
1114 | Agents, Docks and Stock Cards: This Week's Launches
节目备注
From Google's new budget coding model to a WoW voice companion and trading-card stocks, we run through the week's most interesting launches in agentic AI, developer infrastructure, design tools, and consumer apps.
Timeline
- 00:00:04 Opening
- 00:00:49 The agentic coding stack gets cheaper and more parallel
- 00:05:40 Claude moves into your meetings and your content pipeline
- 00:08:42 Owning your stack: site migration, data pipelines, and infra planning
- 00:13:20 Creative tools from the dock to the breadboard
- 00:17:19 Consumer watch: trackers, games, and your own data
- 00:22:19 Closing
Related links
- Gemini 4 Hydrogen
- Expo Simulators
- Burnstate
- qorqbot
- Trainn MCP
- Claude Meetings
- Replicate
- Filament by Galaxy
- Paxper
- Pocket Studio for Supabase
- Ganella
- Templify AI
- Hoverine
- de:volt
- GAtrends
- Kairos
- WHOOP Humans
- Google Pixel Tag
- Willo
- StockRip
- Aven App Agent
This episode is produced by Bri. Bri uses advanced AI technology to turn the feeds you care about into podcasts made for listening. Contact us at hi@bri.so.
文字稿
Mia: Welcome back to the show, everybody. I'm Mia.
Milo: And I'm Milo. And today, honestly, it's one of those weeks where the story almost tells itself, because if you look across what launched in the last 24 hours, there's a thread running through almost everything: agents are getting cheaper, agents are getting more capable, and the tools around them are racing to keep up.
Mia: Yeah, and the other thread is ownership. People are pulling their sites, their data, their infrastructure back out of closed platforms and into things they actually control. We'll get there. But let's start where the money is, or rather, where the money isn't anymore, because Google just cut the price floor for agentic coding.
Milo: Right. Gemini 4 Hydrogen. That's the new model, and it's positioned specifically for cost-effective agentic coding. The pricing is one dollar fifty per million input tokens and seven fifty per million output tokens.
Mia: And those numbers matter more than they might sound like they do, because think about how agents actually work. An agent isn't one call. It's dozens, sometimes hundreds of calls. It reads files, it plans, it retries, it runs tests, it reads the output. All of that is tokens. So when the per-token price drops, what really drops is the cost of letting an agent loop until it gets something right.
Milo: Exactly. The expensive part of agentic coding was never writing the final code. It was all the intermediate thinking, the failed attempts, the context you have to keep feeding back in. So a model priced like this changes the math on whether you can afford to run agents all day, or run several at once.
Mia: Now, the honest question that comes with any cheap model is what you give up. The source here doesn't tell us whether Hydrogen compromises on agent reliability to hit that price, and that's genuinely unknown. Cheaper per token is only a win if the agent still finishes the job without you babysitting it.
Milo: That's the right framing. Price per million tokens is a claim about cost, not about whether the agent actually works. And it's worth holding that skepticism, because this week there are a bunch of tools that only make sense if agents are reliable enough to run unsupervised.
Mia: Like what?
Milo: Okay, take Expo Simulators. This is cloud-based iOS simulators and Android emulators built specifically for coding agents. The idea is the agent installs builds, tests them, and verifies them, and you can run that in parallel.
Mia: So connect that to the pricing.
Milo: Well, that's the point. Parallel agent runs used to be a luxury because each run costs compute and tokens. If the model cost drops, running five agents at once instead of one becomes economically reasonable. But then the agents need somewhere to actually run the software they write. A coding agent that can't launch the app it just wrote is only half an agent. Expo is selling the other half.
Mia: And I like that this is aimed at agents, not at humans clicking around a simulator. Historically, simulators and emulators were developer tools for people. This flips it, the machine is the customer.
Milo: Right. And there's a constraint side to this too, which is funny, because we just said compute is getting cheaper, but there's one resource that isn't: your plan limits. Burnstate is a Mac menu bar app that shows your Claude Code, Codex, and Cursor usage limits against time. So it answers the question "how much of my quota have I burned and how fast."
Mia: And notably, it does this with local logs and no account login. Which for a tool that's watching your AI usage is a nice touch, because the last thing you want is another service that needs access to your accounts just to show you a number.
Milo: Yeah, so within one week you've got the price of the model going down, the infrastructure to run agents in parallel, and a tool that tells you when you're about to hit the ceiling of what your subscription allows. That's a whole stack forming.
