1111 | This Week's AI Launches: Agents, Dev Tools, and Everyday Apps

Mia: Hey everyone, welcome back to the show. It's Mia.

Milo: And Milo. Today we've got a launch roundup, but there's a thread running through basically everything: AI agents are stepping out of the chat window and into the real world. They're getting phone numbers, they're running your ads, they're testing your code, and they're showing up inside the tools you use every day.

Mia: Yeah, and rather than running through thirty products one by one, we picked out the stories with the clearest problems and the most concrete details, and we'll spend real time on each. Let's start with the most literal version of agents getting real-world access: phone calls.

Milo: Okay, so ForAgent. The pitch is private phone numbers for AI agents so agents can call each other. That alone is kind of a wild concept. But the detail that matters is the guardrail: every call connection requires human pickup approval. So a human has to approve before the call actually connects. And there's a free plan with thirty calls a month, so it's easy to try without committing.

Mia: What I like about that is they've clearly thought about the trust problem. Agents calling each other unsupervised is a scenario people worry about, and ForAgent's answer is basically "no call happens without a human saying yes." The open question, though, is how that holds up at scale. If this becomes infrastructure for lots of agent-to-agent traffic, does per-call human approval stay practical, or does it become the bottleneck? And how do you verify who's on the other end of these numbers?

Mia: Those are unanswered.

Milo: And it's not the only one putting agents on the phone. ChromePing is a browser product that calls real phones, with live two-way translation during the call. So you're in your browser, the call goes out, and both sides hear their own language. And on their Plus and Pro tiers, an AI agent can place the call for you, starting at $9.99 a month.

Mia: So ChromePing is more consumer-facing — you need something done over the phone, maybe in a language you don't speak — while ForAgent is infrastructure for agent-to-agent communication. Different layers of the same shift: the phone is becoming an agent channel, not just a human one.

Milo: Right. And then there's the third flavor: agents embedded inside products. SideKick puts an AI agent inside your product. The user states what they need in one sentence, and the agent goes off and completes the task. It's MCP-compatible, which means it can plug into a broader ecosystem of tools.

Mia: That's the "agents as UI" idea. Instead of clicking through menus, the user just says what they want and the product's built-in agent does it. For the maker, it's a way to add agentic behavior without building it from scratch, since MCP compatibility gives it a standard way to connect to things.

Milo: And the last piece of this first thread: AppuBot. Free AI teammates that use your actual Mac or PC applications, running on the AI plan you already have. And if you want them running around the clock, cloud bots start at $19 a month.

Mia: That's agents getting screen access, essentially. Not calling phones, not living inside one product, but sitting on your desktop operating your local apps. So across these four you get the full picture: ForAgent gives agents a phone line, ChromePing gives them outbound calling with translation, SideKick embeds them in products, and AppuBot puts them on your desktop.

Mia: The common thread is real-world channels and access — and the honest unknown across all of them is approval and verification at scale. ForAgent at least makes approval explicit; with the others, it's less clear where the human checkpoint sits.

Milo: Which is a good segue, because once agents have access, the next question is: what do they build, and how do we check their work? Let's talk about agents building and testing code.

Mia: Start with testing. Seldon — or Wren — is behavioral QA. You describe your app's behavior in plain English — the goals — and their agents test the app against that, find paths nobody tested, and hand back replays and repro steps. So instead of writing test scripts, you write down what the app is supposed to do.

Milo: Pricing is concrete too: ten free runs, then $3 per run. So it's priced per exploration, which is an interesting model — you pay when the agent actually goes hunting through your app.

Mia: The interesting claim there is "finds untested paths." That's the thing manual QA always struggles with — coverage. If the agent really surfaces the corners nobody thought to click, that's genuinely useful. But it's a maker claim; how thorough that actually is depends on how well your plain-English goals describe the app. Garbage in, garbage out applies.

Milo: And then the review side. Co-Review is open-source, MIT license, no telemetry, and the idea is you review an entire repository with the agent that built it. You comment on the code, and the agent answers from the code itself.

Mia: That's a smart loop. Normally code review is a human reading a diff cold. Here, the reviewer gets to interrogate the author — and the author is the AI, so it can explain its own choices by pointing at the code. Open source and no telemetry lowers the barrier for teams that care about where their code is going.

Milo: Then there's ZUS, a desktop IDE where the AI agent plans, edits, and runs tests. You review diffs, and — this is a nice detail — there's byte-exact undo. So if the agent makes a mess, you can roll back precisely. The whole thing is built on Tauri and Rust, and the installer is about 20 megabytes.

Mia: That 20MB number is worth pausing on, because Electron-based editors are famously heavy. A lightweight native-feeling IDE with an agent inside is a differentiated angle versus the big players.

Milo: And once the code is written and reviewed, you have to ship it. Ownkube lets you deploy a repo with one prompt to Claude Code, Cursor, or Codex. It sets up the app, Postgres, cache, all on HTTPS. Billing is per minute, and there's a bring-your-own-AWS deployment option.

Mia: So the pipeline is becoming: agent writes code in ZUS or your editor of choice, agent gets reviewed in Co-Review, agent tests behavior in Seldon/Wren, and agent deploys through Ownkube with a single prompt. Humans move to checkpoints.

