1013 | Agents Get Real: Coding, Plumbing, and Everyday AI

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

Milo: And I'm Milo. Today we've got a stack of launches to dig through, and the thing that connects almost all of them is pretty simple: AI agents are finally getting the plumbing they need to do real work — real repos, real tabs, real calendars, real money.

Mia: Yeah, and we're going to be honest about what's a claim and what's proven here, because a lot of what we see is makers describing what their product does, not independent testing. So we'll walk through the coding agents, then the shared infrastructure they run on, how agents reach the live web, the money and sales side, the workspaces, and finally the tools turning saved content into action.

Milo: Let's start with coding, because that's where the shift is most visible. Codemind is a service where a coding agent delegates tasks over MCP, and the pricing model is the interesting part: ten cents per verified build, failed builds are free, and you get a hundred free builds the first month.

Mia: That pricing tells you a lot about where they think the value is. They're not charging per token or per seat — they're charging per successful, verified build. So if the agent produces something broken, that's their cost, not yours. It's a bet that verification is the bottleneck, and that people will pay for outcomes rather than effort.

Milo: Right, and it does suggest the pitch has moved past demo-ware. If you're willing to charge per verified build, you're claiming your agent can reliably produce working code. But we haven't seen data on success rates at scale — that's a genuine unknown. Ten cents a build sounds cheap until you're paying for ten builds to get one working change.

Mia: Now, delegation over MCP is one approach. Dropstone CLI takes a different angle: it's a terminal coding agent that's self-learning, with persistent memory across sessions. So the idea is you work with it today, and tomorrow it remembers what it learned about your project and your preferences.

Milo: Persistent memory is a big deal in coding agents because the classic complaint is you spend half your time re-explaining the codebase every session. If Dropstone actually retains that context — how your project is structured, what conventions you follow, what mistakes it made before — that's a real workflow improvement. It's a free plan, paid tiers run five to a hundred fifty dollars.

Mia: Though again, "self-learning" is the maker's description. The interesting question is what the memory actually stores and whether it transfers well between very different tasks. But between Codemind's pay-per-verified-build and Dropstone's memory-across-sessions, you can see the shape of the change: coding agents are being built for production workflows, where reliability and continuity matter more than a flashy one-shot demo.

Milo: And there's a third piece that completes the picture — what about people who aren't terminal people? Base Code is a browser workspace for your existing GitHub repos. You describe changes in plain language, it works in isolated branches, you get live previews, and it produces standard pull requests.

Mia: That last part is the key differentiator, honestly. Standard PRs means the work lands in the normal review flow your team already uses. It's not a parallel universe — it's your existing GitHub, just with a browser-based interface for people who'd rather not live in the terminal. Isolated branches plus live previews means a manager or designer can look at a change before it ever merges.

Milo: So across these three: Codemind charges for verified outcomes, Dropstone remembers across sessions, and Base Code opens coding agents up to non-terminal people. What we don't know — and should keep asking — is how these hold up on large, messy codebases, because that's where production reliability gets tested.

Mia: Okay, so if agents are going to do all this work, they need somewhere to keep their knowledge and somewhere to show it. That's where the infrastructure layer comes in, and it's honestly the most interesting part of the whole wave to me.

Milo: HexiCloud — the product is called Hexis — calls itself a hosted company brain for AI agents. Shared context, shared skills, and one MCP gateway. So instead of every agent in your company having its own little silo of knowledge, there's one place where the company's context lives, and agents plug into it through a single gateway.

Mia: The pricing is free up to three people, then ten euros per person per month, which is a small-team pitch. And the reason this matters is duplication: right now, if five agents in a company each need to know your product details, someone is maintaining that in five places. A shared brain with one gateway is one answer to that.

Milo: The open question is adoption. Companies are protective of their internal knowledge — will teams actually pool their memory into a shared hosted layer, or keep things local? That's a cultural and trust question, not just a technical one.

Mia: Meanwhile guilds.run attacks the visibility problem from the open-source side. It runs multi-agent teams and gives you full inspection of the prompts, the tool calls, and the memory. And they have a Studio component that drafts fixes when something goes wrong. You can self-host it for free.

Milo: Inspection is the word that matters there. Agent systems have a trust problem precisely because you often can't see why an agent did something. If you can look at the exact prompt, every tool call, and what's in memory, debugging stops being guesswork. And self-hosting means the inspection isn't just for the vendor — it's for you.

Mia: And here's a nice concrete example of the same thinking applied to forms: Cube Forms are agent-native forms. An agent can fill them in over MCP or HTTP — no browser, no CAPTCHA. But a human can open the same link and review what the agent did. Free, too.

Milo: I like that design because it acknowledges both audiences. Agents hate forms — CAPTCHAs and JavaScript-heavy pages exist to block bots — but forms are how businesses collect structured information. Making them fillable directly over an API while keeping the human review path is a sensible compromise. Same link, two ways in.

Mia: And one support piece here: Manifold MCP is a hosted MCP server that gives agents live data — SEO, AI answer results, social, ads, Reddit, and leads data. Free five hundred credits, then pay per call. I mention it now because it's infrastructure in the same sense — plumbing — but it's plumbing that connects outward, to the live web.

