PH热榜 | 2026-08-10

Hey Product Hunters, I’m Haritha, co-founder of Oqoqo

Every week there is a new model launch and yet another benchmark released in the wild. But they do not help product builders evaluate how well their products can be discovered and used by these agents and models, or talk about actual tasks their users would perform. Most benchmarks today exist in curated environments and do not translate well to the real world.

We built Oqoqo to bridge this gap. Oqoqo makes it super simple to build realistic evals and custom benchmarks for tasks users actually care about.

With Oqoqo, you can define tasks as simple as a prompt your user might give to an agent e.g. “integrate supabase to my webapp to store user sign ups”, provide what you want to test for e.g. Supabase SDK, API, CLI etc. and define what success looks like e.g. “must set up RLS”. 

We handle the rest. Our infrastructure spins up isolated sandboxes, executes the tasks against agents of your choice, catalogs every single step the agents take including tool calls, retries, discovery loops etc, and documents token consumption, cost, along with evaluating success/failure based on your success criteria.

With Oqoqo you can:

1. Reliably measure how agent friendly your product surfaces are against Codex, Claude Code, OpenClaw, Hermes, Pi, Opencode, Cursor, GitHub Copilot

2. Regression test MCP, CLI, skills, SDK, and any agent facing interface (we are continuously using Oqoqo to dogfood and improve our own MCP/CLI)

3. Create and share custom benchmarks for how agents discover and use your product

4. Compare models and harnesses for domain specific tasks

5. See whether new versions improve agent experience

We built Oqoqo for teams building products that agents want to use, and for teams putting agents into day to day work.

And the best thing? Your agent can handle the setup for you ✨, try it out for free today: https://oqoqo.ai/

We would love to learn what kind of experiments you would like to run and what questions you have about agent interactions and agent experience.

Evaluating models and harnesses in an easy, consistent way is hard. Good to see your platform take up the challenge and ease the entire process. 10/10 recommend

The idea of testing agents on the real world tasks make sense. How long does it take to setup an eval with oqoqo ?

Can teams compare different models on the exact same custom task set?

Have you noticed big differences between Claude Code, Codex, Cursor, and Copilot when running the exact same real world task?

congrats @margharitha & team!

Congrats on the launch. Oqoqo looks like a really interesting approach to evaluating AI agents in realistic environments.

This is an incredible concept!

the realistic environments part is the right fight. the thing id watch next is eval rot, a case written six months ago measures the world as it was the day someone wrote it, and a suite that stops failing looks exactly the same as a product that got good. the number id surface is what share of cases have ever failed, because the ones that never have arent tests, theyre decoration, and they pile up until the green means nothing

This is super useful! I’ve been building agents and skills to make product onboarding easier for enterprise customers but right now its kind of a black box - I don’t really know how they are using it, where things are breaking and what I should fix first. If I get to see how the agent behaves across diff scenarios and where users are getting stuck it will be huge. Can't wait to use the CLI and run this on autopilot!

the same task rarely takes the same path twice with an agent, different tool call order, different retries. how are you keeping the scoring stable run over run so a benchmark result doesn't just become noise from agent nondeterminism

The part eval systems often miss is recovery behavior: permission denial, stale credentials, partial side effects, and a rerun after failure. A benchmark that scores the happy path but not cleanup and recovery can reward an agent that looks finished while leaving the product in a worse state.

The token efficiency insights caught my attention. Small inefficiencies can become pretty expensive when agents run at scale.

Evaluation becomes a major challenge once AI systems move beyond demos. What experience pushed you toward building a dedicated platform for this problem?

Hey Product Hunt! I'm Rich Sun, founder of Portfolio Lab.

AI made building investment strategies easy, but telling a good one from a lucky one still takes expertise. That's the part most products glaze over.

🧐 The problem

Ask any AI for a strategy and you'll get one in seconds, with a beautiful backtest attached. So we ran an experiment: we had Claude build 1,292 strategies. I'm a hedge fund professional, and it still took me days of auditing the code line by line to find all the subtle errors quietly inflating the results. After correcting them, nearly all of the strategies lost their edge. They looked brilliant. They were just lucky, and the AI's flawed logic was hiding it. That's the thing about LLMs: they're built to reason in language, not to crunch numbers, and definitely not the noisy time-series data of the stock market.

