0923 | Agents, Apps, and Open Weights: This Week's Launches

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

Milo: And I'm Milo. Today we've got a packed one — a whole batch of launches that, interestingly, cluster around a few themes: Mac utilities that make your computer a little smarter, AI agents moving into your work tools, and open-source pushes back against the subscription model.

Mia: And a big one on the AI infrastructure side — open model weights, regional hosting, and what that means for builders. Plus some fun consumer stuff at the end. Let's start where a lot of these tools seem to be converging: the Mac as a home for small, focused AI utilities.

Milo: Right, so there are five tools here that all fit that description. Let's start with Diurnal, because the idea is simple but useful — it's a Mac app for capturing tasks straight into Google Calendar.

Mia: Two keyboard shortcuts: Option-Command-K and Option-Command-L. That's the whole pitch. Capture a task, it lands in Google Calendar. And it works offline, which matters if you're on a plane or just want speed. There's also weekly planning built in, so you can step back and look at your week.

Milo: Who's it for? People who find task managers too heavy. If your to-do list lives inside your calendar anyway, this skips a whole app layer. What it changes versus existing options: most capture tools go to a separate backlog you then have to process into your calendar. This collapses that step.

Mia: What's unknown here — pricing beyond trials, and honestly, whether a two-shortcut tool survives long-term or gets absorbed into something bigger. Keep it in mind as we move to the next Mac tool, Blurt.

Milo: Blurt is open-source, and it's push-to-talk dictation. It uses AssemblyAI's Dictation API, and here's the part that makes it different from the usual voice-typing tools: the text appears right where your cursor is, in whatever app you're typing in. No window switching, no copy-paste.

Mia: So it's a Mac app, open-source, and the flow is: hold a key, talk, release, and your words land in the field you had focused. That's a real workflow improvement over browser-based dictation tools.

Milo: The open-source angle is worth dwelling on because we'll see it again today. It means you can inspect it, and you're not locked into a subscription just to talk to your computer. Open questions: how well the AssemblyAI dependency holds up, and whether the dictation quality is consistent across accents and noise.

Mia: Now the third Mac tool, QuietGlass, and this one has a genuinely interesting trigger mechanism. It's free and open-source, and it blurs your screen in two situations: when you look away — detected through AirPods — or when it detects a nearby face.

Milo: So think of someone leaning over to peek at your screen on a train. QuietGlass catches it and blurs. Or you just turn your head to grab a coffee, and the screen hides itself.

Mia: That's privacy solved at the hardware-peripheral level, which is clever because it requires no camera you have to trust — it's using sensors you already wear. The obvious limits: it depends on AirPods for the look-away detection, and face detection in a busy café might get noisy. But free and open-source is a good starting point.

Milo: Staying on the Mac, Edyt is the "works anywhere" editor. It's AI that lives in any app on Mac or Windows. You hit a shortcut and it rewrites or translates right at the cursor. It can also read what's on your screen, including PDFs.

Mia: The pricing detail we have: a free plan with roughly 130 edits per day. That's a meaningful number — it's enough for a working day of edits before you'd need to pay. What it changes versus existing options: instead of copying text into a chat window, you highlight in place and edit in place.

Milo: Unknowns here: whether the rewrite quality is good enough that 130 edits a day is actually all you need, and how it handles apps that don't play nicely with screen-reading.

Mia: And the last Mac tool is a bit of a niche one but interesting: Robot Voice Bridge. It's a Mac app for people working with ElevenLabs voice generation. It manages your takes with tags and IPA — that's the International Phonetic Alphabet, for getting pronunciations right — and it sends audio straight to the playhead in Pro Tools or Logic.

Milo: So for voice-over workflows, it removes the export-and-drag dance. $12 a month. Small audience, but if you're producing voice content, that integration matters.

Mia: So that's the Mac story: capture, dictate, protect, edit, and produce audio. What connects all of them is that the Mac is becoming a hub for these small, sharp utilities that do one thing well, locally, rather than trying to be a whole platform.

