PH热榜 | 2026-07-27

Hey Product Hunt 👋

Simon here, co-founder and CEO of Adomate.

Before Adomate was a self-serve platform, Lucas and I ran it as a service, building and delivering Meta ad creatives by hand for DTC brands across Belgium and the Netherlands.

We wanted to feel the problem first:

• Strategists losing 60-70% of their week to research instead of making

• Rebuilding context from zero every Monday

• No system connecting last week's winner to what gets briefed next

Creative is the last lever you actually control on Meta now that targeting and bidding are automated, and most teams still run it like it's 2019.

The solution

We built Adomate. It puts your performance data, competitor intel, and consumer signals into one workflow you build and own. A few things it's already done for early teams:

→ Increased number of winners and ROAS

→ Volume: helped Nomige to meet creative demand and deliver the number or creatives required for testing

→ Golden nuggets: automatically found usable insights in thousands of customer reviews

How it works

1. Feed in your performance data, brand brain, competitor ads, reviews, and consumer signals. One place, not ten tabs, so research time drops to minutes.

2. Build the workflow once. Every data point after that becomes an ad concept, so you stop starting from scratch every session.

3. Like the concept, skip it, or refine it in plain language. Point out a reference if you've got one. Every choice trains Adomate and saves a version, so getting from 80% to 100% takes minutes, not a new brief.

Why it's different

No black box. Every concept you see traces back to the exact data point that triggered it, the competitor ad, the review line, the performance metric. You can always see why.

Adomate doesn't replace your judgment, it amplifies it.

Who it's for

Creative strategists and performance marketers at DTC and e-commerce brands running Meta ads, plus agencies managing creative across multiple accounts.

If briefing your designer feels slower than it should, this is for you.

Special for the PH community

→ 30% off your first 3 months: PHMONTHLY30. 300 codes.

→ 50% off your first year: PHANNUAL50. 100 codes.

Apply the code at checkout. Valid until Friday 11:59pm PT, or until the codes run out.

Our ask

If you're running paid social, tell us what's broken in your workflow right now. That's the input that shapes what we build next.

👉 Get started for free to see what you can do with Adomate.

@lucas_desard and I will be here all day. 👇 Big thanks to @rohanrecommends for hunting us 🙏

Congrats on launching! We spend way too long digging through customer reviews to find angles for ads, so pulling Trustpilot and Amazon feedback into the same workspace as ad performance sounds really handy. How does it work for smaller brands that only have a handful of reviews and a modest ad account to learn from?

The traceability is the strongest part here. AI can generate endless ad ideas, but without knowing whether a concept came from a winning campaign, a competitor pattern, or an actual customer review, it is hard to trust or learn from the output. As someone doing more launch and marketing work, I can relate to rebuilding context across too many tabs and starting the creative process from zero every time. Curious how quickly Adomate learns a team's taste from likes, skips, and refinements before the concepts start feeling genuinely tailored :)

Congrats on the launch!

Interesting Product, congratulations on the launch.

Love the focus on solving the biggest bottleneck in Meta ads—creative research. The traceable workflow and data-first approach really stand out. Congrats on the launch, Simon! Wishing the team a strong Product Hunt day 🚀

Congrats on the #1 launch 🎉 The part that stands out to me isn't "AI generates ads" — it's the "fully traceable to the original data trigger, no blackbox" framing. Most gen-creative tools hand you an output with zero lineage, so you can't tell why a concept was made or how to steer the next one. Curious how granular that trace actually is: when a concept links back to a data trigger, does it point to the exact source item — this Trustpilot review, this ad-library creative — and can I act on it, e.g. "this hook came from these 3 reviews, regenerate weighting them differently"? Basically: is the traceability a provenance view, or an actual control surface you steer generation with? Really nice work either way 🙏

Congrats on the launch. The service-first route makes sense here. You usually learn much more by doing the work manually before trying to automate it. I'd like to understand is how Adomate decides what should influence the next brief. Performance data can tell you what won, but not always why it won. Can the team challenge the system’s reasoning or add context when a creative worked because of timing, an offer, or something happening in the market? That seems important if the goal is to avoid turning last week’s winner into next week’s template.