Mia: And one more piece pushes agents even further out of the terminal. qorqbot is a proactive agent for business tasks, and it runs on your existing Claude or Codex subscriptions. It learns your preferences, and for sensitive actions it asks for your sign-off before doing anything.
Milo: That last part is the important design decision. An agent that touches actual business stuff, sending things, making changes, whatever that means in practice, needs a human checkpoint. Sign-offs on sensitive actions is the safety valve. And running on subscriptions you already pay for means it's not adding another bill.
Mia: So the picture from this first block: agents getting cheaper per token, environments to test them in parallel, a meter for your limits, and an agent that steps out of coding into general business work with a hand on the brake. What's still unknown is the reliability question we flagged, and for qorqbot, how well preference-learning works in practice, since that's a maker claim.
Milo: Okay, so let's follow the thread. Coding agents are one place AI shows up in your day. This week, Claude also showed up in two very different places: your meetings and your content pipeline.
Mia: Let's do the content one first, because it's the more industrial version of the idea. Trainn MCP connects Claude or ChatGPT to Trainn, and once they're connected, you can create product videos, help articles, and courses from recordings, just by chatting.
Milo: So the workflow there is, somebody recorded a walkthrough or a demo, and instead of a team manually turning that into documentation, you ask the model, and it produces the deliverable. MCP is the connective tissue, it's the protocol that lets the assistant reach into Trainn and do the work.
Mia: Who's that for? It's for teams that have to produce support content at volume. Every software company has this problem: the product changes, the help center goes stale, nobody has time to redo the videos. If chatting with an assistant can regenerate those from a fresh recording, that's real work being removed.
Milo: And what it changes versus existing options is the interface. Traditionally you'd open a documentation tool, upload footage, edit, write, format. Here the claim is you stay in the conversation. Now, to be fair, that's how Trainn describes it, and we haven't seen the output quality. "From recordings" can mean a lot of things.
Mia: Right, treat it as a claim. Now the meetings side. Claude Meetings. This is Anthropic putting Claude into Google Meet calls. Claude joins the call, takes notes, extracts action items, and builds Docs and Whiteboards live, during the meeting.
Milo: The "live" part is what separates this from the flood of meeting notetakers. Most of those summarize after the fact. This one is building artifacts while the meeting is happening. So by the time the call ends, there's a doc or a whiteboard that reflects what was discussed, not a transcript waiting to be processed.
Mia: Availability-wise, it's in beta, and it's for Team and Enterprise plans. So this isn't something an individual free user can switch on today.
Milo: And the open questions are pretty clear. We don't know pricing beyond the beta. And building Docs and Whiteboards live is a much harder promise than note-taking. Live summarization can be a little wrong and still useful. A whiteboard that's being constructed in real time, if it mishears something, it's building the wrong structure in front of everyone.
Mia: But thematically, put the two together and the direction is obvious. AI isn't just a coding tool anymore this week. It's in your meetings, it's your documentation department. The workday is being colonized.
Milo: And here's the natural next step of that thought. If AI helps you build products and content, at some point you want to own what you built, not rent the platform it lives on. And this week there was a whole cluster of launches about exactly that.
Mia: Start with the most visible one. Replicate. It takes sites built on Webflow, Framer, Squarespace, or Wix and converts them into real Next.js projects on GitHub. For free. And the result comes with a CDN and an open CMS.
Milo: That's a migration tool, and the pricing is the headline. Free. These page builders lock you in with proprietary formats. Your site exists inside their editor and their hosting. Replicate's pitch is an exit: your design and content become actual code in a repository you own, on GitHub, deployable wherever you want.
Mia: Who's it for? Anyone who built on one of those builders, hit the ceiling of what they allow, and doesn't want to rebuild from scratch. The rewrite is always the part that stops migrations from happening. If the conversion is automatic and free, the excuse disappears.
Milo: The limitation we should flag: we don't know how it handles complex sites long-term. A simple marketing page converts cleanly, presumably. A site with intricate interactions, custom integrations, lots of dynamic behavior, that's where converters usually stumble. And the source doesn't promise otherwise.
Mia: Same theme, different layer: data. Filament is an open-source data replication engine, written in Go, Apache 2.0 licensed. It does CDC, change data capture, and it claims up to 2.6 million rows per second, and 91 times faster than Airbyte.
Milo: Okay, unpack that for people who don't live in data pipelines. Replication means keeping a copy of your database in sync somewhere else, like your production database mirrored into a warehouse. CDC means instead of copying everything repeatedly, you stream just the changes. That's how you keep things current without hammering the source.