Milo: Which is actually a perfect bridge, because the next theme is exactly that: agents reaching outside, into the live web and your actual apps. And Manifold is a good example to unpack first.

Mia: Right. What Manifold is claiming to solve is the gap between an agent that only knows its training data or your documents, and an agent that can pull what's happening right now — search rankings, what AI answers say about a topic, what's trending on Reddit. If you're a marketer with an agent, live data is the difference between a generic answer and a useful one.

Milo: The pay-per-call model fits data work — you're buying queries, not seats. Though as with the others, quality of that data is the maker's claim. Garbage in, garbage out still applies.

Mia: Now, from live data to live conversations: Knocket is a free website chat and contact-page widget from Tencent RTC. Messages go to Telegram, and there's an AI agent answering questions from your docs.

Milo: The Telegram routing is the practical hook. A lot of small sites want a chat widget, but the real problem is what happens after someone sends a message — it goes to an inbox nobody checks. Sending it to Telegram means it lands where you actually are. And having the AI answer from your docs first means some questions never need a human at all. Free, which is aggressive.

Mia: Then there's the browser angle, and Strawberry is the boldest take on it. It's a desktop browser for Windows with AI agents built in, and the agents work inside your logged-in tabs.

Milo: That's a meaningful distinction. Most agent tools can't touch your authenticated sessions — the places where you're logged into your work tools, your dashboards, your accounts. Strawberry's bet is that's exactly where agents are most useful: inside the tabs you actually live in. And agents are shareable across teams, so you could build an agent for a workflow and hand it to a colleague.

Mia: Which raises the obvious security question — an agent operating in your logged-in tabs has your credentials in reach. That's powerful and risky at the same time, and the launch copy doesn't tell us how those boundaries are enforced. Worth watching.

Milo: And to show how far this MCP idea is spreading: Milepost is an iOS driving journal — it records your trips as Roadbooks — and it has an MCP server so ChatGPT or Claude can talk to your driving data. Free to record, Pro is twenty-nine ninety-nine a year, with a FIRSTDRIVE offer at fourteen ninety-nine.

Mia: A driving journal with an MCP server. Think about what that means: your agent can ask questions like "how much did I drive for client work this month?" using real trip data. The interesting signal here isn't Milepost specifically — it's that even a personal iOS app is now shipping an MCP endpoint. That protocol is becoming the default way agents touch anything, from search data to your car.

Milo: From infrastructure to commerce. Because if agents can reach live data and your accounts, someone will point them at sales and at your wallet.

Mia: Let's take the sales side first. SalesRep.com puts an AI sales rep on WhatsApp, Instagram, and other channels, with live prices, quotes, and follow-ups, in over thirty languages, starting at thirty-nine dollars a month.

Milo: Live prices and quotes are the substantive part. A chatbot that says "check our website" isn't a sales rep; one that can quote current pricing and follow up on an abandoned conversation actually occupies the role. Multilingual matters too — for a small business, one AI rep covering thirty-plus languages replaces either hiring or losing international leads.

Mia: Then there's the marketing side: Bearconnect 2.0 is a LinkedIn platform combining AI content creation with outreach targeted at people who engage with your posts. It has team workspaces and role-based access.

Milo: The engager-targeting logic is the differentiator — instead of cold-outreach to strangers, you reach people who already interacted with your content, which is a warmer pool. Role-based access tells you it's aimed at teams, not solo creators, which puts it in company-tool territory rather than personal-brand territory.

Mia: And now flip it: AI on the saving-money side. Kudos 3.0 claims three things — it activates card offers worth around eight hundred dollars a year, negotiates bills for about two hundred sixteen a year, and cancels forgotten subscriptions for about two hundred forty-seven a year. And critically, they say the user keeps one hundred percent of it.

Milo: That last claim is the business-model statement, and it's worth pausing on. A lot of bill-negotiation services take a cut of savings — often a hefty percentage. Kudos saying you keep everything means they're monetizing some other way, presumably through the card-offer side or a subscription. That's a genuinely different structure from the standard savings-split model.

Mia: The numbers are, again, maker estimates — your mileage depends entirely on your bills and spending. But the category is real: forgotten subscriptions alone are a well-documented drain, and an agent that notices and cancels them addresses it without you doing anything.

Milo: The open question for this whole group is consumer trust. An AI rep quoting prices, an AI negotiating with your internet provider, an AI holding your card context — these all require handing an agent real financial agency. Some people will love that; some will never. That's not a small adoption hurdle, and the launch copy doesn't resolve it.

Mia: Which actually sets up the next theme nicely, because the tools we're about to discuss are all built around one answer to that trust problem: keeping a human in the review loop.

Milo: Right. ExcaliClaw is a hand-drawn diagram workspace — that Excalidraw feel — with an MCP server for Claude and Cursor. Free tier gives you two hundred drawings with unlimited AI use, and paid plans start at nine euros a month.