Seen a lot of ad generation tools, but nothing works on data. Adomate seems to be the first. Congrats!!

Love the product and the team. From my previous launch, I met with Ozan (he is in Adomate team) and he showed me the demo of the Adomate. The most interesting part they are generate creatives by learning from your competitors. I'm planning to use it for my e-commerce business. Hopefully, in near future I can use for my apps too :)

The traceability is amazing , knowing which data point triggered which concept makes it easier to kill bad directions without guessing

Will it work with other platforms in the future? :)

Pulling from Meta ad data, the ad library, and Trustpilot/Amazon reviews as raw material is a sharp approach, most ad generators start from a blank prompt, but the best-performing creative usually starts from what's already resonating. The full traceability back to the original data trigger is the part that earns trust; "no blackbox" matters a lot when you're trying to explain to a client why a concept exists. Curious how much control you get over brand guardrails so the output stays on-brand at scale. Congrats on the launch!

the traceability and spend-data answers above cover picking winners, but what about the fatigue side - if the engine keeps spinning variants off the same proven winner, how do you stop Meta from showing near-identical creative to the same audience over and over. is pushing for stylistic variance built in, or is that still on the strategist to catch before it burns out an audience

the traceability pitch is what sold me too, but I'm curious about the agency case specifically. if I'm running Adomate across multiple client accounts, is the review/competitor data and brand brain fully siloed per workspace, or is there any shared learning happening under the hood that could let one client's insights bleed into another client's concepts

This looks useful. How do you check the ads before they go live and make sure they do not include anything misleading?

The traceability angle is what caught my eye here. Most ad tools spit out concepts with zero explanation, so seeing exactly which review line or competitor ad triggered a suggestion is the part creative teams will actually care about. Does the attribution work in reverse too, meaning can I start from one specific Trustpilot review and ask it to generate concepts from just that?

Not gonna lie, it looks really good.

Maybe the only thing I'd like to see is to get an analysis and then prepare an ad myself.

Is it possible?

@s_logghe You mention agencies managing creative across multiple accounts as one of your target users. For those teams, is there hard isolation between client brand brains, so one client's reviews, competitor set, and performance data can never surface in another client's generated concepts?

@s_logghe Congratulations. And happy product launch.

I run paid social on the buying side and support on the receiving side, so the thing that is broken for me sits downstream of yours.

Golden Nuggets is the feature I would watch. The review lines that make the best ads are almost always about an experience rather than a product. Arrived in two days. They replaced it, no questions asked. That is exactly why they sound authentic, and it is also the moment the ad stops describing the product and starts making an operational promise.

The trace tells you where the claim came from. It does not tell you whether you can still keep it. That review was true for one customer, in one country, under last year's courier contract. Run it as an ad and you have made that promise to everyone who sees it.

And it does not fail where you are looking. ROAS goes up, because it is a good promise. It shows up two weeks later in the inbox as "your ad said two days", and nobody connects that ticket back to the creative that caused it.

What I would want, and you are closer to it than anyone because you already hold the trigger: flag which concepts make a claim about time, price, returns or availability, as against a claim about how the thing looks or feels. The first kind needs one person to confirm it is still true before it runs. The second does not. Everything you need to tell them apart is already sitting in the data point you traced back to.

Ads game is changing and Adomate came to evolve it even more. Creatives are easier and more effective than ever!

Most AI ad tools generate ads. This feels more like a creative strategy engine. Good stuff.

How much historical Meta data does Adomate need before the recommendations become genuinely useful?

Simon, you asked what is broken, so here is mine. I run paid acquisition in healthcare, and the thing that breaks is upstream of creative.

The loop learns from whatever the ad account calls a winner, and in my market that label is wrong. The conversion I can actually fire is a trial signup, and trials and paying customers are not the same population, so my real cost per paying customer came out roughly double what the platform reported. A creative engine trained on that will confidently scale the concept that produces the cheapest signups, which is not the one that produces revenue.