Milo: And one more in this cluster: Almanzor. You paste a docs URL, and you get a cited AI assistant over those docs — one that's explicitly willing to say "I don't know" instead of making things up. Plus a hosted MCP server, so Cursor or Claude Code can search your docs directly.

Mia: The "I don't know" behavior and the citations are the differentiated parts. Documentation chatbots that hallucinate are worse than no chatbot. And the MCP server part means the docs aren't just a chat widget on your site — your coding agents can pull from them while they work. That ties right back into the agent-coding workflow we just described.

Milo: Okay, so agents build and test. What's the natural next step? Agents sell. Let's move into autonomous marketing and workflow automation — and this is where things get spicy.

Mia: Maxma is an AI teammate that runs end-to-end paid advertising across more than ten platforms. It does creative, tracking, even incrementality measurement. And crucially, it proposes changes and waits for Slack approval before executing. So the human stays in the loop via a channel they already live in.

Milo: End-to-end is the word to watch. Most ad tools optimize bids or generate an image. Maxma's claim is the whole loop: create the ads, deploy them across platforms, measure whether they actually drove incremental results, and propose the next move — with you approving in Slack.

Mia: The unknown there is approval overhead. If an "autonomous" teammate pings you in Slack every twenty minutes, you've just moved the work, not removed it. How much genuinely runs without you is the open question.

Milo: Upstream of ads is knowing who to talk to. Alsona tracks over 32 buying-intent signals — things like hiring activity, funding rounds, podcast appearances, procurement behavior — scores companies against your ideal customer profile, and then AI agents send personalized outreach based on that.

Mia: So the intent data and the outreach are one pipeline. Traditionally you'd buy intent data from one vendor, load it into a CRM, and have humans do the outreach. Alsona compresses that into signal, score, send.

Milo: And to actually send, you need infrastructure. That's Kata Send: cold email infrastructure built for AI agents. Warmed inboxes at $0.04 per email, or you can use your own starting at $29 a month. It automatically pauses when bounces spike — which is the thing that kills deliverability.

Mia: The auto-pause is the detail that shows they understand the failure mode. Cold email at agent speed can torch a domain's reputation in a day if nobody's watching bounce rates. Building the brake into the infrastructure is the right instinct.

Milo: Now, all three of those — Maxma, Alsona, Kata Send — are aggressive autonomy. The counterweight in this topic is Latch.ai, and I think it might be the most interesting one. It's aimed at accountants: they record their screen while narrating a task they're doing, and Latch turns that into a "sidekick" automation that runs alongside them — and when it hits an exception, it stops and asks.

Mia: That's the opposite philosophy. Instead of an agent acting and asking forgiveness, the automation is learned from a human doing the work and explaining it, and it's designed to stop at uncertainty. And they're SOC 2, which matters because accountants deal with sensitive financial data.

Milo: Latch is a template for trustworthy automation, honestly. Watch me work, learn my process, assist me, defer to me when something's off. Compare that with "agent runs your ads across ten platforms" and you see the spectrum in this space — from full autonomy with Slack approvals to ambient assistance that always checks in.

Mia: And the shared unknowns: deliverability for the outreach tools, approval overhead for Maxma, and for all of them — do these agents perform at the level the demos suggest? Treat those as claims until proven.

Milo: So agents build, agents test, agents sell. What's left? The tools humans actually live in — writing, notes, reading. And AI is being woven into all of it, quietly.

Mia: Start with TK by Medium. It's an agentic notebook — AI is inline while you're writing, there's an MCP server to connect to LLMs, a /wikify command, a daily note, and a built-in community.

Milo: The MCP server in a personal notebook is notable. Your notes become something agents can reach. And /wikify suggests turning messy notes into something structured.

Mia: Then the workspace side: Rodeo, a Notion alternative. Free plan includes AI meeting notes with 120 minutes a month, seven database views, and a built-in MCP server. Imports your Notion content. Paid starts at $3 per seat.

Milo: Free AI meeting notes is a sharp wedge against Notion, since AI features there sit behind paid tiers. And again — built-in MCP server. That's now showing up everywhere.

Mia: Twice in this section, actually. Then Coffee, a native Mac Markdown editor. Plain files — no proprietary database — plus inline sums, currencies, time zones. And its AI edits come as suggestions you review, not changes that just happen.

Milo: Plain files is the differentiator. Your notes are just Markdown on disk, portable forever. And "review AI edits" is the same trust pattern as Latch: propose, human approves.

Mia: Now mobile. Opal Reader is a PDF and EPUB reader for iPhone with a notebook beside the page — you read on one side, annotate on the other. Reading is free; Pro is $14.99 one-time for unlimited annotations.

Milo: One-time purchase, nice. LotusQ is a voice keyboard for iPhone — it types into any app in 78 languages and translates as you speak. Free tier is 2,000 words a week, and it's also on Mac, Windows, and Linux.