Milo: Which sets up a nice contrast, because now we're going to look at two tools that are the opposite of that — they're not small Mac apps at all. They're open-source alternatives to subscription software, built for writing and email.

Mia: First, Fulvid. It's an open-source Markdown and MDX desktop editor for Linux, Windows, and macOS. The key idea: your files live plainly on disk. No vault, no cloud, no proprietary format. It's just Markdown files where you put them.

Milo: That's a deliberate push back against tools like Obsidian or Notion, where your notes live inside a vault structure or a cloud workspace. With Fulvid, if you leave, you take your files — they're just files.

Mia: What's unknown: the ecosystem maturity versus hosted rivals. Open-source editors often have the philosophy right and the polish lagging.

Milo: And on the email side, Xem — open-source email marketing. It has a visual editor, automations, and over 200 templates. And here's the important part: you send through your own Amazon SES or SMTP setup.

Mia: So instead of paying a hosted email platform a monthly fee that scales with your list, you run the sending yourself. That's a real cost difference at scale. The trade-off is you're responsible for deliverability — the software doesn't magically fix your sender reputation.

Milo: So both of these — Fulvid and Xem — are betting that some users want ownership over subscriptions convenience. Now, that open-source instinct also shows up in AI infrastructure, which is where we're heading next.

Mia: But before infrastructure, let's talk about AI agents actually joining your work tools, because that's arguably the bigger shift. Three products here that are all variations on the same idea: agents as teammates with real permissions, not chatbots in a window.

Milo: Brev first. It's an AI coworker that lives in Slack and Teams. It tracks goals, preps meetings, and follows up on things — across more than 50 integrations.

Mia: That integration number is the differentiator. A chatbot in Slack is one thing. An AI that can pull context from your calendar, your docs, your project tracker, and then nudge you before a meeting — that's a different category of tool.

Milo: What it changes versus existing options: instead of you remembering to follow up, the agent does. The risk, and this is the honest unknown for the whole category, is reliability in production. If an AI coworker misses a follow-up, that's worse than no coworker.

Mia: Plane takes a slightly different angle. It's a project management platform, and it's adding AI agents as workspace members. These agents have playbooks, skills, and triggers — so they're not just answering questions, they're responding to events in the workspace.

Milo: And the pricing detail is notable: no extra seat cost for the agents. Most SaaS pricing would charge you for an extra member. Plane is treating agents as ambient labor rather than billed headcount.

Mia: Now WeWeb MCP — and this one is the most concrete about what an agent can actually build. WeWeb is a visual web app builder, and the MCP integration lets AI agents build pages, wire up data, and create workflows. Then a human reviews it in the visual editor.

Milo: That review step is the governance story. The agent does the building, the human keeps the veto. That's probably the right pattern for agents touching real production systems, and it's the same open question we flagged with Brev — how much do you trust it without that human in the loop?

Mia: There's a fourth data point that rounds this out: Anomalo Analyst. It's AI agents monitoring data warehouses and explaining trends and anomalies in plain language. We'll come back to it properly in a bit when we talk about monitoring data, but it fits the same pattern — agents embedded in a system, doing work, explaining themselves.

Milo: So what's happening is agents are moving from chatbots to teammates with real permissions. The tools are there. The governance and reliability questions are the open ones. And to make that work at all, you need the layer underneath — the models and the inference. That's where we're going now.

Mia: Three main items here. First, Xiaomi quietly did something notable: released MiMo-V2.6, omnimodal models in Pro and Flash variants, with a 1 million token context window. And importantly, they released the weights and the RL training code publicly.

Milo: Omnimodal means it handles multiple modalities — text, images, presumably audio — in one model. The 1M context is a big deal for long documents or codebases. And the RL code being public means researchers can reproduce and build on the training approach, not just the final artifact.

Mia: Contrast that with xAI's Grok 4.7, which is a commercial model. It targets code and long-horizon work. Pricing is $2 per million input tokens and $6 per million output. It's available in Cursor, Grok Build, and the API.

