0929 | Agents, Screenmates and Slick Mac Apps

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

Milo: And I'm Milo. Today we've got a lineup that actually hangs together surprisingly well — it's all AI, but in very different places in your life: AI that governs coding agents, AI that sits on your screen with you, some really polished little Mac utilities, and then AI pointed at sales and marketing instead of developers.

Mia: Yeah, and the through-line I keep noticing is that AI is moving out of the chat window. It's becoming infrastructure, or a layer that watches what you're doing, or a background worker. So let's start where a lot of this is happening fastest — coding agents.

Milo: Right, so if you've been paying attention, coding agents are everywhere now. People are running these things that write files, run commands, call tools, all semi-autonomously. And once you do that at any scale, two problems show up immediately: cost and control.

Mia: Let's take cost first, because there's a whole cluster of products attacking it. Harness Router is an open-source layer that sits in the middle of your agent's tool calls and decides where to route them. Their claim — and I want to stress this is their claim — is roughly 9.5 times faster and about 294 times cheaper, using something they call Jev.

Milo: A 294 times cheaper claim is the kind of number that should make you raise an eyebrow, right? That's not an incremental improvement. But the shape of the idea is real: when an agent makes hundreds or thousands of tool calls, each one is a decision point. If you can route the trivial ones to a cheap path and only escalate the hard ones, the math changes completely.

Mia: And it being open source matters, because if this routing layer is going to sit in the middle of your entire agent stack, most developers won't accept a black box there.

Milo: Now, the adjacent product is Zerg Router, and it solves a slightly different slice of the same problem. It gives you an OpenAI-compatible endpoint for coding agents — so you can point your existing setup at it without rewriting anything — and then it adds two very practical things: daily budgets per API key, and explicit fallbacks.

Mia: Those sound boring but they're not. Daily budgets per key means a runaway agent can't burn through your whole month's credits overnight. And explicit fallbacks means you decide in advance what happens when a model fails or is unavailable, instead of your agent silently doing something weird.

Milo: They're also running a 14-day trial of DeepSeek, so you can test that cheaper model in the loop without committing.

Mia: So Harness is about the routing decision itself, Zerg is about the budgeting and reliability envelope around it. Both are betting that cost control becomes a first-class concern as agents scale.

Milo: Okay, control in the other sense — governance. VibeDefend, from CybeDefend, governs coding agents at the moment they write code. You give it rules — business rules, security rules — and it enforces them in real time as the agent is writing, not in a review after the fact.

Mia: That timing matters. If you catch the agent before the bad code lands, you've saved the whole review-and-rework cycle. They cite compliance going from 88 percent to 89 percent, and I'll be honest — that's a one-point improvement, and it's a vendor-supplied number.

Milo: Yeah, treat that as a claim, not a verified result. The interesting part is the positioning: "governance at write time" is a genuinely different angle than scanning repositories after agents have touched them. And it's free, so the barrier to trying it is low.

Mia: The open question with all three is the same: these gains are claimed, not demonstrated at scale. Real-world performance when you've got teams, multiple agents, messy codebases — we don't have that evidence yet.

Milo: Then there's GenCode from Genspark, which approaches from the interface side. It bundles a Chat UI and a Terminal UI into one agent, and it runs on Claude, GPT, Gemini, DeepSeek, and others — including open-weight models that they say cost about a tenth of what you'd pay otherwise.

Mia: So that ties back to the cost theme. If open-weight models genuinely deliver at one-tenth the cost, then routing layers, budget endpoints, and model choice all compound — you're stacking savings.

Milo: And there are monitoring tools growing around all of this, which is what mature infrastructure looks like. Vantage is an open-source monitor for coding agents: live cost tracking, approvals for files and commands — so the agent has to ask before it touches something — and alerts if secrets appear in what the agent is doing.

Mia: That secrets alert is not a nice-to-have. Agents that write code and run commands will eventually encounter credentials, and you want a tripwire, not a surprise.

