0811 | Product Hunt roundup: Paritok, Portfolio Lab, Prime Agent, SecondBrain Note

Mia: Hey there, and welcome back to ProductHunt Daily on Bri. I'm Mia, and as always, I'm joined by my co-host Milo. Hi Milo!

Milo: Hey Mia! Great to be here. We've got a really packed show today, so let's get right into it.

Mia: Absolutely. We're covering everything from a new kind of AI investing platform to open-source coding tools, smart voice recorders, and even a tiny task manager that lives in your Mac's menu bar.

Milo: Plus a longevity dashboard, a team-based AI discussion tool, and a company building custom benchmarks for real-world tasks. Lots to unpack, so stay with us.

Mia: Let's start with Paritok, a launch from engineers Jiayu and Luzhuo that was hunted by Chris Messina on Product Hunt. It's basically a smart gateway that sits between a coding agent and its model API, and its whole job is to make those conversations cheaper without throwing anything away.

Milo: And the savings claim is pretty striking — the makers say you can shave about 25 percent off your token bill from the very first turn, and in long sessions where the context window is getting full, they say you can cut more than 85 percent. Essentially, you can run roughly three times as many turns inside the same context window.

Mia: The magic comes from a small code-native model trained on tens of thousands of real agent work sessions. It's doing a few things at once. Tool schemas — the instructions describing what tools the agent can call — shrink dramatically each turn, and that part is done locally, on CPU only, with no model involved. File reads and tool output get compressed to about a quarter of their original size while keeping the identifiers, paths, and errors intact. And once you set a context budget, older history gets summarized once that budget fills up.

Milo: I love their tagline — "lossy on the wire, recoverable when it counts." Nothing is actually discarded. The original bytes are recoverable locally through a read original function, so you get the compression benefit on the wire but the full fidelity stays available when you need it.

Mia: Integration is a single environment variable pointing at a local address, and their site lists compatibility with Claude Code, Cursor, Codex, OpenHands, and anything that speaks OpenAI's protocol. Self-hosting is free, and they also offer an API key. Their session math shows a single turn going from 40,000 tokens down to 12,000 — that's the 25 percent saving — and a longer 20-turn session compressing from 250,000 tokens down to 95,000, about a 63 percent cut. Their stated ceilings run higher for tool-heavy setups, and their site calculator throws out a roughly six and a half thousand dollars saved per year estimate.

Milo: Next up, Portfolio Lab, which launches as an "AI investing, done responsibly" platform for self-directed investors who want AI-built systematic strategies without having to build the system themselves.

Mia: The founder, Rich Sun, comes from a hedge fund background, and he tells a cautionary story behind the launch. They ran an experiment where Claude built nearly 1,300 strategies, and even he needed days of line-by-line code auditing to find subtle errors that were inflating results. After corrections, almost all those strategies lost their edge. His point is that LLMs reason in language, not numbers, so prompting an agent and trusting the output is essentially a coin flip.

Milo: So the remediation is a three-step workflow. First you Build — you set a goal and their proprietary quantitative models construct systematic strategies. Then Validate — every strategy gets tested on unseen data and runs live in paper trading before real money moves. Finally Deploy — you connect Claude, ChatGPT, or any MCP agent to trade in your own brokerage account.

Mia: The site claims the company is SEC registered and a SIPC member, and says the system is powered by seven AI models, more than 200 predictors, over a billion data points, and more than 35 years of market history, drawing on fundamental, technical, estimate, macroeconomic, and alternative data categories.

Milo: One build generates multiple strategies tuned to different objectives — the examples shown are optimized for return, for Sharpe ratio, and for minimum drawdown, though these are explicitly labeled simulated and not a guarantee of future performance. There are a few strategy types on offer, from a risk-aware hedged long-only setup to tactical long and short approaches. So the pitch is: let the machines draft the strategies, but put them through rigorous validation before a cent moves.

Mia: Then we have Prime Agent, an open-source, self-improving coding harness from Prime Intellect. They built it around two abstractions — a recursive language model and what they call a Continual Harness — and the headline number is that with Opus 5 it scores 95.5 percent on the ARC-AGI-3 benchmark, which they say surpasses the reported human expert baseline.

Milo: The key design difference is how it treats the context window. Conventional harnesses rely on fixed tool-calling schemas and static, hand-engineered sub-agents, prompts, skills, and memory. Prime Agent instead treats context as a variable, and sub-agent delegation becomes function calls inside a persistent Python REPL — and that REPL is the agent's only tool. The Continual Harness lets the agent create, update, and even delete its own prompts, skills, memory, and sub-agents, and agent-to-agent messaging lets you orchestrate across sub-agents and separate sessions.

Mia: So it's genuinely self-modifying rather than a fixed toolkit. On the engineering side, sessions run in a background daemon over a local socket, with attach and detach support, and each root session tree runs in a recoverable worker process — crashes get restored from session logs and kernel snapshots. History is append-only, so nothing gets silently overwritten, and it supports branching, forking, and cloning. Sub-agents idle out after 30 minutes and reload when you address them again.

