1116 | Launch Roundup: AI Agents, Email Apps and Smarter Work

||下载

节目备注

From AI coding agents and the data market feeding them, to native Mac apps, privacy paperwork and tools for small businesses — a quick tour of new launches worth knowing about.

Timeline

  • 00:00:04 Opening
  • 00:00:25 AI coding agents that touch real code
  • 00:05:27 The market behind AI agents: data and shared platforms
  • 00:07:52 Native email done differently
  • 00:09:29 Less admin: meetings, clients and small-business workflows
  • 00:13:22 Closing

Related links

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, everyone. I'm Mia.

Milo: And I'm Milo. Today we've got a stack of products that all circle around one idea: software and AI that actually does the work — not just talks about it. Agents that fix real code, marketplaces that feed the models, and a bunch of apps that quietly take admin work off your plate.

Mia: Let's start with security, because Polylane is one of the more interesting ideas here. It reviews production pull requests looking for exploitable flaws — not just style issues or linting complaints, but actual security problems.

Milo: And here's the part that makes it more than a static analysis toy: it traces an input all the way to the vulnerable line. So it can show you, concretely, how untrusted data flows from somewhere out in the world into the exact line of code where things go wrong. That's the hard part of security review. Anyone can flag a suspicious function; connecting an attacker-controlled input to it is the real work.

Mia: And then it doesn't just report — it opens fix PRs, including fixes for known CVEs. So the output is a pull request you can review, not a scary dashboard entry that lands in someone's backlog and dies there.

Milo: Which is why it matters. Security review has historically lived in a separate process — an audit once a quarter, a pen test before launch, a scanner that files tickets nobody reads. Polylane is trying to move that into the normal development loop, where the fix shows up as a PR next to everything else you're merging anyway.

Mia: Now, we should be careful here. Everything we're describing is what the makers claim. What we don't know is coverage and accuracy on real codebases. Does it actually find exploitable bugs at a useful rate, or does it bury reviewers in false positives? Does the trace hold up on a big, messy production repo? Those are exactly the questions that decide whether this is a tool or a demo.

Milo: And it fits a bigger pattern in today's episode. AI agents are increasingly acting on real code, not just suggesting completions in an editor. Polylane writes fix PRs. And on that same theme, there's Talyn.

Mia: Talyn runs Claude Code and Codex — the coding agents — in parallel sandboxes in the cloud. And the detail I like: it uses your existing plan. So you're not paying for a separate pool of model credits; your subscription does the work.

Milo: The free tier is 10 box-hours a month, which gives you a way to try it. Think about what that enables: instead of running one agent task at a time on your laptop, you can spin up several boxes in parallel, have them work on different tasks, and they're isolated from each other.

Mia: So Polylane and Talyn are two ends of the same shift. Polylane is agents reviewing and fixing code that's already in production; Talyn is infrastructure for running agents on your code at scale, in parallel, safely sandboxed.

Milo: There's a supporting cast here too. ByteAsk is an AI agent focused on C and C++ — which matters because those are languages where generic agents tend to struggle. Its pitch is that it indexes your repo according to your build, so it understands the actual structure of your project, not just the files. And the claim is that it produces the same fixes as Codex with 59 to 70 percent less cost.

Mia: Same caution as before: that's the maker's number, not something we've verified. But the interesting angle is the indexing-by-build approach. If you're working in C++, your build system defines what code actually gets compiled and how — an agent that respects that is making a real bet about what matters.

Milo: Then there's the visual end of the agent world: AgentGrid's Canvas. It's a free visual canvas for designing real HTML alongside AI agents, connected via MCP. It's local, no registration. So instead of an agent dumping code into a chat window and you squinting at it, you get an actual canvas where the HTML you're building exists visually, next to the agent working on it.

Mia: And AgentGrid's Rakazo pushes even further. It's an open source alternative to Grok Bot — about 8,400 stars — and the pitch is bots with memory, routines, and their own computer, self-hosted for free. That's the full self-hosting philosophy: your agent lives on your machine, your data stays with you.

Milo: One more in this family: Pixel Office. It's open source and runs locally with your own keys, and it presents AI agents — Claude, Codex, Gemini, via OpenRouter — as employees in a pixel-art office. Which sounds like a gimmick, but there's a real design idea underneath: instead of a chat sidebar, you see agents as distinct entities in a workspace, each doing their thing.

Mia: Okay, so that's the agent layer. But agents need two things to become an economy: something to eat and somewhere to live. And that's where the next two products come in.

Milo: First, the food. Perit Marketplace, from Y Combinator's F26 batch, sells data and RL tasks to AI labs. The headline numbers: $11 million in revenue in four and a half months, and 29 datasets listed.

Mia: Those revenue numbers are the company's own claims — we can't verify them. But if they're even roughly right, it tells you something about demand. Labs are buying data and reinforcement-learning tasks from third parties at real scale, which means the training-data supply chain has become an actual market with sellers, listings, and prices.

Milo: And the distribution side is Frevana. It's a platform for building AI agent apps, and it's backed by people from a16z and OpenAI. What's new is team features: you can now share agents and skills across a team.

Mia: That's a meaningful product step. Before, agents were personal — one person builds an agent, it lives in their account, and if they leave, the knowledge leaves with them. Sharing agents and skills means an agent becomes organizational infrastructure, like a shared codebase or a shared playbook.