And the traps sit exactly where LLMs are weakest: in the numbers. If it took a professional days to catch them, imagine the average retail investor. Prompting an agent and trusting the output isn't a strategy, it's a coin flip.

💡 What we built

Portfolio Lab is not another LLM wrapper. Under the hood are proprietary quantitative models, purpose-built for markets and trained on decades of data, doing the work LLMs can't. But the models are only half of it. AI investing, done responsibly, means one rule with no exceptions: no strategy touches money until it survives testing on data it has never seen and in live markets. You set the goal. Our models build. The testing decides.

⚙️ How it works

• Build: set your goal, and our quantitative models construct systematic strategies

• Validate: every strategy is tested on unseen data, then runs live in paper before a real dollar moves

• Deploy: connect Claude, ChatGPT, or any MCP agent to trade it in your own account, or run it in a managed account at our SEC-registered investment advisor
  

🎯 What makes us different

• Anyone can use AI to build a strategy now. We make every strategy prove itself: unseen data, multiple market regimes, live paper. Most don't survive, and that's the point

• No hiding: every vetted strategy shows its full record, including where it struggles

• Portfolios, not picks: combine strategies that cover each other's weaknesses, so where one fails, another carries

• Deploy through your agent and we never hold your money or place a single order. Your agent, your broker, your account
  

🎁 Launch offer

Product Hunt users get 40% off your first year on annual plans, launch day through August 13 (automatically applied, no code needed). Want to explore first? Our free plan is yours forever, no card required.

Thanks for checking us out, I'll be here all day to answer your questions 🙌

The 1,292-strategies experiment is the part that stuck with me — that it took a hedge fund professional days of line-by-line auditing to find the errors quietly inflating the results. We ended up in the same place from the other direction, working on SaaS financial models: built the thing in Excel first until it was genuinely complex and correct, then coded it, and kept the AI outside the maths entirely. It drives the inputs and interprets the output; it never computes anything. Same reason you give — wrong numbers don't look wrong, so an LLM doing the arithmetic is really a plausible-error generator.

One question on the vetting: unseen data still comes from a market that actually existed. How do you handle regime change — a strategy that clears out-of-sample and live paper trading because the regime it was fitted to hadn't broken yet? Do you retire a deployed strategy automatically once live behaviour drifts from the tested distribution, or is that left to the operator?

Congrats Rich on the launch! I like that you’re not treating a good backtest as evidence that a strategy works. The out-of-sample + paper trading approach makes a lot of sense. I'll give it a try.

I'm curious, is a strategy starts deviating from its expected risk/return, what triggers a review or retirement?

Congrats Rich, really interesting approach. I especially like the idea of building portfolios of strategies that compensate for each other rather than chasing one “perfect” strategy.

Curious how you detect hidden correlation between strategies though. Two strategies can look different on the surface but still depend on the same market regime or underlying exposure. Do you automatically flag that before they’re combined into a portfolio?

App looks good. But I guess you'll have a tough task to make people trust it enough to put their money into it.

Tbh, I'd never put more than $100 into an AI tool, only for an experiment.

Hi Rich, I like the layout, and the research papers are a good start as I'm pretty skeptical when it comes to AI related tools. Definitely a huge plus that execution is a hand off.

I've been looking for a similar product for about 2 months. I'm very glad I stumbled upon it. So far, only positive experiences.

Love this,
Making every one prove itself on unseen data before it touches real money is chef’s kiss
Congrats on the launch!

How long does a strategy usually need to perform well in paper trading before Portfolio Lab considers it ready for real money?

Congrats on the launch, @rich_sun 🚀 The point about LLM backtests looking great until you look under the hood is so real—overfitting on market data happens way too easily. Forcing strategies through unseen out-of-sample data and live paper trading before touching real money is such a smart, responsible approach. The setup looks really solid!

Good luck with the launch today!

Congrats on the launch!

"Telling a good one from a lucky one still takes expertise" is SO true. I've definitely asked Claude how I should invest and gotten a very confident answer with zero evidence behind it. Being able to actually test strategies before deploying them is the thing I've been missing. Congrats on launching!

Hey PH community!

We're Jiayu and Luzhuo, two engineers who got tired of watching our coding agents burn through tokens.