Can Adomate learn from an outcome uploaded back later, a paid conversion at day 30 for instance, or does winner mean whatever Meta says it means?

I am thinkin this could help brands scale creative production without losing consistency. What reporting features help users understand which research sources influence successful ads the most?

Hi Product Hunt.

I’m Pete from Databox. Today we’re shipping Artifacts inside Genie, our AI Analyst. An artifact is a report Genie generates from your connected data. It combines the charts you’d normally build by hand with the analysis of what actually drove the numbers, in one shareable document.

Reporting used to work like this... You’d build a bunch of dashboards. You’d compare performance to the previous period and to goal, try to understand how the work you did turned into the result you got. Then, you'd write out what happened and what you planned to do next.

Whether you hit the goal, missed it, or crushed it, the manual process was the same.

With Artifacts, none of that is necessary. It’s a one-time setup. You connect your data and define your metrics once. After that, you tell Genie what you did, and it compares performance to the previous period, checks whether the work correlated with what you achieved, digs into why through statistical analysis, and recommends what to do next. Then it builds the report.

A “what happened last month” question becomes a finished report in minutes.

A few very specific things it can do:

• Turn the same report into a slide deck. One follow-up request, nothing rebuilt.

• Restyle a report to your own brand or a client’s, mid-conversation.

• If you're a professional services firm, build a cross-client report spanning every account, not one client at a time.

• Ship as a public link, with no login for whoever you send it to, or a PDF. Turn the link off whenever you want.

You might be asking, why not just use Claude or ChatGPT? Because Databox already has your data connected, understands how to calculate your metrics, and knows how to run statistical analysis on them.

Three things the general chat tools don’t have or don't do consistently. Ask one of them for a report and you feed it the data yourself, every time, and hope it calculates and analyzes correctly. And when it gets something wrong, you’re stuck.

Databox artifacts are editable, so you add your own context and Genie refines it. Fine tuning, not redoing.

Less time spent reporting means more time doing the work and more time thinking about how to do it better.
If you’re not on Databox yet, I challenge you to automate your next report in three steps.

1. Start a trial,
2. connect your data,
3. and ask Genie to “create a report analyzing what happened last month and why.”

No coding, no spreadsheet-wrangling required.

This is the part of the job I've spent years fixing manually.

Reporting systems that actually get used are rare, most of them just create more work downstream. Being able to ask for a report and get back something client-ready, built on the live data, is the direction this space needed to go. Excited to see how agencies and marketing teams use this to stop rebuilding the same deck every month.

Marketing analytics person here — the biggest benefit of this is not being the human BI layer anymore. You can ask for a report and get something narrative-ready back. And it's based on verified data (something that really matters but many marketers forget about). Super useful!

One thing I'd add from the marketing perspective: the most useful thing about Artifacts is that it closes the "I have an answer" to "I have something shareable" gap.

For example, ask the AI Analyst to build you a report or deck, and it does so using your real data. Then you can hand it off via public link or PDF and there's no login needed on the other end.

I know this is HUGE for anyone who's ever had to rebuild the same client report by hand every month, and it's a great way to add a supporting doc to your quick answer. Happy to answer questions if useful!

The problem with most AI generated reports is you still have to check them before sending, since the AI didn't actually see your numbers. Artifacts skips that step because Genie pulls straight from your connected data. I asked for a quick performance overview for a stakeholder update and the numbers matched the dashboard exactly, because it's the same data.

Hey Product Hunt!

We kept seeing the same problem: AI analysts got good at answering questions, but the answer dies in the chat. You still have to copy numbers into a deck, rebuild charts, format a report, and by the time you share it, the data is stale.

Artifacts closes that gap. You ask Genie a question, and instead of just an answer, you get back a finished, ready-to-share output. A report, a slide deck, an interactive document. Built from your live data, generated from a single prompt, shareable via public link or PDF.