Mia: And the numbers are claims from the maker, we should say that plainly. 2.6 million rows a second and 91x faster than Airbyte, which is one of the established tools in this space. If it's anywhere near that, it matters, because replication speed is often the bottleneck that decides how fresh your analytics can be.
Milo: And the open-source angle fits the theme. Apache 2.0 means you can run it, modify it, embed it, without a vendor holding the keys. Same instinct as Replicate: own the pipeline, don't rent it.
Mia: Next layer up: infrastructure planning. Paxper lets you plan cloud infrastructure visually, as diagrams. You review the changes, and then you apply them, and they get applied in your own AWS account. There's a cost estimate built in too.
Milo: The "your own AWS account" part is doing a lot of work there. Plenty of infra tools want to be the layer between you and your cloud. Paxper's pitch is a visual planning surface that ultimately makes changes where your resources actually live. And cost estimates upfront address the classic fear of infrastructure-as-code: did I just provision something expensive?
Mia: And then the last piece of this block is the most practical, almost humble one. Pocket Studio for Supabase. It's an Android dashboard for your Supabase project: tables, a SQL editor, logs, and alerts. And it has a free plan with no ads and no tracking.
Milo: That's the "check on your backend from your phone" tool. Nothing glamorous, but think about when it's useful. Something breaks at night, you're not at your desk, you want to peek at the logs or run a quick query. A native app with alerts covers that.
Mia: And again, no ads, no tracking on the free plan, that's the ownership-and-privacy instinct again. So this block, end to end: get your site out of a proprietary builder, run your own data pipeline, plan your own infra visually, and manage your database from your pocket. It's one coherent argument about control.
Milo: Alright. You've built things, you own the stack. Now, the week also had a bunch of tools that are about making things, and some of them are delightfully weird. Let's start at the top of the screen.
Mia: Ganella. It's a native macOS dock replacement, and the hook is live widgets. Spotify, Calendar, Xcode, and Claude or ChatGPT. So instead of a dock that just holds icons, you've got actual functioning surfaces right there.
Milo: The AI widgets are the interesting bit to me. Having Claude or ChatGPT as a live dock widget means asking something doesn't require opening an app, finding a tab, whatever. It's just there. And it's native, which matters on the Mac, because native usually means it behaves right with the rest of the OS.
Mia: Pricing: 9 dollars 99, one-time, with a 7-day trial. A one-time purchase in a world of subscriptions is itself a small statement.
Milo: The open question, and it's a real one: how does it perform against Apple's own widget ecosystem? Apple has been investing in widgets and desktop stuff heavily. A third-party dock has to justify replacing something built into the OS, and whether it does is something the source can't tell us.
Mia: From the desktop to the phone. Templify AI is an iPhone editor for videos, photos, and carousels, with over 1200 templates, BeatSync for matching cuts to music, and auto-captions.
Milo: And the evidence here is more than a claim about features: over a million downloads, and it's bootstrapped. Meaning no raised money, it grew on its own. Now, downloads are not the same as happy long-term users, but a million downloads on a bootstrapped budget at least tells you it found an audience without a marketing war chest.
Mia: Who's it for? Social content creators, basically anyone who needs to produce decent-looking short video regularly and can't afford an editor or an hour per post. Templates plus beat-sync plus captions is the whole "make it look professional fast" bundle.
Milo: Then there's a tool for a slightly more niche audience, but an important one: motion designers. Hoverine ships about 100 Rive animation templates, and the detail that makes it smart is that the templates come with documented state-machine inputs.
Mia: Explain why that detail matters.
Milo: Because in Rive, animations aren't just videos, they're interactive, driven by state machines. If you buy a template and the inputs are documented, you can actually wire it into your app and control it. A beautiful animation you can't trigger is decoration. Documented inputs make it a component.
Mia: And they also ship Riv Eye, a free inspector for .riv files. So you can open and inspect the files themselves, which is a nice free companion to the paid template library.
Milo: And then the fourth one in this group goes back to fundamentals in the best way. de:volt is a browser-based breadboard simulator, with real physics, 105 components, and you can program Arduino and micro:bit boards in it, plus guided learning courses.
Mia: I love this one. A breadboard simulator with actual physics means you learn electronics by making the same mistakes you'd make on a real desk, minus the burnt components. And bridging into Arduino and micro:bit means the path continues from "I wired an LED" to "I wrote code that runs on a board."