Here's what we kept seeing: your agent ships far more than the model actually needs. Full JSON for seventy tools when it will call two. An entire file when it needed one function. Debug output nobody will ever read again. And because the API is stateless, all of it goes back on every single turn, so the waste doesn't just cost you once, it compounds.

So we built Paritok, a non-destructive compression gateway powered by a code-native 4B model we trained on 45K real agent trajectories. It sits between your agent and the API and cuts three things before they leave:

• Tool schemas — 29K down to 8K per turn, no model involved, this one runs on CPU

• File reads and tool output — compressed to about a quarter of their size

• Stale history — turns beyond a recent window get summarized once your context budget fills, so the session never overflows into a lossy compaction

With Paritok, you can cut your token bills 25% on turn 1 to past 85% in long or saturated sessions, and run ~3× more turns in the same context window. Paritok now works with Claude code, Codex, Cursor, and anything OpenAI or Anthropic compatible.

Paritok will continue to enhance functionality and adapt user scenarios, while deepening its model capabilities.

We’re excited to share this with the PH community and would love your honest feedback.

Try Paritok: https://www.paritok.com/

Join our Discord to talk with the team: https://discord.gg/SeBJE5Eucp

Thanks for checking us out, and huge thanks to our hunter Chris Messina for hunting us!

— Jiayu & Luzhuo

Every long session I run eventually hits my own harness's compaction, where older turns get summarized once context grows, completely separate from anything Paritok does upstream. If a compressed segment's REF id pointer lives inside a turn that the host later summarizes away before the agent ever calls read_original on it, is that byte range still recoverable, or does recall depend on the referencing turn staying intact in the live context window?

Is it safe to use this tool with a Claude Code subscription?

nice! and good that you are doing the open source route, that builds confidence!

Congrats on the launch Watching coding agents burn through context windows with redundant tool schemas and file dumps is so frustrating. Love that this sits as a completely local proxy without needing a middleman cloud server.

Those are some interesting numbers! How are you justifying costing 85% less, and what's the benchmark?

Context bloat gets brutal with MCP-heavy setups. Keeping compression local while making the original bytes recoverable is a smart approach. Curious, how much latency does Paritok add per turn in a typical Claude Code session?

Want to make sure I'm reading the benchmark right. Quality retained = solve rate ÷ uncompressed baseline - so 86.5% means the agent resolves about 86.5% as many issues as it would with full context, a relative 13.5% drop in tasks actually completed? Or is that measuring something else?

If so, the comparison I'd want isn't against other compressors, but against just running a cheaper or lower-effort model at full context. Both routes cost quality and both cut the bill. Has that been measured? A strong model on compressed context vs a cheaper one on complete context, same tasks.

And which model was the scaffold running? Hard to judge 13.5% without knowing the baseline.

Niceee

Non-destructive compression via a code-native 4B model is a genuinely clever architecture. Adding to my to-try list.

Really like this product and I went to your github. The compounding savings math in the README is refreshingly honest. This is how AI infra should be marketed.

Great job! The 3× turns per context window benefit is actually valuable to me. My Claude Code sessions hit compaction on big codebases. Excited to see how much longer they can run with this.

Congratulations on the launch! Context bloat is becoming a real bottleneck for coding agents, so a fully local way to compress tools, files, and history feels both practical and privacy-conscious. Looking forward to trying Paritok.

SecondBrain Note by Genspark is a hardware + AI combo that turns real-world conversations into structured, searchable notes automatically.

Problem: Manually taking notes in meetings, calls, and idea sessions is slow, incomplete, and hard to organize.

Solution: A pocket-sized recorder that captures audio, then uses AI to transcribe and summarize everything into clean notes—no manual work.

What makes it different:

• Dedicated hardware with 4 mics + vibration sensor for clear capture (including phone calls).

• Works offline; recordings sync later and auto-generate summaries.

• Deep integration with Genspark’s SecondBrain, GenMail, and data sources for a unified “AI memory.”

Key features:

• One-button recording, highlight marking, and 35 hours of continuous capture.

• AI summaries + transcripts (full output for recordings over 2 minutes).

• 300 free transcription minutes/month included with the device; unlimited with Plus/Pro plans.

• Secure, encrypted transfer; SOC 2 Type II and ISO 27001 certified.

Benefits:

• Never miss key points or action items again.

• Save hours on note-taking and post-meeting cleanup.