The chat was never the deliverable. The report is. Now you get both in one step.

This is our fourth launch this year (Genie in March, MCP in June, Skills Marketplace in July), and they all stack: connect your data, chat with it anywhere, apply expert skills, and now walk away with something you can actually put in front of a client or your exec team.

Artifacts is also another important building block in our agentic analytics direction. We're building toward a world where the analyst work happens end to end: the agent pulls the data, runs the analysis, applies the right expertise, and delivers the finished output, with you directing the work instead of doing it. Every launch this year has been a piece of that. Artifacts is the delivery layer.

Try it and tell us where it breaks. That feedback is worth even more to us than the upvote.

Every team I talk to has some version of the same Friday afternoon task: turn the week's numbers into something presentable and pass it around. Artifacts replaces that with a prompt. You ask Genie for the report, get back a finished document, and either share the link or download the PDF. You skip the whole assemble-and-format slog in the middle.

The hardest question we had to answer while building Artifacts wasn't technical, it was about control: when AI writes your report, what still belongs to you? Our answer: Genie builds the document from your live data (with a real design system behind it and a deterministic engine doing the math, so numbers are computed, never guessed), and you keep the last word, literally. Edit any text directly, and Genie knows about your edits, so its next update builds on your version instead of overwriting it. Would love your take on whether we drew that line in the right place. I'll be around all day for questions.

I tested this by asking Genie for a report, then asking it to turn that same report into a slide deck. Same data, same analysis, different format, one follow-up prompt. That's the kind of thing that used to mean rebuilding the whole thing in a different tool. Here it just carried over.

Turning live data into a polished, shareable document in seconds is pretty incredible.

I don’t touch reporting for clients, so I wasn’t expecting to use this much. I asked Genie to put together a slide deck for our all-hands update and it just built it, no design help needed. Took me a few minutes instead of the usual back and forth with someone on the team.

Congrats on shipping this.

The gap between 'I have a question about my data' and 'I have something polished to hand my boss' has always been the part no AI tool solved well.

This closes it. Looking forward to seeing where the format list goes next.

Congratulations on the launch. The connection to live data is what separates this from a generic AI document writer. What you share is grounded in the same numbers your dashboards already show. Well done to the team.

This would save client-facing teams hours every month rebuilding the same presentation for recurring client meetings. Solid update.

I really like that the reports stay connected to live data instead of becoming outdated static documents or PDFs.

How customizable are the generated reports for agencies with strict client branding requirements?

The part worth explaining is how the artifact actually gets built. It's not assembled from a template or a databoard, Genie authors the layout, structure, and visuals fresh for each request, following a shared design spec so it comes out on brand by default. The data inside it is pulled live from your connected metrics at generation time, so what you get is a real document backed by your real sources, not a paraphrase of a chart you pasted in.

I saw the artifacts preview when you had created the PH forum post last week. Good update!

One implementation detail I'm particularly happy with is how Artifacts evolve through a conversation.

When Genie generates a report, you don't have to start over if you want something different.

Need another section? Just ask.

Want a different chart? Ask.

Need to adjust the branding for a client? Ask.

Those kinds of changes happen directly on the existing artifact, so you're refining what you've already built instead of starting from scratch.

For bigger transformations, like turning a report into a slide deck, Genie creates a new artifact while keeping the original report intact. Both versions are available to share, and because they're built from the same connected data and conversation, you don't lose the context that got you there.

We didn't want AI to feel like a series of disconnected prompts. We wanted it to feel like working on a living document that evolves with you.

Congrats on the launch!

For teams migrating from existing BI tools, how seamless is the process of importing historical datasets and custom metrics, and does Databox offer any AI assisted mapping to speed up that transition?

Congrats on the launch!

Worth flagging for anyone who tried an earlier version of this: you can now edit text and titles directly inside the artifact panel, no prompt needed for small fixes. And every artifact you make is saved to its own gallery with search and favorites, so finding last month's report doesn't mean digging through chat history anymore.

