this for? Honestly, it's for teams that already have mature incident processes and want to shorten the time between "pager fires" and "we know what's wrong." The value proposition isn't replacing the engineer; it's replacing the first forty-five minutes of bleary-eyed investigation.

Mia: Let's stay in that world of trust and AI touching your work, because there's a companion story here about coding agents specifically.

Milo: This is shim-cli. So the context: a lot of developers are now running coding agents — things like Claude Code — that read your codebase and run tools, and the outputs of those tools can contain secrets. API keys, tokens, credentials that happen to be sitting in environment variables or config or test output.

Mia: And the agent doesn't necessarily know what's sensitive. It's just reading text.

Milo: Exactly. So shim-cli is a hook and a local proxy — it sits between the agent and the world — and what it does is mask secrets in tool results before they get anywhere. The claims that matter here: it's local, there's no telemetry, and it's open source.

Mia: That trio is the whole pitch. Local means the masking happens on your machine. No telemetry means your code and your secrets aren't phoning home. And open source means you can actually verify those first two claims instead of taking them on faith. That's a real difference from a lot of "trust us" security products.

Milo: And it's not limited to Claude Code — the description says it hooks into Claude Code "and others," so it's positioned as infrastructure for the agent era generally, not a single-vendor add-on.

Mia: I like putting these two stories side by side, the 3AM AI and shim-cli, because they're solving the same problem from opposite ends. The on-call AI says "the AI can act, but a human approves the action." Shim says "the AI can read, but the sensitive parts get hidden first."

Milo: One is about constraining what the agent does; the other is about constraining what the agent sees. And both are responses to the same underlying reality, which is that these agents have broad access by default and the industry is scrambling to add guardrails.

Mia: Now, the open questions here: with shim, the big unknown is coverage. How far does the local masking extend to other agents beyond the ones it explicitly hooks? Secret detection is a hard problem — there are so many formats, and things that look like secrets but aren't, and things that are secrets but don't match any pattern.

Milo: Right. A masking tool is only as good as its detection, and false negatives there are quietly catastrophic — a secret slips through once and it's in a log somewhere forever. And the honest answer is we don't have real-world data on how well it catches things across different agents and tool types.

Mia: Which connects to where this whole discussion goes next, and that's the reality that agents read everything. Your code, your configs, your tool output, your notes. So the question "how do developers keep secrets safe when agents read everything" is now a first-order engineering problem, not a nice-to-have. Tools like shim are one answer; policies and scoping are another. It's going to be a busy area.

Milo: And actually, that agent-era framing is a good bridge to our next subject, because version control tools are starting to adapt to the same reality — agents writing code that humans then have to review.

Mia: So, Tower 18. This is a Mac client for version control, and the headline features are about its CLI. You can open views directly from the command line — History, Blame, Working Copy — so instead of clicking around the GUI, you jump straight to the view you need.

Milo: That's a workflow improvement in itself, but the more interesting part for this conversation is the Agent Skill it ships with, which is aimed at reviewing AI-generated work.

Mia: Which is a very 2025 feature. The insight is that reviewing AI code is becoming a distinct activity with its own needs. An agent can produce a plausible-looking change, and the human reviewer needs to interrogate it — see the history, see who changed what, see the state of the working copy — and Tower is saying "here's a skill designed for that review loop."

Milo: It's a small feature but it signals a real shift. Version control tools were built on the assumption that humans write code and other humans review it. Now the writer is often a machine, and the review step is where humans still add the most value, so the tools are reorganizing around review.

Mia: And there's a supporting example in the same space worth mentioning: an open source email client, Inbox Zero Mail. It's AGPL-licensed, works with Gmail and Outlook, has keyboard shortcuts, and includes AI that classifies your mail and drafts responses for twenty dollars a month.

Milo: It's a different domain — email versus version control — but the pattern rhymes: take a high-volume workflow, add AI that handles the triage and the first draft, keep the human in the approval seat. Same shape as the 3AM on-call AI, honestly. Draft, classify, suggest — human decides.