Mia: Why give a drawing tool an MCP server? Because diagrams are a great agent interface. An agent can generate or update a diagram from a conversation, and a human can see it, sketch on it, correct it. It's visual collaboration between human and machine — the person stays in the loop by literally drawing on top of what the agent produced.

Milo: And at the data end of the pipeline, Scrape Anything by Context.dev is one API that turns any URL into Markdown, HTML, screenshots, or structured data, claiming a ninety-nine point nine six percent success rate, with a thousand free credits to try it.

Mia: Agents constantly need web content in a clean form — raw HTML is a mess for a model, Markdown is perfect. A single scraping API with a high success claim is the kind of boring-but-essential piece that agent workflows sit on top of. The success rate is their claim; scraping is notoriously fragile, so that number deserves scrutiny, but the thousand free credits make it testable.

Milo: And at the output end: SquadRoom runs AI employee agents that actually finish tasks. It's bring-your-own AI key — you pay forty-nine dollars a month for the platform and supply your own model costs — and nothing gets sent without your approval. It's built by a SpeedVitals cofounder.

Mia: That approval gate is the answer to the trust question we just raised. For a lot of teams, "the agent does the work, but a human approves anything that leaves the building" is exactly the compromise that makes AI employees acceptable. Bring-your-own-key is also a real cost model — you're not paying a markup on tokens, just for the orchestration layer.

Milo: The open question there, which Mia mentioned earlier in a different context: how much does approval friction slow teams down? If every outbound action waits for a human click, you've traded speed for safety, and the right balance probably differs by team. The maker bet is that forty-nine a month plus your own key costs makes that trade worth it.

Mia: And one related piece from the notes side: Understory Notes is a Mac app for going from note-taking to writing. Subnotes can have multiple parents, so they form outlines organically, and you can publish to PDF, Markdown, or web. It's a one-time purchase, not a subscription.

Milo: Worth mentioning here because it's the human counterpart to all these agent tools — the place where your own thinking gets shaped before it ever becomes an agent task or a published document. And the one-time-purchase pricing is a nice contrast to the sea of subscriptions in everything else we've discussed today.

Mia: Okay, last theme: content you've already got — documents, saved articles, recordings — turned into something you can act on. And this is a story in two halves: consumer apps at the front, and one foundational model underneath.

Milo: Start with Invited. You scan a bill, an invitation, or a cheque, and it turns the dates into reminders. It can find multiple dates in a single document, it works in over eighty languages, syncs across nine calendar services, and there's no signup.

Mia: The multiple-dates-per-document detail is what makes it more than a gimmick. An invoice has a due date, maybe a discount deadline; an invitation has an RSVP date and the event date. Extracting all of them is the real-world problem. And no-signup plus eighty-plus languages tells you it's aimed at maximum frictionlessness — point your camera, done.

Milo: Then Ancher 2.0 solves the hoarding problem. It captures articles, videos, and posts you've saved, lets you search them by rough description — not exact keywords — and turns them into briefs and posts for agents.

Mia: Rough-description search is the important phrase. Everyone has a saved-for-later graveyard; the reason it's useless is that you can't remember the exact wording. If you can say "that piece about pricing pages I saved in the spring" and get it back, the pile becomes usable again. And feeding the results into agent briefs closes the loop: saved content becomes raw material for actual work.

Milo: Then ngram 2.0 goes the other direction — it takes papers, decks, and recordings and turns them into narrated explainer videos. You choose the audience and the depth: Overview, Detailed, or Comprehensive.

Mia: The audience-and-depth choice is the thoughtful part. The same source document explained to an engineer versus an executive should be two different videos, and the three depth levels make that an explicit decision rather than a black box. It's essentially a knowledge-transfer tool — dense material into something a person can actually watch.

Milo: And underneath all three of these — the scanner reading a cheque, Ancher finding the article you half-remember, presumably ngram understanding a paper — is the same underlying capability: understanding content across formats. Which brings us to EmbeddingGemma 2.

Mia: It's Google's open embedding model, seven hundred forty million parameters, multimodal — it handles text, image, video, and audio — and it's released under Apache 2.0, designed for on-device use.

Milo: Apache 2.0 means genuinely permissive — commercial use included. And on-device is the headline: embeddings computed locally mean search and understanding over your documents without shipping them to a server. For the tools we just discussed — especially anything touching private bills or company papers — that's the difference between usable and not usable for a lot of people.

Mia: So it's the quiet foundation story of the day. The apps are the visible layer; a small, open, multimodal model running locally is what could make this category work everywhere, cheaply, and privately.

Milo: Let's tie it together. Today's pattern: coding agents getting paid per verified result and remembering across sessions; shared brains and inspectable agent teams making the work visible; MCP connecting agents to live data, browsers, and even your car trips; agents selling and saving real money; workspaces keeping humans in the loop; and open models making it all work on-device.

Mia: And the honest caveats throughout: nearly everything we covered was described by its maker, not independently verified. The reliability claims, the success rates, the savings figures — treat them as claims. The open questions worth keeping an eye on: do coding agents hold up on real codebases at scale, do teams adopt shared memory, and do consumers trust agents with money?

Milo: That's the show. Thanks for listening — we'll see you next time.

Mia: Bye, everyone.