0817 | Onyx, Chert, HarnessRouter, and Expeditione: PH's New Wave

Mia: Welcome back to ProductHunt Daily on Bri Radio. I'm Mia.

Milo: And I'm Milo. Today we've got a launch that helps builders price out construction jobs, a startup promising FaceTime that answers itself, and an open-source toolkit for running AI agents at scale.

Mia: Plus an interactive 3D encyclopedia, an AI-powered longevity dashboard, and an iPhone alarm that works with your AirPods. Lots to dig into.

Mia: A new browser-based construction workspace called CostLogic has launched on Product Hunt, built around an AI agent named Onyx that the makers say can act on the user's behalf. According to the listing, Onyx detects rooms on floor plans, prices out takeoffs, drafts invoices, and answers questions about jobs, letting a construction team go from blueprint to invoice in one connected workflow instead of stitching together five different tools. The pitch is that what used to take days can now take hours or even minutes. Behind that positioning is a concrete workflow: upload construction PDFs, and pages are named automatically while scale is read from the title block; measure directly on the drawing with linear, area, count, and angle tools, or let the Auto Room feature detect rooms and areas; build out line-item estimates with markup, waste, tax, and deposit controls; and turn an approved estimate into a branded PDF invoice in one click. The product also records payments, tracks outstanding invoices, and exports CSV for a bookkeeper or a QuickBooks import. The maker emphasizes that measured quantities flow straight into pricing and billing, so nothing gets retyped along the way. The intended user is construction professionals who currently move plans, estimates, and invoices between separate takeoff and accounting tools. In community discussion, co-founder Micah explained that he and his co-founder saw few construction tools harnessing AI the way other industries have, describing the alternatives as either software built back in 2005 or thin GPT wrappers prone to costly mistakes, and said CostLogic aims to strike a balance between those two extremes.

Milo: So what does that workflow actually look like in practice?

Mia: You're uploading construction PDFs, and the pages get named automatically with the scale read straight from the title block. Then you can either measure directly on the drawing using line, area, count, and angle tools, or let the Auto Room feature detect rooms and areas for you. From there you build out line-item estimates with markup, waste, tax, and deposit controls, and when the estimate gets approved, it turns into a branded PDF invoice in one click. The system also records payments, tracks outstanding invoices, and can export a CSV for your bookkeeper or for importing into QuickBooks.

Milo: And the key thing, from what the makers emphasize, is that measured quantities flow straight into pricing and billing so nothing gets retyped. That's really the part that would save someone sitting with a paper plan and a separate estimating tool.

Mia: Exactly. That retyping is exactly where errors creep in. So the intended user here is construction professionals who are currently shuffling plans, estimates, and invoices between separate takeoff and accounting tools.

Milo: And why do they think nobody's built this before?

Mia: In the community discussion, co-founder Micah said that he and his co-founder saw very few construction tools actually harnessing AI the way other industries already are. He described the current alternatives as either software essentially built back in 2005, or thin GPT wrappers that are prone to costly mistakes. So CostLogic says it's aiming to strike a balance between those two, and Onyx is how it's trying to do that.

Milo: A startup called Chert, which the co-founder describes as being in Y Combinator's P26 batch, has launched what it calls Vapi for FaceTime on Product Hunt — a service for building and deploying interactive AI video agents that can answer and place FaceTime calls, in their words, in a few lines of code. Co-founder Gary frames the product around the problem that every voice agent shipping today is blind, since voice agents can only handle what a customer can describe, so visual problems like which cable goes where or an error on a screen force a human to step in. Chert's agents can see what a caller shows the camera and respond in real time, which the maker says is the interface AI has been missing. Claimed use cases include remote support, field service, telehealth intake, guided onboarding, visual inspections, and customer success. Per the product page, deploying an agent involves three steps: configure a prompt defining the job, instructions, model, voice, and behavior; choose an avatar and framing; and publish the assistant to a provisioned FaceTime line. A browser preview lets builders test the prompt, voice, microphone, interruption behavior, avatar, and cleanup without placing a FaceTime call. The website says the control plane supports bounded inbound and outbound test workflows, with live execution provisioned and explicitly authorized, and accepting mode is not enabled by default. The FaceTime API is not generally available yet — Chert says it is onboarding the product through a controlled private preview while production media and line readiness are validated. A live test is available by calling a number that the team has posted.

Mia: And why FaceTime specifically? What's the problem they're solving?

Milo: The co-founder Gary frames it this way: every voice agent shipping today is blind. A voice-only agent can only handle what a customer can describe out loud. So visual problems — things like which cable goes where, or an error message showing up on a screen — instantly force a human to step in. Chert's agents can see what a caller shows the camera and respond in real time, which Gary says is the interface AI has been missing.

Mia: So the use cases that makes possible would be things like remote IT support, field service, telehealth intake, guiding someone through onboarding, visual inspections, customer success work.

