1014 | Agents Everywhere: The Tooling Wave

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

Milo: And I'm Milo. Today we've got a stack of launches to walk through, and honestly, the thing that ties most of them together is one idea: agents are no longer just answering questions — they're running other agents, running servers, running your design files, your video edits, your books.

Mia: Right. It's less "AI app" and more "AI supervising AI," and every layer of the stack is getting that treatment. So let's start right at the top of that stack, with orchestration and coding agents.

Milo: Let's do it. First up is Remocn Studio. This is a free, MIT-licensed macOS app, and the pitch is pretty specific: it takes what your Claude Code or Codex agents build and turns it into a real Remotion project, which you can then export as MP4 up to 4K.

Mia: For anyone who doesn't know Remotion — it's that framework where you write video as code. So the value here is that your coding agent isn't just producing snippets or demos; its output becomes an actual video project you can render and ship. That's a real workflow, not a toy.

Milo: And the fact that it's free and open-source matters. There's no subscription gate on the path from agent output to a finished 4K video. The limitation, obviously, is that you're on macOS, and you're depending on your agent writing good Remotion code in the first place — the studio isn't going to fix a bad script.

Mia: Okay, so that's one agent's output becoming something concrete. But Scape takes it further — it's about running many agents at once. It's agent orchestration on macOS, built around something called the Argus loop orchestrator, which drives an agent fleet with a heartbeat, and uses SQLite tickets as memory.

Milo: I like how concrete that is. Instead of a vague "orchestrate your agents" promise, you get a mechanism: a heartbeat keeps the fleet moving, and each agent's state lives in a ticket in SQLite. That's auditable. You can open the database and see what's happening. And it's open-source with examples included, which means you can actually study how the loop works rather than trusting a black box.

Mia: The open question for me is scale — how many agents can one Mac drive before the heartbeat model strains? But as a pattern, "fleet plus tickets plus heartbeat" is a genuinely different take from the chat-window model.

Milo: Now, within that orchestration world, the coding agent itself is maturing. OpenCode 2.0 is the big open-source example here: 208,000 stars, 16 million monthly developers, and it runs in the terminal, on desktop, and in the IDE, supporting more than 75 model providers.

Mia: Those numbers come from the makers, so treat them as claims — but the design point stands. A coding agent that isn't locked to one model vendor changes your leverage. You can swap the brain without changing your workflow. That's why 75-plus providers is the detail I'd highlight, more than the star count.

Milo: And then there are two tools that are almost the same idea from different angles. MausBot is an open-source, Apache 2.0 alternative to Grok Bot — agent teams inside your chat app, running on your own API keys, and it can operate the Claude, Codex, and Grok CLIs.

Mia: Meanwhile CLI Manager added an AI Control API, which lets one agent drive other agents in a visible terminal. It's local, with REST and MCP interfaces, and it's off by default. That "off by default" detail is worth pausing on — they clearly know handing an agent control of other agents is sensitive.

Milo: So MausBot is agent teams in chat, CLI Manager is one agent visibly steering others in your terminal. Different surfaces, same question: who's accountable when an agent delegates to another agent? Both tools are at least making the delegation visible, which is the right instinct.

Mia: Alright, so agents are orchestrating agents, coding agents are everywhere. Naturally the next layer down is: what do those agents touch? Which brings us to infrastructure and ops.

Milo: Servo is the headline here — one console for all your servers. Web terminal, file editor, monitoring, status pages. And the security story is the interesting part: no open ports, and commands issued in the browser are signature-verified.

Mia: That's a real design decision. Normally remote server management means either exposed SSH or some tunnel. Servo says no open inbound ports at all, and instead verifies that commands actually came from you via browser signatures. If that holds up, it meaningfully shrinks your attack surface while still giving you a full control room.

Milo: And if agents are going to touch your servers — which, given everything we just talked about, they will — you want exactly this kind of centralized, verified chokepoint rather than scattered credentials.

Mia: Now the security theme gets sharper with Upsolve. They've added row-level security that's enforced server-side, before queries run. The key claim: agents can't see the rules, so they can't bypass them. You can test the rules as any user, and one proxy can serve thousands of tenants.

Milo: That's clever. A lot of AI safety talk is prompt-level — "please don't leak data." Upsolve moves enforcement below the agent entirely, into the query path. The agent literally doesn't have the rules in its context, so there's nothing to jailbreak. And testing as any user is what you need to actually audit that.

Mia: The thing to flag is that this is the maker's description of their own system. "Agents can't bypass it" is a strong claim, and the honest response is: good design, now let's see it audited. But architecturally, putting permissions below the model is the right direction.

Milo: Rounding out ops is vibesmonitor — an operations hub covering traffic, errors, user feedback, and cookie consent across all your projects, with one script per site and an MCP server included.

Mia: The MCP part is the thread back to our earlier conversation. Monitoring that speaks MCP means your agents can read your ops data directly. So the loop closes: agents build the sites, agents watch the sites.

Milo: And OpenFlowKit fits here too, though it really belongs to the next topic. Actually, let's bridge naturally — because the idea of "one central surface an agent works against" leads straight into documents and design.