Milo: Also worth mentioning in this same neighborhood: MuM, which is a native macOS Markdown reader — pure AppKit, no web engine, starts in about 0.3 seconds, weighs 1.7 megabytes, MIT licensed. The detail that connects it to our coding-agent discussion: it has CLI hooks so agents can interact with it. Even a reading app is getting an agent interface.

Mia: And Statable Analytics, in the same spirit — it's cookieless web analytics hosted in the EU, and it exposes an MCP server so AI agents can read your analytics or set up reports themselves. Again, not an agent itself, but built so agents can use it.

Milo: Okay, let's shift — same idea of AI watching, but now it's watching *your* screen, not your codebase. This is the copilot-on-your-screen category.

Mia: The first one is Sayble, which is a real-time copilot for phone calls. While you're on a call, it suggests the actual line to say next, in about half a second, and afterwards it writes up a summary and a follow-up email.

Milo: Half a second is the number that matters there. A suggestion that arrives after the pause in conversation is useless — the moment's gone. If it really lands in 0.5 seconds, it's usable in live speech. And the automatic summary plus follow-up email is the part that saves you real time, because that post-call admin work is what people actually skip.

Mia: Then Pip, which is a Mac companion that literally sees your screen and points at the exact spot you need to click. So instead of telling you "open the settings menu," it points at the pixel.

Milo: Pricing tiers: free gets you 30 questions a month, Pro is $15, Max is $69.99. That free tier is a genuine way to find out whether pointing-at-the-screen is actually better than describing, which is the core bet of the product.

Mia: And Arc takes the broadest swing: a free assistant that works over any screen, on Android and Mac. It summarizes what you're looking at, reads it aloud, rewrites it, and automates actions — all without you copying and pasting anything into a chat window.

Milo: That "without copy-paste" is the whole category thesis in four words. The reason we all paste things into chatbots is that the chatbot can't see our screen. These products remove that step.

Mia: The open questions here are privacy and accuracy. These tools need to see your screen — possibly during calls, possibly with sensitive material on it — and if they misread what's on screen, the suggestion is wrong in a way that might be worse than no suggestion. We don't have independent testing on any of that.

Milo: Related, staying on the Mac: Stash. It's not an AI product, but it's in the same "your Mac knows more than you think" space. It gives you hidden controls — sliders along the edge of the screen, dials in the corners, a hidden dock — plus something called Stash Deck, where your iPhone becomes a tactile remote control for those controls.

Mia: Okay, so from screen companions to Mac utilities that keep things tidy. And this batch is unusually charming.

Milo: Start with the charming one: Mochi. It's a Mac mascot — a pink blob — that watches your desktop and files loose items into folders for you.

Mia: And the design constraints are what make it trustworthy: it never deletes anything, every move is undoable, and if you pay $5 for Pro, the blob gets a simulated life, like a little Tamagotchi.

Milo: That "never deletes" guarantee is doing a lot of work. The fear with any auto-organizing tool is that it files something somewhere you can't find it. Undoable moves plus no deletion means the worst case is fully reversible.

Mia: Then Vitals, which is a Mac activity monitor with a specific reframe: it tracks activity per app, not per process. If you've ever opened Activity Monitor and seen ninety anonymous background processes, you understand why that matters — you want to know "this app is the problem," not "process 4472 is spiking."

Milo: It keeps 30 days of history, sends alerts, and — this is the detail I like — it detects forgotten dev servers. Every developer has left a server running overnight at least once. A tool that notices that for you is solving a real, slightly embarrassing problem.

Mia: Pricing: $5 for the first 1,000 licenses. Cheap enough to just buy.

Milo: Third main one: Dina 4.5, which consolidates screen recording, screenshots, and 3D motion capture into a single flow. That combination is usually three separate tools with three separate exports. Free for 7 days, and there's a discount code DINAPH for 10 percent off.