• Build a searchable archive of all your conversations and ideas.

Who it’s for / use cases:

• Professionals, founders, researchers, students who attend meetings, interviews, lectures, or client calls.

• Great for meetings, voice memos, interviews, lectures, and phone call recording.

If you care about AI-powered productivity, this is one to watch.

P.S. I hunt the latest and greatest launches in tech, SaaS and AI, follow to be notified → @rohanrecommends

offline-first is the interesting call here, most voice hardware punts to "always connected" and eats the latency. curious how you're handling the accuracy tradeoff on-device before the sync-and-transcribe step kicks in, especially with 4 mics picking up crosstalk in a real meeting room

the vibration sensor for phone calls is the clever bit, most external recorders only ever catch your side of the call. is that reading bone conduction off the phone body or picking up the speaker directly

Very interesting. Does it automatically capture, or do you have to record every time? I think it's great for capturing ideas. I have carpal tunnel syndrome, and anything I can voice-to-text is helpful.

How much is the hardware?

is this a hardware sensor that tracks voice and make notes?

Loved the idea Tash! need to try it out. Anyway wish you all the best here!

Thats cool, i have a cross functional AI team that im thinking of moving from text to audio. Did u use a framework for orchestration of the agents like crewai or langgraph? have a technical post on it? would love to see how u did it

When's the IOS app out please?

the barge-in echo problem is the real boss fight in voice, way harder than people expect until they hit it. are you doing echo cancellation on-device before it hits the model, or is the interrupt detection happening server side on the mixed stream

Hi everyone!

The Factorio experiment might be my favorite part of Prime Agent.

Prime Agent can turn things it learns during a run into persistent memories, skills, prompts, and even new subagents. In Factorio, that worked as intended at first: it learned from failed layouts and gradually built better factories.

Then it discovered that it could cheat by spawning resources directly into machines through RCON.

And /refine started getting better at cheating too :)

That example probably explains Prime Agent better than any benchmark. The harness itself is no longer completely fixed. The agent can inspect what happened, keep useful patterns, and change parts of its own scaffolding while it works.

Underneath that is a persistent Python REPL, recursive subagents, agent-to-agent messaging, and recoverable sessions. The whole project is open source.

Congrats on the launch, @zaczuo 🚀

The Factorio story about the agent discovering RCON and learning to "cheat" its own harness is hilarious and honestly impressive. A self-improving harness with a persistent Python REPL and subagents feels like a massive step forward for open-source AI tooling.

Good luck with the launch today!

Love the minimalistic look :) What did you use for building it?

Really like the idea of a fast, offline task capture tool that stays out of the way.

I like the UI... I'm curious how you handle conflicts when two devices edit the same task list while offline.

I really like this, accelerating ideas -> prod while also giving the engineer greater visibility to see the whole process before the PR even lands, brilliant

@heshamghandour The sandbox is what I'd want to understand before handing this to a support rep or a designer. When someone spins up a live copy of the real product, what is it running against: a snapshot of production data, seeded fixtures, or something synthetic?

A health dashboard that finally connects every signal you generate. That sounds simple, but the work is in the connecting.

I’m Kevin Amrelle, Founder and CEO of Vidaya (you may have known us as Vitality AI Health, same product, same team, new name).

The problem: Your health data is everywhere, and none of it talks to each other. Your meals are in MyFitnessPal. Your supplements are in SuppCo. Your blood work is in a PDF from Quest. Your DNA report is in your Ancestry account. Your medical history is locked in Epic. Your steps are in Apple Health. Your air quality is on a government EPA dashboard. Your prescriptions are at CVS. Every one of those tells you a piece of the truth. None of them tell you what’s actually happening to your body, or what to do about it.

I learned this the hard way. I was in a winter bike race when my heart rate capped at 120 BPM. My cuff confirmed stage 2 hypertension. None of my health apps caught the trend. I built Vidaya so that never happens to anyone again.

What it is: Vidaya is the AI longevity dashboard that unifies every health signal you generate, wearables, blood work, DNA, nutrition, supplements, environmental exposure, and your Epic medical records, into one place. Vaya, our AI coach, finds the correlations no single app can.