Mia: OpenFlowKit 2 is an MIT-licensed diagram canvas, local-first, no account needed. You can import code, Mermaid, or D2, AI agents can edit it in real time through MCP, and there's a CLI for checking drift in CI.

Milo: That CI drift check is the sleeper feature. Diagrams rot — the picture stops matching reality. If a CLI in your pipeline flags when the diagram diverges from the code, diagrams become living artifacts instead of stale documentation.

Mia: And compare that with HarnessHive, which does the same "one live document" idea for hardware. AI-native wire harness design: schematics, wiring, bill of materials, and 1:1 formed boards all in a single real-time document, collaborative, with an MCP server.

Milo: Physical stuff! Schematics and BOMs and actual formed boards in one shared surface, with an MCP endpoint so agents can work on it too. The parallel with OpenFlowKit is direct: same thesis, that the collaboration unit should be a single live document agents can touch, whether that's a wiring diagram or a software architecture map.

Mia: Then the third piece of this cluster is The Drive AI — a file agent that works across Google Drive, SharePoint, Dropbox, and others to organize, dedupe, and generate documents. And crucially, it shows you a plan first and waits for approval.

Milo: That approval step is the trust design. A file agent moving things around in your Drive is exactly where you'd want a human checkpoint. It's the same instinct we saw in CLI Manager being off by default — these tools are learning that visibility and consent are features, not friction.

Mia: So: diagrams, hardware docs, and files, all becoming agent-editable live surfaces. What happens when the thing being created is content itself? Let's talk creators.

Milo: Takes for Mac first — it's free and open-source, a teleprompter app where the prompter follows your voice. But then Claude Code can cut the footage, generate vertical clips, and create posts. And there's a neat touch: your notes become rules — so your editing preferences get encoded as instructions the agent follows.

Mia: The voice-following teleprompter alone solves a real problem — reading a script while sounding natural is hard because you're always slightly out of sync. If the prompter tracks your speech, that friction disappears. And then the same tool chains into editing and publishing. One app covering record, edit, distribute.

Milo: PostNitro for Agents is the publishing layer generalized: it lets agents design and publish social posts via MCP, CLI, or REST, with one-click OAuth, covering LinkedIn and five other platforms.

Mia: So your agent can go from "here's my raw footage" to "here are the posts" without you touching a composer window. Six platforms through one integration.

Milo: And Agent Craft zooms out to the whole routine: a daily three-minute voice or video note becomes posts across eight platforms. They price it from $95 a month, and they make a claim worth repeating carefully: personal accounts get roughly seven times the engagement of company pages.

Mia: That seven-times figure is the maker's own claim, not something we can verify. But it points at a real shift people keep talking about — personal brand as the channel, versus the faceless company account. If there's any truth to it, tools like this are betting that the founder's voice, fed daily, beats polished corporate content.

Milo: And that raises a genuinely open question: if everyone's agent posts daily on eight platforms, does the advantage just cancel out? Worth watching. Which leads to the follow-on — if personal presence matters, how do people even find you now?

Mia: Two launches point at new distribution surfaces. MentionBird calls itself "Ahrefs for AI search" — it tracks which sources ChatGPT, Gemini, and Claude actually cite, then generates an action plan to raise your brand's mention rate.

Milo: The logic is sound: if people ask AI assistants instead of search engines, the citations those assistants give become the new page-one. A tool that shows you what's being cited and what to do about it fills a real gap. The caveat is that the "action plan" quality depends entirely on the underlying analysis — that's where I'd want to see examples before judging.

Mia: Vincentt goes somewhere completely different: WebAR and camera experiences. You build with your own agent plus an open SDK, preview live on your phone, and publish to a link — no app store involved.

Milo: Publishing to a link is the whole point. No review process, no install friction. Between MentionBird and Vincentt, you've got two answers to the same question — where does attention live next — one inside the AI assistant, one outside the app store entirely.

Mia: And from new surfaces, it's a short hop to the tools sitting closest to daily life: assistants. Kai comes from the Morgen team, it's an AI admin assistant at $20 a month, doing inbox triage, drafting replies, meeting transcription and summaries, and task tracking.

Milo: That's the classic "get a chief of staff" pitch, priced like a software subscription. The Morgen team has calendaring pedigree, which gives the task-and-meeting side some credibility. The open question is whether inbox triage is good enough to actually trust — that's the make-or-break for any assistant.

Mia: GotEmail approaches email from the opposite end: a native Mac app, one-time payment of $9.99 for all your Macs, with one-click copying of login verification codes and on-demand AI drafting and summarizing.

Milo: The one-click OTP copy is such a specific, honest feature. It targets the most annoying two minutes of email — hunting for a code — rather than promising to run your whole life. And the pricing model says something: buy it once, it's yours. That's a statement against subscriptions.

Mia: Then Toby, which is a different kind of assistant: an AI mock interviewer. You paste in the job description, it researches the company, interviews you round by round, and after the call gives you a score and improved versions of your answers. It bills per minute, with the first five minutes free.