Mia: Quick hits in this same neighborhood — Shotcandy is a free, open-source browser tool for making screenshots look good: frames, arrows, blur, and video up to 4K. MIT licensed, no registration, no watermark. If you've ever paid for a screenshot beautifier or fought a watermark, that's your tool.

Milo: FaveNest is a visual bookmarks library that uses Apple Intelligence for titles, summaries, and tags — all processed on-device. No account required; it syncs through your own iCloud. So the intelligence is local and the data never touches their servers, because there are no servers.

Mia: And Ryu Journal on iOS — a minimalist diary built around the idea of "write and let go." No accounts, no feed, no social anything. Everything stays local on the device, no tracking, no data collected. In a world where every app wants to be a platform, a diary that's just a diary almost feels radical.

Milo: So the pattern in this whole section: small, focused tools that respect your data, mostly local-first, mostly cheap. Okay — from personal tools to tools aimed at customers and pipeline. This is where AI is heading for most businesses.

Mia: The incumbent play first: Microsoft Copilot. AI embedded across Word, Excel, Outlook, and Teams, with two newer pieces — Work IQ, and an Agent Builder for building agents. Plans start at $9.99 a month.

Milo: What Microsoft has that nobody else does is the surface area. Copilot doesn't have to convince you to open a new app — it's already inside the documents, spreadsheets, and email you use all day. And at $9.99 a month, the entry price is calibrated for broad adoption, not just enterprise.

Mia: The challenger angle is SaleSmartly: an omnichannel inbox that pulls in WhatsApp, Instagram, TikTok, LINE, and more, with a CRM attached, AI agents answering 24/7, and real-time translation across languages.

Milo: The translation piece is bigger than it sounds. If you're selling across regions, live translation in the inbox means one team can genuinely serve customers in multiple languages. They say over 300,000 businesses use it — that's a maker claim, take it as such, but even if it's inflated, it tells you the omnichannel-plus-AI-agent model has real demand.

Mia: Then Okara, which positions itself as an "AI CMO." It audits your website and then deploys more than ten agents to handle SEO, AI search visibility, Reddit, X, LinkedIn, and outreach to creators. They claim 100,000-plus businesses use it.

Milo: The "AI CMO" framing is the interesting part. Instead of buying a point tool per channel, you're hiring a system that runs all the channels at once. The obvious question — and it's an open one for everything in this section — is whether AI-generated outreach actually converts, or whether it just fills pipelines with noise that customers learn to ignore.

Mia: Which brings in Lattice as a supporting piece: it rewrites text by passing it through multiple languages, which makes it sound more natural and strips out the telltale markers of AI writing — while preserving names, figures, and technical terms.

Milo: That's directly relevant here. If agents are writing outreach at scale, and everyone's inbox fills with machine-sounding pitches, then a tool whose entire job is de-machine-ifying text becomes part of the go-to-market stack. Whether that's a good thing for recipients is another question entirely.

Mia: And one more that echoes all the way back to our first topic: Databox's MCP Connectors. They connect Genie — their AI analyst — to your CRM, Slack, and other tools, so the analyst has real context and can actually execute actions, with permissions set per tool.

Milo: Notice the echo: per-tool permissions is governance, exactly like what VibeDefend does for coding agents and what Vantage's approvals do. The same pattern — route, budget, approve, audit — is showing up everywhere, from dev tooling to marketing to analytics.

Mia: So where does this all leave us? A few honest unknowns. For the agent infrastructure: the speed and cost numbers are vendor claims, unproven at real scale. For the screen copilots: privacy and accuracy in everyday messy use. For the business AI: whether generated outreach converts or just adds noise.

Milo: And for the little utilities, honestly, the risk is lower — they're cheap, mostly local, and do one thing. Mochi never deletes, Ryu never phones home, Vitals costs five bucks. Those are low-stakes bets.

Mia: That's the show for today. Thanks for listening — we'll catch you next time.

Milo: See you then.