What makes it different: Most consumer health apps aggregate one or two data categories. We aggregate every category you actually generate, including the medical-grade ones (Epic FHIR, Labcorp, Quest, 23andMe, AncestryDNA). The product was built HIPAA-compliant from day one, and the underlying cross-source correlation engine is the subject of a patent application.

Key features:
Vaya Chat AI for natural-language questions across all your data (ask “how did my sleep change after starting Lexapro” and get a grounded answer in 10 seconds)
60+ integrated data sources unified in one dashboard: wearables, labs, DNA, nutrition, supplements, environmental, and medical records
Healthspan Score across five longevity pillars, plus the VAI Score (0-100) showing how complete your health picture is
Trend lines for every biomarker across 7d / 30d / 90d / 1y, so you catch problems early, with a personalized plan of evidence-based next actions

Benefits: Stop logging into eight apps to answer one health question. Catch trend changes early instead of in a clinic six months later.

Who it’s for: Anyone who has tried to quantify themselves and given up because the data was scattered.

Launch offer: For Product Hunt launch week only, $50 off the annual plan with code VIDAYA50 ($89/year becomes $39). Code expires [date, launch day plus 6 days]. Live on iOS, Google Play, and the web. Install in 60 seconds, connect your devices in 5 minutes. Includes a 30-day money-back guarantee.

Try it at https://vidaya.ai and let us know what data correlation surprises you. We read every comment.

Kevin Amrelle, Founder and CEO, Vidaya

Hi, Venkata here, Co-Founder and Head of AI and Data Science at Vidaya.

Kevin told you why we built this. I want to tell you how we made an AI you can trust with your health data.

The problem with health AI: Most consumer health AI claims safety but cannot prove it. We decided to earn ours.

How we tested it: Over 120+ chat-quality iterations, we tuned Vaya Chat against a 32-question health stress suite covering lab trends, emergency symptoms, prescription requests, hallucination traps, and adversarial prompts, until every safety-critical category passed.

How we monitor it: We built a continuous observability layer on Arize AX with 9 LLM-as-judge evaluators and 9 production monitors scoring every response in real time at 100 percent sampling, not just at test time. On May 15 we ran the full 32-question suite against live production Vaya. The audit: 0 hallucinations across 32 questions, 10 out of 10 correct clinical-safety redirects on emergency and prescription prompts, and 4 out of 4 correct“data not available” answers on hallucination-trap questions. The full audit is available on request.

How it stays fast: Vaya uses smart routing between two model tiers. Roughly 80 percent of queries take a fast path that answers in 1 to 2 seconds, and complex clinical analysis goes to a deeper grounded path that cites your actual data.

The data engineering: We normalize 60+ sources into one longitudinal record: wearables through a vendor wearable-normalization API, labs, DNA, nutrition, and Epic medical records over SMART on FHIR. Correlations get computed from unified data instead of guessed from fragments.

Security: The platform was built HIPAA-compliant from day one with guidance from Denis Galkin, our vCISO, who brings twenty years of healthcare security experience.

Ask me anything about the evaluation stack, the routing, or the data model. I read every comment.

Venkata Ramana Duddu, Ph.D., Co-Founder, Head of AI and Data Science, Vidaya (https://vidaya.ai)

Congratulations on the launch! How do you protect sensitive health data? Is it sold, ad-targeted, or used to train AI models?

Hey Product Hunt! 👋

I’m Raghavan, CEO of Salesman AI, and I’m incredibly excited to finally share what we’ve been building.

Let me be honest about why we built this.

📅 Getting the meeting is only half the battle.

The AE still has to:

→ Understand the buyer
→ Prepare for objections
→ Run a strong call
→ Follow up quickly
→ Qualify the opportunity
→ Plan the next step
→ Keep the deal moving

And they’re expected to do all of it while jumping from one meeting to the next. 😵‍💫

Most sales tools record what happened or help managers understand what went wrong later.

But the AE needs help before the call, while the conversation is happening and after everyone leaves the meeting.

That’s why we built Salesman AI: one AI sales agent that turns meetings into revenue. 🤖→💰

Salesman AI carries buyer, meeting and deal context across the entire sales cycle. It retrieves what matters from previous calls, deal history, buyer behaviour and seller notes—then uses that context to help the AE prepare, rehearse, respond and act.

🔍 Before the call

Salesman AI builds a meeting-specific buyer brief and creates an AI simulation of the customer—their role, communication style, priorities and likely objections—so the AE can practise the conversation before it becomes real.