Milo: So that's the two paths side by side: open weights you can host yourself and modify, versus a priced API from a frontier lab. Neither is obviously right — open weights give you control, API gives you someone else managing the GPUs.

Mia: And then there's a third option emerging: regional hosted inference. 2BA.AI is EU-hosted, OpenAI-compatible, €20 a month for 4,500 requests in a 5-hour window, with zero logs.

Milo: That zero-logs claim matters for European companies with data residency concerns. And OpenAI-compatible means you can point your existing code at it with a config change.

Mia: So the picture: open weights, priced APIs, and regional hosting. Add two more pieces to that layer. vgpu is Vercel's TypeScript library for WebGPU, aimed at agents — typed WGSL shaders, runs in the browser, in headless Node, and in CI. 25 kilobytes.

Milo: That's GPU compute without a server, which matters for agents running locally. And then Valori, a memory layer for AI — vectors, graphs, deterministic state using Q16.16 fixed-point math, and BLAKE3 verification. SDKs in Python and TypeScript.

Mia: So the stack is forming: open models at the bottom, memory layers above them, GPU access locally or in the cloud, and regional inference options for compliance. The open question across all of it is real benchmark performance at scale — claims are easy, sustained throughput is hard.

Milo: Now, staying in the model layer but moving to the ecosystem around it — there's a Jev model appearing in multiple places, and that's worth noticing.

Mia: Fez is the anchor here. It's a Mac app — so back to our first theme briefly — where agents are members of chat rooms. The Jev model routes and decides which agent responds or what happens next. It's built on nostr, the decentralized protocol, and it's MIT-licensed.

Milo: So it's a chat client where some of the members are AI agents, and the routing is handled by a model rather than hardcoded rules. Open protocol, open license — that's the developer-friendly combination.

Mia: And the same Jev model shows up in two other products. gg-friggin-ez is a free Node.js profanity and toxicity screener — multilingual, and evasion-aware, meaning it catches people trying to sneak past filters with creative spelling. Free, open-source presumably, given the ecosystem pattern.

Milo: And Jev Wrapped analyzes Telegram channels for the share of ads and clickbait. It scored 1,500 posts. So the same model family is powering an agent router, a moderation tool, and an analytics tool.

Mia: That's an ecosystem forming — a model with a philosophy being applied across agents, moderation, and content analysis. Worth watching whether it becomes a standard or stays niche.

Milo: There's one more item in this cluster that's a bit uncomfortable but worth discussing: Freebuff Ads. It sells ads inside coding agents, reaching 500,000 developers, pay per click, with a $500 credit offer.

Mia: Ads inside the tool a developer uses to write code. That's a new surface. The honest unknown: how agent-native ads will be received. Developers are famously ad-averse, and there's a trust question when an AI assistant is also an ad surface.

Milo: And to close this cluster on a practical note: ResumeContext is an MCP that merges coding-agent sessions — from Claude, Codex, Cursor — into one file per project. Free beta. If you're juggling multiple agents on the same codebase, that's a real continuity problem it solves.

Mia: So that's the ecosystem: a Jev model showing up in agents and moderation, developers as a new ad audience, and tooling to manage multi-agent work. Now let's shift to the creative side — AI for video, design, and coaching.

Milo: Six products here, and there's a common thread: AI output you can still edit and direct. Start with VideoFlow Studio. It's an npx tool — so you run it from the terminal — and a coding agent turns your site's URL into an editable launch video.

Mia: The "editable" part is the differentiator. Most AI video tools give you a render. This gives you something you can keep changing, because it's being generated by a coding agent, presumably with the source available.

Milo: Clueso MCP is similar in spirit but different in mechanism. You create and edit videos through Claude or ChatGPT — from an idea or a screen recording — and the output is fully editable and on-brand.

Mia: So the pattern: video as a structured artifact, not a fixed render. And the same logic carries into design. PixelCrew is a crew of AI design agents that go from research to direction to wireframes to production HTML. Free alpha, bring your own API key.