Mia: Yes! Draft-and-approve keeps recurring. And on Tower, the honest unknowns are adoption and imitation. Will developers actually adopt agent-review skills inside GUI clients? And will competing clients copy the pattern? The feature only matters if it fits into how people actually work.

Milo: And what comes next beyond review — that's the open question. Review is the first agent-era feature version control has grown, but if agents are writing most of the code, what else do they need? Better scoping of what an agent is allowed to touch? Branch policies designed for machine authors? We don't know yet.

Mia: Okay, let's change gears — same era, different corner of it. From agents that write and review code to agents that can now generate media. And the through-line is that generation is getting dramatically cheaper and faster.

Milo: First one is Krea 3, which does AI image editing. The claims: native 2K output, an edit in under eight seconds, at six cents per edit. Plus up to eight reference images, and regional editing that preserves the rest of the image.

Mia: Let's unpack those numbers because they're the story. Under eight seconds and six cents means iteration becomes nearly free. When a creative iteration costs a minute and a dollar, you think twice before trying an idea. When it costs eight seconds and a nickel, you just try it. That changes the creative process itself, not just the speed.

Milo: And regional editing is a meaningful capability distinction. It means you can change one part of an image — swap an object, adjust a background area — without the model regenerating and subtly wrecking everything else. Anyone who's used image generators knows the pain of "I asked it to fix the hat and it changed the face."

Mia: Right. And eight references suggests it can work from a set of source images rather than one, which matters for style consistency. Now, who's this for? Working designers and creators who need editing control, not just text-to-image dice rolls. The differentiation versus existing options is exactly that: speed, price per edit, and surgical regional control.

Milo: Then the other half of this story is Popcraft MCP, and this one's aimed squarely at agents. MCP, as in Model Context Protocol — it exposes a single endpoint that an agent can call to generate video, image, music, and sound effects. And under the hood it routes to multiple models — Seedance, Veo, Kling, and others.

Mia: So the shape of this is different from Krea. Krea is a tool a human uses. Popcraft is plumbing an agent uses. One endpoint, many modalities, multiple underlying models. That's an abstraction layer — the agent doesn't need to know the differences between the video models; it asks for "a video" and Popcraft picks.

Milo: There are 100 free credits to start, so there's a low-friction on-ramp. And the reason this matters is the direction of travel: media generation isn't just getting cheaper and faster, it's becoming callable. An agent that can generate text and now also images, video, music, and SFX through one interface can compose whole media outputs, not just assist with pieces.

Mia: Which is exactly where this goes next — agents composing full media pipelines. Script, visuals, audio, assembled end to end. And I want to be careful here, because that's the trajectory the tools point toward, not something demonstrated in these sources.

Milo: Agreed. And the unknowns are significant. Quality at scale — a six-cent edit sounds great until you're reviewing hundreds of them and the error rate matters. And licensing of outputs — who owns what an agent generates through one of these endpoints, and what can you legally use it for — is genuinely unresolved territory. Treat all of these capability claims as the makers' claims.

Mia: Okay, from agents making media to consumer AI — things you wear and things that help you present. And again, trust and privacy are the selling points.

Milo: Amazon's got a line called Bee — two products. The Bee Frontier is a wearable at $74.99, and the Bee Eclipse is $99.99, and instead of a wearable it's a magnet for your phone.

Mia: So two form factors, one idea. The idea is personal-memory AI — the device captures your day and the AI helps you remember it. Conversations, context, what happened when. The Frontier is the thing you wear; the Eclipse turns your existing phone into the capture device via a magnet, so you don't need another gadget on your wrist.

Milo: And the privacy claim attached: audio gets processed and then deleted. Not stored indefinitely, not shipped to a server forever — processed and deleted.

Mia: Which is the right promise for a device that is, functionally, recording your life. But I want to flag the boundary of that claim. The statement covers audio. What about the transcripts, the derived summaries, the memory graph the AI builds from your day? "The audio is deleted" doesn't automatically answer "what happens to everything the audio was turned into." That's the meaningful open question, and it's the one anyone considering these should ask.