Milo: Connect those two and you get the shape of the agent economy: Perit is the supply side — data and tasks that make agents capable — and Frevana is the distribution side — teams actually deploying and sharing them.

Mia: And there's a natural adjacency: if your team is sharing agents and skills, you probably want a visible, organized place for all that work. That's Tulvo — a workspace it calls adaptive, with connected data. The pitch is you describe your business, and the AI creates bases, boards, docs, and sites for you. So the workspace itself is generated around what your team does.

Milo: Like everything in AI right now, the question is whether what it generates actually fits your workflow or just looks impressive on first run. But the direction is clear: tools that assemble themselves from a description instead of you configuring them from scratch.

Mia: From agents acting on code, let's shift to something more personal: email. GotEmail is a native mail app for Mac and iPhone, and the defining choice is that it has no servers of its own.

Milo: That's a big architectural statement. Most modern mail apps route your mail through their infrastructure — that's how they add features like snoozing and scheduling across devices. GotEmail skips that entirely: the app talks directly to your mail servers, so there's no third party sitting in the middle of your inbox.

Mia: And the business model is unusual too: one license covers all your Macs. No subscription, no per-device upsell. For Apple users who are tired of paying monthly for everything, that's the pitch — pay once, own it.

Milo: What we don't know: any plans for Windows or Android. Right now it's Apple-native, which is also its strength — it can behave like a real Mac and iPhone app rather than a cross-platform compromise.

Mia: Email pricing is exactly the battle Emayler is fighting from the other side. Emayler gives you email on your own domain with a fixed price per account — they explicitly don't charge per user — with plans starting at $3 a month, and there's a free plan.

Milo: So GotEmail is attacking the client side — the app you read mail in — with no subscriptions, while Emayler attacks the hosting side — where your mail lives — with predictable per-account pricing. Two different layers of the same stack, both pushing against the subscription-everything model.

Mia: Alright, from inboxes to the rest of the admin pile. Spiik is a Mac app that transcribes meetings and classes in over 100 languages, identifies who's speaking, and — this is the differentiator — records Zoom, Meet, and Teams directly on your Mac without a bot.

Milo: That no-bot detail is genuinely important. A lot of transcription services work by joining your call as a participant, which means everyone in the meeting sees a weird guest named something like "Notetaker" and has to consent to an extra stranger in the room. Recording locally on the Mac avoids all of that.

Mia: Speaker identification matters too, because a transcript where nobody knows who said what is only half useful. The open question is the one that plagues every transcription tool: accuracy in noisy settings, overlapping speakers, bad microphones. Clean audio is easy; real meetings are not.

Milo: And Spiik is one side of admin relief — capturing what happened. TwoBucket is the other side — handling clients. It's mobile software for detailing businesses, $20 a month, with quoting and payments that go through the business's own Stripe or Square.

Mia: The "your own Stripe" part is the interesting claim. A lot of vertical SaaS takes a cut of every payment as the real business model. TwoBucket says: we charge you a flat $20, the payments run through your own payment accounts, no commission from us.

Milo: Detailing is a smart wedge — mobile car care is full of solo operators and small shops doing quotes over text messages and chasing invoices. But that raises the adoption question: does this generalize beyond detailing, or is it built so specifically for that workflow that other trades would need something else?

Mia: The admin-cutting theme extends further. Privacyboard detects what tools your business uses by looking at your web presence, and then drafts your privacy policy, a DPA, and a cookie banner — in about five minutes.

Milo: Five minutes for compliance paperwork that normally takes an afternoon with a lawyer and a template hunt. Now, the obvious caution: auto-generated legal documents are a starting point, not a substitute for advice on anything genuinely sensitive. But for a small site that just needs the basics to exist, the gap it fills is real.

Mia: Similar spirit in collecto.rs — it catalogs any collection using AI from photos. Over 14 million items and more than 60 collection types are supported. Free on iPhone and web, Android coming.

Milo: So instead of manually entering your records, stamps, coins, whatever — you photograph things and it does the cataloging. It's the same pattern as Spiik and Privacyboard: take the tedious data-entry part of a hobby or a job and hand it to the machine.

Mia: Remindly fits here too — notes with reminders built in: one-time, repeating, or location-based reminders. It works on iPhone, iPad, Mac, and web, and it can import from Evernote. That's clearly aimed at people leaving Evernote who want their notes and their to-do-style reminders in one app instead of two.

Milo: And a quick community-facing one: HuntMonitor is a live dashboard for the daily Product Hunt race — points, voting velocity, charts refreshed every 15 minutes. It's for makers and watchers who follow launch day in real time. We'd say: it's a window into the day, not a verdict on quality — votes measure attention, not how good a product is.

Mia: So let's zoom out. Today we saw agents writing security fix PRs and running in parallel sandboxes; a marketplace feeding labs data while team platforms distribute agents; mail apps pushing back on subscriptions; and a whole cluster of tools — transcription, quoting, compliance, cataloging — whose shared promise is simply: less admin.

Milo: The honest through-line: almost every number and capability we cited is a maker's claim, not a verified result. The interesting open questions — Polylane's accuracy on real codebases, Perit's long-term demand, Spiik in noisy rooms, TwoBucket beyond detailing — are the ones that will separate tools that stick from demos that shine.

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