Mia: Voice as a system-level input rather than a single app. And Clive — this one's fun — an iPhone app that lets Claude set real alarms through Apple's AlarmKit. You add it as a connector to Claude, it's iOS 26, free for 5 alarms, Pro at $1.99 a month.

Milo: "Hey Claude, wake me up at seven" — and an actual alarm appears in the system clock app. It's a small thing but a real-world action from a chat agent. This whole section is AI woven into everyday capture and reading: TK for thinking, Rodeo for teams, Coffee for writing, Opal for reading, LotusQ for speaking, Clive for the most mundane task imaginable.

Mia: From there it's a short hop to the fun stuff — simulation, creation, play. And the headline here is MuJoCo 4.

Milo: DeepMind's open-source physics engine, first major version in three years. Adds a 3D Studio, better deformable bodies — simulating squishy, flexible materials, which is notoriously hard — and it runs in the browser.

Mia: MuJoCo is foundational — robotics research runs on it. Three years between major versions is a long time, and "runs in the browser" lowers the barrier massively: no install, just open a tab and simulate.

Milo: On the creation side, Lottie AI: text prompt to editable vector Lottie animations. The AI writes actual animation code that compiles to Lottie, you get controls to tweak, you refine through chat, and you export.

Mia: The editable part is what separates this from "AI generates a video." You get a vector animation file your designers can keep adjusting.

Milo: MoreFrame fits here too — free dual-screen recorder and editor for Mac and Windows, 500+ graphics, Smart Zoom that follows your cursor, AI editing via MCP, no watermark.

Mia: For anyone making tutorials or launch videos, that's a real stack: record two screens, get automatic zooms, and have an agent do the edit through MCP. And it's free with no watermark, which is generous for this category.

Milo: Then Ember — open-source, Slack and Campfire style chat. One-click deploy to Cloudflare's free tier, accepts Slack-compatible payloads, and there's an AI assistant built in.

Mia: Slack-compatible payloads is clever — integrations written for Slack can potentially point at Ember instead. Free tier hosting plus open source is a very low-cost self-hosted chat option.

Milo: And closing this thread on a lighter note: CollaboTokyo. A virtual Tokyo with nine districts and 81 spots, a space for creators, builders, and brands to network and collaborate. Free to explore.

Mia: A virtual world as a networking venue instead of another video-call platform. Ambitious, and whether people show up regularly is always the question with virtual spaces — but it's free to wander and see for yourself.

Milo: The theme of this whole stretch: simulation and creation keep getting cheaper and more accessible. A world-class physics engine in your browser, animations from a sentence, a video editor that's free.

Mia: Okay, last thread — the small tools that punch above their weight. We'll go a bit quicker, but each one solves a specific, real gap.

Milo: Quillo — voice-first. Reads PDFs, Word docs, and even photos aloud with sentence highlighting so you can follow along, and Quillo Ai answers questions about what you're reading. Free PDF tools included.

Mia: Photos is the surprising one — snap a page and have it read to you. Accessibility angle is obvious and important.

Milo: Obsiva, for field work: take site photos, add voice notes, and it produces an AI-written PDF inspection report. Works offline, bilingual English and Spanish, free on Android.

Mia: Offline plus bilingual tells you exactly who it's for — inspectors on job sites without connectivity, in the field, not at a desk. Converting photos and rambling voice notes into a structured report is a genuine time saver.

Milo: Inspo by Context.dev, for the design-to-code gap: you describe a site, pick real design references you like, and it exports everything as a prompt for your coding agent. Free MCP, no API key needed.

Mia: "Make it look like this" has always been the hardest part of prompting a coding agent. Giving it real visual references makes that instruction concrete.

Milo: Tokens by UI Colors — a visual playground for design tokens: color, typography, spacing, radius, shadows. Export as CSS, or as a DESIGN.md file you can hand to an AI agent.

Mia: DESIGN.md is the interesting export — design tokens as a document your coding agent can read and respect. That's design systems meeting the agent workflow.

Milo: FigSchema — a FigJam plugin for database schema design. Relations anchor to columns, you can export SQL or DBML. Free canvas, Pro is a $29 one-time payment.

Mia: Schema design on an infinite canvas instead of inside a database tool — good for planning before you build.

Milo: And AppTraction — a native Mac app for App Store Connect analytics. It uses AI to review your app's listing and write reports. Key detail: data stays local on your Mac. There's a 50% off code, HUNT50.

Mia: Local-only data is the differentiator. App Store Connect data is sensitive business information, and a lot of analytics tools want it in their cloud. Keeping it on your machine is a real answer to that concern.

Milo: So there's the full sweep. Agents with phone numbers and desktop access, agents building and testing code, agents running marketing with humans approving in Slack, AI woven through writing and reading, simulation and creation getting cheaper, and a pile of small sharp tools.

Mia: And the honest caveat to carry out of this episode: most of what we described is maker claims. The most credible signals are the concrete things — pricing, guardrails like ForAgent's human approval, Latch's stop-and-ask design, AppTraction keeping data local, Kata Send's auto-pause. Where you see those, you see teams who've thought about failure. Where you don't, ask questions.

Milo: Well said. That's the episode — thanks for listening, we'll catch you next time.

Mia: Bye everyone!