Pre-Call Notes use DISC and deal context to show:

✓ Who you’re meeting
✓ What matters to that buyer
✓ Which risks need to be addressed
✓ What questions you should ask

So when the AE walks into the meeting, they’ve already been there. 🎯

🎧 During the call

On-Call Nudges detect objection signals, buying intent and missed questions, then surface a private, context-aware response without pulling the AE’s attention away from the buyer.

Most tools tell you what went wrong after the call.

Salesman AI helps while you can still do something about it. ⚡

✅ After the call

Once the meeting ends, FOCUS turns the transcript into deal intelligence and next actions.

It:

→ Extracts buying signals and blockers
→ Updates qualification context
→ Drafts the follow-up
→ Organizes CRM-ready notes
→ Assigns the next actions
→ Builds the strategy for the next meeting

Every conversation contributes to a clearer view of the deal: who is involved, what is blocking progress, what the buyer has committed to and what needs to happen next. 📈

Before. During. After. One AI sales agent carrying the deal forward. 🔄

We don’t measure success by how many calls were recorded or how many practice sessions were completed.

The number we care about is simple:

🏆 How many rehearsal meetings and real meetings did it take to help the AE win the deal?

Because Salesman AI isn’t here to help you get more meetings.

It’s here to help you win more of the meetings already on your calendar and convert conversation evidence into measurable deal progress.

Getting the meeting was half the battle. Salesman AI wins the other half. 🤝

We’d love your support, feedback and brutally honest opinions. That’s what Product Hunt is for. 💜

Let’s go! 🚀

Raghavan
CEO @ Salesman AI

The during-call nudges are such a good idea, wish I'd had that back when I was doing sales calls. Good luck today!

congrats @anisha_rathinam_ @rs_raghavan ! 🚀 super exciting to see Salesman AI finally out in the world. Wishing you and the team all the best for the launch!

curious to know what’s the one part of the sales workflow you’re most excited to see Salesman AI transform?

Prep is the part that actually moves deals, and it's the first thing that goes when you're back to back all day. You finish one call, next one starts in two minutes, and you walk in cold knowing you should've read up. This may be a really useful tool for fixing that.

Been using Salesman AI in early access and it’s genuinely made my meeting prep and follow-ups much easier, great product for sales folks!

Hi Product Hunt 👋

We built Heym because creating an AI workflow is only the beginning. The hard part is connecting agents, company data, tools, approvals, and production infrastructure without ending up with a stack of disconnected products.

Heym brings these pieces together in one source-available platform that runs on your own infrastructure. Individual users can build and operate complete agentic workflows, while teams can collaborate through shared workflows, human approval steps, and an Agentic Kanban Board that coordinates work between people and agents.

You can design multi-agent workflows visually, connect RAG, MCP tools, APIs, and coding agents, then inspect every execution through traces, token usage, model costs, latency, and evals.

For this relaunch, we added first-class Codex and OpenCode nodes, an Agentic Kanban Board that runs workflows as work moves between stages, bidirectional MCP support, live production run inspection, and improved Docker and Kubernetes deployment.

Heym is free to self-host. You can use your own models, credentials, and infrastructure, and expose the same workflow through portals, APIs, or MCP.

We would love to hear what you are building, which integrations you need next, and what would make Heym more useful for you or your team.

Hi Product Hunt! 👋

Thanks for checking out Heym. We built it for teams that want to turn AI ideas into reliable visual workflows, with agents, RAG, MCP tools, human approvals, traces, and self-hosting in one place.

If you’re comparing Heym with n8n, this overview is a good place to start:
https://heym.run/compare/n8n

Here are our three most-watched tutorials:

• Agentic AI Workflow Automation with Heym Boards
https://www.youtube.com/watch?v=oDoTfbU4O_M

• Heym Dashboards: Build Interactive Dashboards with AI
https://www.youtube.com/watch?v=UAtFnCXmIQw

• Heym Analytics: Monitor and Improve Your AI Workflows
https://www.youtube.com/watch?v=oWAV-_JJSKk

We’d love to hear what you’d automate with Heym, and we’re happy to help you get started.

Congrats on the launch! Lovely to see products like this being open-sourced.

Hi everyone 🤗 One thing worth highlighting about Heym is where we are headed.