Milo: And the audience is learners, students, hobbyists, maybe engineers who want to sketch circuits without a drawer full of parts. It's a very different kind of creative tool than video templates, but it's the same underlying urge: give people a sandbox to make things in.
Mia: Okay. We've covered the work stack and the making stack. What's left is the stuff you actually pick up or subscribe to as a consumer. And there's a lot of it, so let's take it in related pairs rather than one at a time.
Milo: Good, because some of these genuinely pair up. First pair: two things that watch things for you. GAtrends monitors your own keywords across 7 sources, and the mechanism is the clever part. It learns what "normal" looks like for each keyword, and when there's a spike above that normal level, it sends you a Telegram alert.
Mia: The "learns a normal level" part is what keeps it useful. A naive alerting tool pings you constantly and you mute it in a week. Baseline plus spike detection is the difference between noise and signal. And the free plan covers 3 keywords, so you can try the model without paying. Whether 3 is enough depends entirely on how many terms you actually care about.
Milo: The other watcher is a bit more personal. Kairos is a personal analytics engine that spans your health data, your finances, and your calendar. Every morning it gives you scores showing what's driving your energy and your mood.
Mia: And that's ambitious, because the interesting correlations are cross-domain. Maybe your mood dips when your spending spikes, or your energy tracks with meeting density. A single-domain app can't see that. A combined engine can, at least in principle. The unknown, as always with these, is how good the correlations actually are once it's looking at your messy real data.
Milo: Next pair: social and physical. WHOOP Humans turns the WHOOP app into a social network. You follow friends, you share sleep, recovery, and strain numbers, and you can post moments with a photo and a note. And there's granular privacy control over what you share.
Mia: Granular privacy is essential here, because health data is intimate. Maybe you'll share recovery with your training partner but not your sleep with your whole office. Letting people dial that per-thing is the difference between adoption and instant ghosting.
Milo: And the physical half of the pair is Google's Pixel Tag. 29 dollars 99, it's an Android tracker, it uses the Find Hub network, which Google says spans a billion devices, it has precision finding, and the battery is replaceable.
Mia: Replaceable battery is an underrated feature for a tracker, because trackers live in bags and coats and die silently. If you can swap the battery instead of the tag, the total cost of ownership changes completely.
Milo: The open question is accuracy. Find Hub's network is huge, but how it compares to Apple's AirTags in real-world precision finding, that's something we don't have evidence for in the source. Network size on paper doesn't automatically mean better finding in a parking garage.
Mia: And then three quick ones that all make something complicated feel easier. Willo is a voice companion for World of Warcraft. You ask your questions out loud, and it answers in your own voice. Free to start, with 50 answers a day.
Milo: Answering in your own voice is a strange, charming detail. WoW has years of accumulated systems and lore, and asking a question mid-game without alt-tabbing to a wiki is the actual user problem here. 50 answers a day on the free tier is a real trial, not a token gesture.
Mia: StockRip is the weirdest of the bunch. Digital trading cards backed by tokenized shares of companies like Nvidia, Tesla, and SpaceX. You open packs, like physical card collecting, but the cards represent actual stock, and it runs on Robinhood Chain.
Milo: So it's collecting culture meeting ownership culture. And honestly, the open question is the demand question: does turning equities into collectibles find a real audience, or is it a novelty? The mechanics are clear from the description, but whether people want their portfolio to also be a card album, that's unproven.
Mia: And the last one is Aven App Agent. An AI assistant that lives inside apps. You tell it what task you're trying to accomplish, and it gives you a walkthrough with highlighted UI elements, so you literally see where to tap.
Milo: That's the "nobody reads onboarding" solution. Instead of documentation, the app points at itself. If it works as described, it's a genuinely better support experience. The usual caveat applies: an in-app agent highlighting UI is easy to demo and hard to make reliable across updates and edge cases.
Mia: Alright, let's land this. If you zoom out on the last 24 hours, a few products had the clearest case. Gemini 4 Hydrogen, because the price floor for agents is a real, checkable number, and parallel agent infrastructure like Expo Simulators only makes sense because of it. Replicate, because free migration off proprietary builders with open output is as concrete as it gets. Filament, because open source plus CDC plus specific performance claims invites verification rather than hype.
Milo: And the running theme of caveats: model reliability at low prices is unproven, converters stumble on complex sites, performance claims are maker claims until benchmarked, and the consumer stuff, from Pixel Tag to StockRip, has open questions about accuracy and demand that only usage will answer.
Mia: That's the briefing. Thanks for listening, I'm Mia.
Milo: And I'm Milo. We'll see you on the next one.