Milo: Production HTML as the end state is notable — it's not a mockup, it's something a developer could actually use. The alpha status means expect rough edges, but the pipeline approach — research before design before code — is the interesting part.

Mia: Two more, both in the coaching-and-learning space. WZRD turns documents, slides, forms, and sheets into conversational AI experiences users can talk to.

Milo: So instead of reading a 40-page PDF, you ask it questions. It's a way of making static knowledge interactive.

Mia: And thestory.run is a writing coach for LinkedIn posts. The mechanism is interesting: the coach interviews you, and you write. So it's not generating posts for you — it's drawing out what you actually think through questions.

Milo: That's a real philosophical stance in a market full of auto-generate tools. And then Keet, an iOS app from YC, generates Duolingo-style video courses on any topic. Two free courses, then paid credits.

Mia: The common thread across all six: the human stays in the loop as editor or author. The open question, honestly, is whether AI-assisted output — a launch video, a design, a LinkedIn post — can match the quality of human-made work, or whether it just becomes fast and forgettable.

Milo: Now let's shift to data and monitoring, and this connects back to something we touched on — Anomalo Analyst, which we'll get to properly now.

Mia: Three products that all deal with observing systems and data. Pulsetic RUM adds real-visitor Core Web Vitals and error tracking by page and by country on top of uptime monitoring.

Milo: So traditional uptime monitoring says "the site is up." RUM says "the site is up, but visitors in Germany on mobile are having a terrible time on the checkout page." That's a much more useful signal.

Mia: SereneDB is open-source, Postgres and Elastic compatible, and claims top benchmarks for fast full-text search and analytics. One engine doing both.

Milo: The compatibility is the hook — if you already speak Postgres or Elastic, you can point your existing tooling at it. The unknown with any "claims top benchmarks" statement is, of course, whether it holds up under your workload, not theirs.

Mia: And then Anomalo Analyst, which we previewed earlier. AI agents monitor warehouse data and explain trends and anomalies in plain language.

Milo: So instead of an alert saying "metric X dropped 14%," you get an explanation of why. That closes the loop on what we said earlier about trusting AI in operations — observability is shifting from raw metrics to explanations.

Mia: And the risk is obvious too: an AI explaining an anomaly can be confidently wrong. So the same governance question from the agents-in-workspace discussion applies here.

Milo: Let's finish with consumer AI — the stuff that touches everyday purchases and everyday browsing.

Mia: Hola AI answers your calls with AI voicemail, filters spam, and sends summaries. $4.99 a month for the first six months. It's basically an AI receptionist for your phone.

Milo: The spam filtering is the part most people would actually pay for, given how bad robocalls are. And the summary means you don't have to listen to a voicemail at all.

Mia: Then the big one: Googlebook. A laptop designed for Gemini, syncing with Android phones, from $899, with preorders open.

Milo: "Designed for Gemini" is the phrase doing the work there. It's a laptop where the AI is the organizing principle, not a feature. And the Android sync is the lock-in play — your phone and laptop as one experience.

Mia: Whether that pairing actually changes buying habits — whether people pick a laptop because of how it syncs with their phone — that's the open question.

Milo: And the last one is a small browser extension with a nice idea: ReallyFree. It colors your search results green, yellow, or red to indicate whether the thing you're looking at is actually free, with no tracking.

Mia: Because "free" online often means "free trial that converts to a charge" or "free tier with nothing usable in it." A color-coded honesty layer is genuinely useful, and the no-tracking promise matters for an extension that watches your searches.

Milo: That's the show. A lot of ground today — Mac utilities, open-source alternatives, agents in work tools, model and infra layers, a Jev ecosystem, creative AI, monitoring, and consumer AI.

Mia: The through-line, if there is one: AI is moving from a thing you visit — a chat window, a website — into a thing embedded everywhere, from your keyboard shortcuts to your phone calls to your laptop hardware. What we don't know yet, across almost every product we discussed, is whether the quality and reliability hold up outside the demo.

Milo: Thanks for listening — we'll catch you next time.