Milo: Right, all the situations where you actually need to see what's in front of the person. And on the builder side, the process is three steps: you configure a prompt that defines the job, instructions, model, voice, and behavior; you pick an avatar and framing; and then you publish the assistant to a provisioned FaceTime line. There's also a browser preview so you can test the prompt, the voice, the microphone, how it handles interruptions, the avatar, and cleanup — all without placing an actual FaceTime call.

Mia: Is it fully live for anyone to use yet?

Milo: Not quite. The website says the control plane supports bounded inbound and outbound test workflows, with live execution provisioned and explicitly authorized, and accepting mode is not enabled by default. And the FaceTime API itself isn't generally available yet — Chert says it's onboarding through a controlled private preview while production media and line readiness are validated. There is a live test folks can try right now, though, by calling a number that's been published.

Mia: A new open-source project called HarnessRouter Community Edition just launched on Product Hunt, and its makers bill it as "OpenRouter, but for agent harnesses instead of models" — meaning one unified interface where you run multiple AI coding agents instead of juggling tools. It puts OpenAI's Codex, Anthropic's Claude Code, and Nous Research's Hermes behind a single API that handles sessions, streaming, files, artifacts, cancellation, and failure handling. The software is Apache 2.0 licensed and runs in one Docker container with Gateway, Runner, and Console bundled together, using local SQLite and a workspace volume — no managed database, external vault, or cloud account involved, and the makers state that credentials, state, and files never leave your infrastructure. The reasoning behind the tool, per the makers, is that they previously rebuilt the same backend every time they needed an agent runtime and found their self-built harness couldn't compete with the established players. After switching to off-the-shelf harnesses, they claim agent feature delivery time shrank from weeks to hours and agent quality improved dramatically — though these are founder claims, not independently verified numbers. Alongside the product itself, the team published the Unified Harness Protocol, or UHP, a public, versioned HTTP contract governing how applications talk to an agent harness. It standardizes harness selection and configuration, task execution, streaming events, sessions, file management, cancellation, error handling, and result delivery, and ships with OpenAPI 3.1 definitions, JSON Schema 2020-12, and 47 runnable conformance checks.

Milo: So what does that actually look like in practice?

Mia: "emph"So it puts several big-name agent harnesses — OpenAI's Codex, Anthropic's Claude Code, and Nous Research's Hermes — behind a single API. That one interface handles things like sessions, streaming output, file management, artifacts, cancellation, and what happens when something fails.

Milo: And it's meant to be self-hosted and self-contained, right?

Mia: Exactly. It's licensed under Apache 2.0, and it runs in a single Docker container that bundles the gateway, the runner, and the console together. It uses a local SQLite database and a workspace volume — no managed database, no external vault, no cloud account required.

Milo: That's a big part of the appeal for people who care about data staying in-house.

Mia: Yes — the makers state that credentials, state, and files never leave your own infrastructure. And they explain the motivation this way: every time they needed an agent runtime, they found themselves rebuilding the same backend from scratch. But their self-built harness struggled to keep up with what the big polished products already offer. Once they switched to off-the-shelf harnesses instead, they claim feature delivery went from weeks down to hours, and agent quality improved dramatically. Now, those are their own claims, not independently verified results.

Milo: And there's also a protocol that shipped alongside it.

Mia: Right — the Unified Harness Protocol, or UHP. It's a public, versioned HTTP contract that standardizes how an application talks to an agent harness. It covers how you pick and configure a harness, how you run a task, streaming events, sessions, file management, cancellation, error handling, and delivering results. It ships with OpenAPI definitions and 47 runnable conformance checks, so other tools can actually verify they speak the same protocol.

Milo: So the idea is, this isn't just a product — it's trying to become the standard way agent harnesses get integrated.

Mia: That's the ambition. A unified interface that works across the big agents, hosted entirely on your own machine.

Mia: Another Product Hunt launch takes a very different approach to education: Expeditione is an interactive 3D encyclopedia built by solo developer Aureon, where instead of reading about a subject, you step into handcrafted, explorable worlds. It runs directly in the browser with no login, ads, or tracking, and the first three expeditions are free forever, with free educator resources for classroom use. The flagship expedition is Ancient Egypt, which Aureon says took more than 1,000 hours of research, reconstruction, 3D modelling, animation, interaction design, writing, and optimization. It contains five handcrafted scenes: family life, quarrying, temple life, the Nile Delta, and the Pyramids. A notable engineering constraint — the landing page and the first four scenes are optimized to roughly 1MB, including all custom 3D models, audio tracks, textures, shaders, and code, with the stated goal of proving that immersive doesn't have to mean bloated. The other free expeditions are Layers of Soil, described on the site as a journey through five ancient layers beneath the surface, plus an upcoming Cell Biology expedition listed as TBA and described as the final foundational gift of the collection. Community reactions were positive but impression-based. One commenter specifically praised the quarrying scene as an odd and good choice, arguing that most Egypt material skips the labor of cutting and moving stone, and that giving labor its own scene teaches kids more than the monument does. Another asked how long a single scene takes to walk through — a question the launch discussion doesn't answer.

Milo: And it runs entirely in the browser?