Milo: Right, and it's the question for the whole category of memory wearables: they're only useful if they capture a lot, and they're only acceptable if they retain almost nothing. Whether Amazon has threaded that needle, we can't verify from the product description.

Mia: The related story is MyLens, which is AI for presenting. Specifically, it presents live, one-to-one — so think sales calls or personalized walkthroughs — but at scale. It builds new slides in real time from your sources, so the deck you show one prospect is assembled on the fly for that prospect, and it reports on each session afterward.

Milo: That's a genuinely different model from the shared-deck world. Instead of one deck everyone gets, every session gets a deck generated from your source material, tailored to the audience, and you get analytics on how each one went.

Mia: Who's that for? Anyone doing personalized one-to-one presentations at volume — sales, customer success, that kind of role. And it pairs interestingly with the memory theme: both Bee and MyLens are about AI that knows your context — your day, your sources — and turns it into something useful in the moment.

Milo: And then two supporting items round out the consumer-helper theme. First, Tomo in ChatGPT: you point it at any topic or even a chat, and it turns it into a Duolingo-style learning game. Free to start, no registration required.

Mia: So the mechanic is borrowed from language apps — streaks, levels, gamified practice — but applied to anything you want to learn. It's a small thing but a nice example of AI lowering the friction between "I have material" and "I have a study routine for it."

Milo: The other is Latch, which is a self-control tool. It blocks apps on your phone until you scan a physical key — NFC or QR — or literally walk a certain number of steps. No account, no ads, and a Pro tier that's a one-time $9.99.

Mia: I love that the unlock mechanisms are physical. The whole point is adding friction to distractions, so the escape hatch is deliberately inconvenient — you have to get up and walk, or go find a physical key. And the pricing model is refreshing: one-time purchase, not a subscription, and no account needed, which also means minimal data trail.

Milo: Which loops back to our trust theme — even in the consumer space, "no account, no ads, one-time payment" is a trust pitch.

Mia: Now let's get into something I know you'll enjoy — the play-anywhere corner of the episode.

Milo: The Vampire Survivors studio — poncle — has a new game: Vampire Crawlers. And it's a different genre for them. This one is a turn-based deckbuilder — they're calling it "hyper" — with rogue-lite elements. And the availability is the headline for a lot of people: PS5, Xbox, Switch, Steam, and Epic. Essentially everywhere on day one.

Mia: "Hyper" is their word, so treat it as marketing flavor, but the structural description is clear: cards, turns, runs, and presumably builds that evolve as you play, which is where the rogue-lite DNA comes in. What's notable is poncle expanding beyond the survivors-like genre they essentially defined, into deckbuilding, which is a crowded but passionate space.

Milo: And the reason this matters beyond one game: Vampire Survivors showed that a cheap, deeply addictive loop could be a phenomenon. If that studio applies its loop-design sensibility to deckbuilders, that's a real signal for where the genre goes — and other studios will be watching how the market responds.

Mia: Which is honestly the main unknown here. We don't have reviews yet, and availability details beyond the platform list are thin. So: promising premise, big distribution, unproven reception.

Milo: And the natural companion piece is a pocket controller — the Xbox Nano Controller. It's a wireless gamepad that's small enough for a pocket, and the claim that matters is that it has all eighteen standard inputs. So it's tiny, but it's not a stripped-down toy — full button set, Bluetooth, rechargeable battery, $49.99.

Mia: That "all eighteen standard inputs" line is the whole product. The usual compromise with compact controllers is missing inputs — no gyro, truncated triggers, whatever. This one says: full controller, shrunk. And with Vampire Crawlers launching across every platform including handhelds and consoles, you can see how these two stories sit next to each other — games going everywhere, and hardware shrinking to follow.

Milo: $49.99 is mainstream controller pricing, so it's not a premium accessory bet. Whether the ergonomics work at that size — that's the unknown a spec sheet can't answer. Small controllers live or die on how they feel after an hour.