Milo: Pay per minute is an interesting pricing fit — an interview runs a predictable length, so you're paying for actual practice time, not a monthly habit you'll abandon. And the improved-answers feedback is the part with real value. The limitation is obvious though: an AI interviewer isn't the real interviewer, so calibration to actual hiring outcomes is unknown.

Mia: And then the model layer underneath all of these: Meta's Muse Wonder, a new model family aimed at personal and professional agents, with proactive memory, proactive focus, and specializations in legal, finance, and health.

Milo: Proactive memory and focus are the features that would change assistant apps. Today's assistants mostly wait to be asked; a model that notices what matters without prompting is a different category. And regulated-domain specializations — legal, finance, health — suggest Meta wants agents operating in high-stakes territory, which raises the evaluation question all over again.

Mia: Which is a good bridge, because the next batch deals with exactly high-stakes mundane stuff: your body, your money, your team. Fitbit Edge first — Google's new slim band with a screen, built-in GPS, ECG, seven-day battery life, and a replaceable battery. $179.99, on shelves October 27.

Milo: The replaceable battery is the angle I'd underline. In a market where everything is glued shut, "you can repair this yourself" is a genuine differentiator, and it's coming from Google, which is notable. Seven days of battery plus GPS plus ECG at that price is a competitive spec sheet on paper.

Mia: From hardware to money: Fina Business, at $20 a month, is a QuickBooks alternative — double-entry bookkeeping, automated invoicing, Stripe payments, and notably, one workspace where you can manage your personal finances alongside multiple businesses.

Milo: That last part is the differentiator. Solo founders and small operators constantly juggle "which entity paid for this?" and QuickBooks makes that clunky. A single workspace spanning personal plus multiple businesses matches how messy small-business finance actually is. Whether it matches QuickBooks on depth — payroll, accountant handoff, all that — is the open question.

Mia: Then two team tools. Genchi runs anonymous one-click confidence votes in Slack, and aggregates them by project to predict delays. There's Jira and Asana integration, MCP support, and it's free for teams under ten people.

Milo: Anonymous confidence voting is a smart signal — people won't say "this project is slipping" in a standup, but they'll tap a button privately. Turning that into delay predictions via Jira data is a genuinely useful synthesis. The obvious caveat: anonymous sentiment can be noisy and gamed, and "predicts delays" is a claim we'd want to see validated over real timelines.

Mia: And Articos MCP does simulated research: inside Claude or ChatGPT, it runs simulated buyer tests against your messaging and pages, claiming 86% agreement with human research and reports in 30 minutes.

Milo: Eighty-six percent agreement is the number to hold carefully — it's the maker's figure, and agreement with "human research" depends entirely on which humans and which research. But the concept is provocative: if simulated buyers cluster reasonably close to real feedback, you can iterate messaging in minutes instead of weeks, then confirm with real tests.

Mia: And Genchi plus Articos actually rhyme: both are about engineering trust and honest signal into workflows — anonymous truth-telling in one, simulated truth-testing in the other. Neither replaces the real thing, but both make the expensive version cheaper to approximate.

Milo: Which brings us to our final thread: open source and local-first, privacy as a feature rather than a compliance checkbox.

Mia: OpenWhispr on iOS is 100% open-source voice transcription running entirely on device. Free, unlimited, works offline, built on Parakeet and Whisper, with optional cloud — or bring your own API key if you want it.

Milo: Everything about that positioning is a statement. Free and unlimited, because there's no server cost when it runs locally. Open-source, because with transcription, people rightly want to know their voice isn't being stored. And "bring your own key" is the escape hatch for people who want cloud quality without handing data to the vendor's cloud.

Mia: The parallel is GotEmail — we talked about it earlier, but it fits this theme exactly. A native Mac app, one-time purchase, on-demand AI rather than always-on. Both products are betting that on-device is a differentiator people will pay attention to.

Milo: And it's a bet worth watching. If enough tools go local-first, the "privacy-first" label stops being a niche and becomes table stakes. We don't know yet whether the market actually rewards it — that's the honest unknown.

Mia: So let's try to pull the whole day together. Agents orchestrating agents with Scape, MausBot, and CLI Manager, with Remocn turning agent output into real video and OpenCode as the dominant open coding agent.

Milo: Then the infrastructure those agents touch — Servo, Upsolve, vibesmonitor — all converging on centralized, verified, below-the-agent control. The live-document pattern with OpenFlowKit, HarnessHive, and The Drive AI. The creator stack with Takes, PostNitro, and Agent Craft pushing personal brands onto eight platforms.

Mia: New surfaces in MentionBird and Vincentt, assistants from Kai to Muse Wonder, mundane high-stakes workflows in Fitbit Edge, Fina, Genchi, and Articos, and the local-first privacy push with OpenWhispr.

Milo: If there's one question to carry out of today, it's the accountability one: as agents delegate to agents, every one of these tools is making its own choice about visibility — heartbeat tickets, signed commands, approval plans, off-by-default switches. Whether that's enough is genuinely unknown.

Mia: Agreed. That's the show for today — thanks for listening, and we'll catch you next time.

Milo: See you then.