Mia: Yes — no login, no ads, no tracking. The first three expeditions are free forever, and there are free educator resources for classroom use.

Milo: So what's the flagship expedition?

Mia: Ancient Egypt, which Aureon says took more than a thousand hours of research, reconstruction, 3D modeling, animation, interaction design, writing, and optimization. It's built around five handcrafted scenes: family life, quarrying, temple life, the Nile Delta, and the Pyramids.

Milo: And there's a technical constraint here that's unusual to hear about.

Mia: Definitely — the landing page and the first four scenes are optimized to roughly one megabyte, including all the custom 3D models, audio tracks, textures, shaders, and code. The stated goal is to prove that immersive content doesn't have to mean bloated content. So you're getting rich 3D worlds that still load quickly.

Milo: What are the other free expeditions?

Mia: There's Layers of Soil, described as a journey through five ancient layers beneath the surface, and an upcoming Cell Biology expedition that's listed as coming soon, described as the final foundational gift of the collection.

Milo: Have people had strong reactions to it?

Mia: Early reactions were positive but impression-based. One commenter specifically praised the quarrying scene as an odd but good choice — pointing out that most Egypt content skips the labor of cutting and moving stone, and that giving that work its own scene teaches kids more than the monument itself does. Another commenter asked how long a single scene takes to walk through, and the launch discussion doesn't actually answer that yet.

Mia: There's a new app on Product Hunt called Vidaya, formerly known as Vitality AI Health, and it's being pitched as an AI-driven longevity dashboard. The idea is to pull together all the scattered pieces of your health data—wearables, blood work, DNA results, nutrition, supplements, even air quality exposure and medical records from Epic—and turn them into a single Healthspan score, plus a personalized plan of what to do next.

Milo: And the founder, Kevin Amrelle, has a pretty personal story behind it. He says he built it after a winter bike race when his heart rate topped out at 120 beats per minute, and a blood pressure cuff flagged stage 2 hypertension. But none of his health apps caught the trend. His complaint is that your data lives everywhere—things like calorie trackers, genetic tests, lab results, the Apple Health data, even the EPA's air quality dashboard—and each one only gives you a fragment of the story.

Mia: So this is aimed at people who are tired of another dashboard and want the numbers converted into an actual next action. The maker argues the main difference from existing apps is that most consumer health apps combine one or two data categories, while Vidaya claims to unify every category a user generates, including medical-grade sources like Epic FHIR, Labcorp, Quest, 23andMe, and AncestryDNA.

Milo: Unsurprisingly, with medical records involved, privacy is a headline. The company says it was built to be HIPAA-compliant from day one, with guidance from a virtual CISO named Denis Galkin, and that the cross-source correlation engine is the subject of a pending patent application.

Mia: There's a chat interface called Vaya that answers natural-language questions across all your connected data. The example they give is asking how your sleep changed after starting Lexapro, and getting a grounded answer in around ten seconds.

Milo: So the real pitch here is less about giving you one more number and more about doing the integration work for you—finding the connection between, say, a supplement, a lab result, and how you actually slept. Whether it lives up to that depends on how well those correlations hold up, but it's a clear step beyond just aggregating a few fitness stats.

Mia: Another wellness launch on Product Hunt this week is AirAlarm, a straightforward iPhone alarm app that pairs with AirPods. Instead of one fixed alarm time, you pick a wake window, and the app aims to get you out of bed more gently—without requiring an Apple Watch, an account, ads, analytics, or a cloud sleep profile.

Milo: It's built in SwiftUI and uses Apple's AlarmKit to handle the actual system-level alarm scheduling. Sleep records and your preferences all stay on the iPhone, which fits the no-account, no-tracking angle. It currently supports English, Simplified Chinese, and Japanese.

Mia: The flow is interesting. You choose your wake window, fall asleep wearing the AirPods while listening to built-in sounds—rain, ocean, forest, fan, or white noise—or audio from another app. Then AirAlarm uses that bedtime flow to schedule the end of a 90-minute sleep cycle to land somewhere inside your window.

Milo: And the maker is pretty honest about the framing—calling it "experience first, not a lesson in sleep science," while noting the sleep-cycle timing is still working under the hood. The FAQ is careful to describe it as an alarm and bedtime companion, not a medical device or diagnostic sleep tracker. It's designed for the AirPods Pro, AirPods 3, and newer models you can comfortably wear while falling asleep.

Mia: So the target audience is clear: iPhone users who want a less abrupt wake-up and are comfortable sleeping with their AirPods in, without adding another wearable, account, or cloud profile to the routine.

Milo: And the maker is inviting community feedback on a couple of open questions—specifically whether choosing a wake window feels more natural than a fixed alarm time. It's an early-stage app, but the privacy-first, no-hardware-required approach is what sets it apart in the crowded alarm space.

Mia: And that wraps up today's episode. Thank you so much for listening and for sticking with us through all of it.

Milo: We appreciate every one of you tuning in. Until next time, take care of yourselves, and we'll catch you right back here soon.