Mia: Okay, last major subject, and it might be my favorite: local-first AI on the Mac, and what happens when your notes become readable by agents.

Milo: There are three products here that make a coherent argument together, so let me take them in turn. First, Openbroca. It's AI dictation for Mac, and the size claim is remarkable: about six and a half megabytes. It removes filler words from your speech — the "ums" and "uhs" — while keeping your tone, it has a personal dictionary so it learns your terms, and it translates to twenty-eight languages.

Mia: Six and a half megabytes is a statement. For comparison, that's smaller than most photos people take. It says the intelligence runs in a genuinely tiny local model rather than calling the cloud.

Milo: Second, SpeechMac, which is local speech-to-text built on Whisper — with a Parakeet model — running entirely on your Mac. Fully offline, audio never leaves the machine, twenty-five languages, one-time purchase.

Mia: So the overlap with Openbroca is obvious — both are Mac, both are voice-to-text, both emphasize local. The difference is the emphasis: Openbroca is about polished dictation output — clean prose, your tone, translations. SpeechMac is about raw fidelity and privacy — offline transcription of anything, audio never leaving the machine, and a one-time price rather than a subscription.

Milo: And they're both part of a larger wave: local models got good enough that "offline, private, on-device" stopped being a compromise and started being a selling point. The open question for both, honestly, is accuracy versus the cloud giants — local models are close now, but on hard accents, noisy rooms, specialized vocabulary, does local hold up? We don't have independent benchmarks here, only claims.

Mia: Third, and this is where it gets interesting: Gravity Notes. It's a note-taking app built around a single list — the description calls it Karpathy-style, referencing Andrej Karpathy's known preference for one flat list of notes instead of folders and hierarchies. One-time payment, notes live on your devices. And — here's the hook — it ships with a CLI so that agents can read and write your notes.

Milo: That CLI is the bridge between two of today's themes. All episode we've been talking about agents that read everything — your code, your tool output. Gravity Notes says: if agents are going to be part of your workflow, your notes should be part of what they can work with. Your agent reads your notes, does something useful, writes back.

Mia: And the design choices reinforce it. A single flat list is the most machine-friendly structure there is — no hierarchy for an agent to misinterpret, just notes. Local storage means the agent accesses data that never left your devices, which is consistent with the privacy posture of everything else in this segment.

Milo: And there are two supporting items that flesh out this local-first Mac world. MeetPaw is a free, open source app that lives in your Mac menu bar and creates or copies meeting links for Meet, Zoom, or Teams with one click.

Mia: Small, single-purpose, open source, and — notably — local in the sense that a meeting-link tool doesn't need to know anything about you. And then there's Latch again, from our consumer segment, which fits here too as a Mac-adjacent utility — actually no, Latch is phone-side. The honest connection is thematic: small, accountable, single-purpose tools.

Milo: Fair — MeetPaw is the Mac-native support item here. And what comes next in this segment is exactly what Gravity Notes points at: the connection between local-first tools and the agent CLI trend. If your dictation, your transcription, and your notes are all local, an agent with local access can orchestrate them — transcribe the meeting, extract the decisions, append them to your notes — all without anything touching the cloud.

Mia: Which would be the full-stack version of the trust story we opened with. The 3AM AI keeps your infrastructure data on-prem and air-gapped; shim keeps your secrets masked locally; the Mac tools keep your words on your machine; Gravity Notes gives agents a doorway into your notes without a doorway into everything else. Same principle at every scale: capability is fine, but the data boundary is the product.

Milo: And the honest caveat across all of it: these are claims from makers, and the open questions are the ones we flagged — false-positive rates, masking coverage, local accuracy, and privacy beyond the audio. The direction is clear, though. The agent era isn't just asking "what can AI do," it's asking "what is AI allowed to see, and who says yes."

Mia: That's a good place to land. Thanks for listening, everyone — I'm Mia.

Milo: And I'm Milo. We'll see you next time.