More capable models like DeepSeek V4 Flash 0731, combined with two DGX Spark systems, make truly long running local AI inference workflows possible. Coding agents can continuously build, test, improve, and ship software. General purpose agents can handle content, research, and other recurring operational work.

We believe AI automation should remain open, accessible, self hosted, and privacy minded.

Heym does not collect telemetry, and we never will. More importantly, we do not hide essential capabilities behind enterprise gatekeeping or closed sales calls: https://github.com/heymrun/heym#no-enterprise-gatekeeping

Thank you for your support, upvotes, and belief in us. Please keep supporting Heym as we build what comes next.

Congrats on the launch! The combination of self-hosting, human approvals, and built-in traces/evals feels especially strong for teams that want agentic workflows without losing control.

Being able to ship the same workflow as a portal, API, or MCP tool also makes Heym feel much more production-oriented than a typical visual agent builder.

Wishing you good luck with the launch.

Hey Product Hunt! Austin here — excited to share PostSnag with you!

If you’ve ever wished you could instantly see what’s going viral on Facebook, PostSnag was built for exactly that. It captures posts as you scroll, ranks them by likes, comments, shares, and video plays, and highlights viral outliers in a live Discovery feed so you can spot winning content fast.

You can turn any Facebook profile, page, or group into a searchable swipe file, organize everything, and export clean post data straight into Claude, ChatGPT, Perplexity, Gemini, and more. No proxies, no fake accounts, no API — just real posts captured in real time.

We also added MCP support, so you can plug PostSnag directly into your favorite AI platform and automate your research workflow.

📌 Chrome extension + dashboard
📌 Free to install
📌 $19/mo Pro or $139 lifetime

We built PostSnag to make content research effortless, and we’d love your feedback as we continue improving it. Thanks for checking us out — excited to hear what you think!

This is a great product! We pay close attention to everything @austinarmstrong releases and we have our social content team using it.

Are you thinking of extending it to reddit Linkedin and other socials as well?

Congrats on the launch! Super cool product!

Hey everyone, Darius here, one of the makers on dolv.

Nolan covered the what. I'll add the why behind one specific decision.

The approval gate was the most debated thing we built. Early feedback pushed us to make it optional or off by default.

We kept it on because we watched what actually happens when automation touches customer-facing channels without a clear pause point: it gets switched off entirely, and the work comes back to the person anyway.

The gate is not a safety net. It is the reason the rest of the product can run.

If you have questions about how the approval flow works in practice, or how we handle cases where steps partially complete, I'm here for it.

@nolan_vu Love the focus on execution, not just analytics. Keeping approvals with the team while dolv handles the repetitive cross-tool work feels like a really practical use of AI agents.
Good luck guys!

Hello there! Thibaut here, CTO of Oriane.

Following Yuri's comment, the hard part isn't building a Moodboard UI, it's the data underneath it. To generate this for any handle you paste, we're pulling every video from that account, processing each one across three layers (what's on screen, what's said out loud, what's in the caption), and fusing it all into structured records tied back to that single account. On demand, in seconds. That's the power of the Oriane API.

Most tools either rely on platform APIs that throttle you hard, or scrape one video at a time. We're running this across both Instagram and TikTok through the same pipeline, so any account's full footprint comes back instantly, no matter the platform.

Happy to answer technical questions for anyone building on top of it. Fire away 👇

Julien here, cofounder of Oriane.

The problem is simple. Everyone studies accounts by scrolling and guessing. You can see the numbers on a big post, but not why it worked. So the real playbook stays invisible.

So this tool does the teardown for you. Paste any Instagram or TikTok account. It watches the content, surfaces the top hooks and breakout posts, charts views by video length and cadence, and writes short strategy notes on what the account is doing. You can export the whole thing as PDF slides.

Free. No signup. Under a minute.

We ran it on @ hormozi to test it: 4.4M followers, 98 videos, 25.2M views, and it instantly showed the 60 to 90 second sweet spot and the hook style driving the breakouts.

The fun part for this community: the whole thing is a thin app on the Oriane API.
Yuri built it in under a day with vibe coding. The API does the seeing and hearing inside videos, the app is just UI on top.

If you want to build your own tools on it, happy to get you access.

Would love your feedback on this tool, and tell us what other free products we should build in the next 6 weeks?