Milo: Prime Intellect positions it for three uses: a general coding assistant, a runtime for long-horizon autonomous evaluation, and a collaborator for research and autoresearch. And they say it works immediately with modern open and closed frontier models, so you're not locked into one provider. It's a compelling pitch — an agent that literally rewrites its own operating instructions and keeps full recoverable history as it does.

Milo: Finally, SecondBrain Note from Genspark, which is a MagSafe, card-thin AI voice recorder — a genuine hardware and AI combo that turns real-world conversations into structured, searchable notes automatically.

Mia: The workflow is deliberately frictionless. One press-and-hold of two seconds starts continuous recording on a standalone device with no settings to configure, and it can capture up to 35 hours. The Genspark app then produces AI summaries and transcripts saved straight into your SecondBrain. The launch description says it picks up voices from more than five meters away and never misses a word.

Milo: The hardware itself is remarkably small — roughly the size of a credit card and light enough to forget you're carrying it, in aircraft-grade aluminum with a single button and a single light. It has a burst of storage, a battery the maker rates for those 35 hours of continuous recording, and Bluetooth 6.0. Audio capture uses four microphones plus a bone-conduction mic, and there's a vibration sensor that handles phone-call capture. The MagSafe wallet accessory has three modes — it works as a card wallet holding two cards plus the device, snaps onto an iPhone, or props up as a kickstand.

Mia: Who's it for? The community description targets professionals, founders, researchers, and students who live in meetings, interviews, lectures, and client calls, with voice memos and phone-call recording as the headline use cases. The differentiators they lean on are the dedicated hardware, the offline-first operation where recordings sync later and auto-generate summaries, and the deep integration with Genspark's SecondBrain and related tools for a unified notes workflow.

Milo: So the pitch is that your meetings and conversations stop being things you scramble to reconstruct — the recorder captures them, the AI structures them, and they land in your notes system automatically. That closes out our look at four launches turning raw AI interaction into something people can actually use.

Mia: Let's start with a launch that's pitching itself as a new way to do a big slice of the work behind AI-assisted product development. It's called Remix, now live on Product Hunt, and the maker frames it as "Figma, but on your production app." The idea is that any team member — a designer, a product person, not just an engineer — can spin up a safe, sandboxed copy of the actual product, then modify it just by describing what they want. No project setup, no risk to the live app.

Milo: So the pitch is covering the parts the maker argues are still hard about AI-assisted development: getting the project stood up, running it, previewing it, keeping it in sync with the main codebase, reviewing the work, and ultimately shipping it. Each of these remixes gets its own sandbox and a live preview, gets checked against the team's design system, security, and compliance rules, and can be merged with other variants or opened up as a pull request straight to GitHub.

Mia: And every single prompt is recorded, so a reviewer can see exactly how a change was built. There's also a live link so anyone can test before the code ever touches the main branch. The maker frames engineering's role as shifting toward architecture and approval — engineers see the prompts, review a clean pull request, and keep the final say.

Milo: For evidence of impact, which is self-reported and not independently verified, the maker cites one early team where a designer reworked the app's empty states — something that had apparently sat in the backlog for months — prompted it, previewed it, and shipped it the same day through a normal pull-request review. One commenter praised the concept as accelerating the path from idea to production while giving engineers far more visibility into the whole process before the pull request even lands. The big open question from the discussion: someone asked what the sandbox actually runs against — whether it's a snapshot of production data, which the maker hasn't fully answered yet.

Mia: Next up, a tiny tool for people who want to capture tasks without leaving whatever they're doing. It's called Gutta, and it lives in the Mac menu bar, described by its developer as a first public release. The killer shortcut is Command-Shift-Space, which pops open a quick panel, and you just type tasks the way you'd say them — something like "send the invoice Friday at ten thirty," or "book tickets tomorrow; email Alex at four PM."

Milo: So it parses natural language. Semicolon-separated input turns into multiple tasks at once, dates and times get picked up, and tasks without any scheduling detail default to today. Words like morning, afternoon, or evening add default times, and an exact time always wins over those. Everything is stored locally on the Mac using SwiftData, so there's no Gutta account, no subscription, no tracking, and no developer-owned cloud.

Mia: But there is optional sync, and it's a clever approach: you pick a folder in your own iCloud Drive, Dropbox, or OneDrive, and Gutta exchanges one small JSON record per task in a dedicated subfolder, rather than syncing the live database. It merges added, completed, changed, and deleted tasks while the app is running, offline changes catch up when the folder service reconnects, and if the same task is edited concurrently on two devices, the most recently edited copy wins.

Milo: The developer notes you only have to pick that folder once per Mac, because the App Sandbox blocks silent access to arbitrary iCloud folders. And reminders are local, based on timed tasks, with a global default lead time you can set — anywhere from fifteen, twenty, or thirty minutes, up to an hour or two hours, or off entirely — plus per-task overrides on top of that.

Mia: Want some tougher feedback than a single chatbot can give? There's a Product Hunt launch for that too. It's called AI Group Call, and the maker, Tash, frames it as an alternative to consulting one chatbot: you type a goal, and you get a live voice call with six AI participants cast for that goal, each with a name, role, and personality. They speak one at a time, build on each other's points, disagree when they should, and stop the moment you speak.

Milo: And that barge-in is the hard part. The maker says that over a phone speaker, the microphone also hears the agents, so the app has to handle echo and use raw audio voice detection to make interrupting feel like a real call. Just to be clear — every participant is AI, there are no humans on the call. The idea is to give you "the room": a skeptic to poke holes in a pitch, a strategist to reframe it, and someone playing the customer who has to sign off. The example casts include a host, a skeptic, a strategist, an ops lead, a researcher, and a customer, with at least one participant who will push back.

Mia: On the capabilities side, you get a live captioned transcript saved to history, one-tap summaries with key points and action items, the option to rejoin any room with the same cast, and the ability to edit any agent's name, role, bio, and traits mid-call. You can also pause to freeze your minutes, and flag an AI response from any transcript.

Milo: The example transcript on the site shows the group pressure-testing a pitch and agreeing the retention story should lead — with a key point of "lead with 68% week-four retention" and a benchmark of a 24% category median. That's illustrative sample output, not a verified result. On availability, Android is live today, and iOS is built and waiting on submission. New accounts get one free minute with no card, and then monthly bundles of minutes kick in after that.

Mia: Finally, a launch aimed at anyone trying to make sense of their own health data. It's Vidaya, previously known as Vitality AI Health — an AI longevity dashboard that pulls in wearable data, lab results, DNA, nutrition, supplements, environmental exposure, and Epic medical records, then turns it all into a Healthspan score and a personalized longevity plan. Founder and CEO Kevin Amrelle says the product exists because health data is fragmented across tools like MyFitnessPal, Quest PDFs, Ancestry, Apple Health, and CVS, none of which connect the signals.

Milo: And he cites a personal example: a stage 2 hypertension diagnosis that no app caught. The service includes Vaya Chat, an AI coach that answers natural-language questions across all your data. The example given is asking how your sleep changed after starting Lexington — wait, after starting Lexapro — and getting a grounded answer in about ten seconds, unified across more than sixty integrated data sources.

Mia: The founder claims most consumer health apps aggregate only one or two data categories, whereas Vidaya pulls in medical-grade sources like Epic's health record standard, Labcorp, Quest, and the DNA testing services, was built HIPAA-compliant from day one, and uses a cross-source correlation engine that's the subject of a pending patent. Those are the maker's claims, not independently verified.

Milo: On safety, the co-founder and head of AI describes self-reported testing: more than 120 chat-quality iterations tuned against a 32-question health stress suite covering lab trends, emergency symptoms, prescription requests, and hallucination traps, plus ongoing production monitoring. A May 15 live production audit reportedly showed zero hallucinations across the 32 questions and full marks on correctness. Again, that's self-reported. So the through-line here is a big ambition — unifying all your health data into one score and one plan — but also a heavy reliance on the makers' own claims about accuracy and safety.

Mia: A new tool called Oqoqo just launched on Product Hunt, pitched as the easiest way to build evals and custom benchmarks for real-world tasks. Co-founder Haritha frames the underlying problem pretty clearly: new models and new benchmarks seem to drop every week, but most benchmarks run in carefully curated environments. They don't actually tell a product team whether an agent can discover their product and use it on a real user task.

Milo: So what does Oqoqo actually let you do with that?

Mia: You define a task as a simple prompt a user might give an agent, like asking it to integrate Supabase into a web app to store user signups. Then you specify what to test, say the Supabase SDK or its API, and define success criteria, such as requiring row-level security to be set up. The infrastructure spins up isolated sandboxes, runs the task against whatever agents you choose, and catalogs every step, including tool calls, retries, and discovery loops, while tracking token consumption and cost. It evaluates pass or fail against your own criteria.

Milo: And the whole point is measuring how agent friendly your product actually is.

Mia: Exactly. Oqoqo says you can gauge how agent friendly your product surfaces are against the major coding agents, including Codex, Claude Code, Cursor, and GitHub Copilot. You can regression-test agent-facing interfaces like MCP, CLI, and SDKs, build and share custom benchmarks, compare models and harnesses for domain-specific tasks, and even trigger these experiments from your CI pipeline. The site shows a sample evaluation on an Apify Actor, running tasks like recommending a budget mechanical keyboard against Claude Code and Codex, and it compared a version with an added treatment against the baseline.

Milo: So the takeaway is that instead of relying on generic benchmarks that may not reflect real usage, teams can now measure agents against their own actual product workflows, and apparently even see a clear difference when they improve how agent friendly an interface is. Pricing wasn't shared in the launch material, but the pitch is that this closes the gap between benchmark results and whether an agent can genuinely use your product on real tasks.

Mia: That's our snapshot of today's Product Hunt launches — from Paritok's non-destructive compressed storage to Oqoqo making custom benchmarks easy, and everything in between. Plenty to dig into.

Milo: Absolutely. Whether it's AI investing done responsibly, a self-improving coding harness, or a tiny keyboard-first task list in your menu bar, there's something for every kind of builder.

Mia: And a card-thin AI voice recorder, plus what's being described as Figma for your production app. Thanks for listening — catch you next time.