Looks cool!!

The question I get most on calls is some version of 'so once Genie answers, then what.' Before, the answer was a chat message. Now it's a finished report or slide deck with your actual data in it, shareable with a link. For prospects who need to show something to their team after a demo, that's a much easier story to tell.

This is interesting. What's the most surprising query users didn't think to ask?

I find this release shows how adaptable and mature Databox is becoming as a company. It already has a pretty sophisticated Databoards feature that has been polished to be pixel-perfect over the years. When users shared data outside of Databox, they shared those Databoards. And we made sure they always made an impression. That they always looked Databox-ish. Years were spent improving them. Loud conference room meetings were spent over positioning pixels on them. It was something we were proud of, and, of course, still are.

Today, we handed some if that control over to our customers through AI (Genie). Building what they share with the outside world. The Artifacts still look Databox-ish, colleagues made sure of that. And it does come with lots of upsides - building Artifact with a prompt is way faster than placing and configuring blocks on a Databoard, plus AI has a lot of knowledge built-in. But still, we had to let go of some of the control we had over what can be shared. What carries our logo.

At the end of the day, that's what customers wanted. And that's what "bites" in today's fast-paced world. Adapting to that need, while knowing it might cannibalise our own prized feature.
That has to mean maturity.

GG Team!

What makes Genie useful is that it's working from the source data, not trying to interpret a screenshot. It connects to 130+ tools and pulls the numbers directly. So when the report is going to a client or the board, you don't have to manually check every figure before sending it.

Big congrats to the team. What stands out is that the document is built fresh each time, from your real connected data, instead of assembled from a template. That's the part that makes it trustworthy enough to send without double checking.

To get the most out of this model, here's the new prompting guide:

• Agentic coding: Claude Opus 5 is strongest on difficult coding tasks: multi-file features, larger refactors, and end-to-end feature work. It completes full tasks rather than leaving stubs or placeholders, and it performs best when given the complete task specification up front and left to run. It also performs well on easier tasks like single-turn edits, where the difference from prior models is smaller.

• Code review and bug-finding: Claude Opus 5 reviews code with high precision and recall: it finds real bugs at a high rate per pass, and its additional findings are mostly real issues rather than false positives. Accuracy holds at lower effort settings, which supports a fast pass at review time and a more thorough pass later. If your review prompt says "only report high-severity issues" or "be conservative," the model may follow that instruction literally and report less; ask it to report everything and filter in a separate pass instead.

• Efficiency at lower effort: low and medium effort produce strong quality at a fraction of the tokens and latency of higher settings, and they perform above the same settings on prior Opus models. Use them liberally as your primary control for token cost and response time wherever your evals show quality holds; for coding and agentic work, xhigh remains the recommended starting point. If you carried effort defaults over from a prior model, re-run an effort sweep on your own evals. See Effort for the full recommendations.

• Vision: Claude Opus 5 is strong on chart, document, and diagram understanding, and on UI and frontend visual replication. Re-validate any prompt-side vision workarounds you tuned for prior models; they may no longer be needed. Vision performance is strongest when the model has tools to iteratively analyze, crop, and visually verify its work, and tool use is a more cost-effective lever than thinking alone.

• Long-context work: Claude Opus 5 has a 1M token context window as both the default and the maximum, and its instruction following, tool calling, and reasoning stay consistent throughout the window.

• Office and document tasks: Claude Opus 5 generates and works with complex, multi-sheet spreadsheets with non-trivial formulas, and it produces well-structured slide decks. Prompt it with any specific styles or templates it needs to follow.

• Multi-agent coordination: Claude Opus 5 coordinates teams of subagents well, with effective writer-verifier patterns and few cases of agents overwriting each other's work. For cost-sensitive workloads, cap delegation; see Controlling subagent spawning.

Been running it in Claude Code. Not sure yet if it's better or worse than Fable 5 for what I do, but the token efficiency alone is a big win.

Very nice. Here's a comment generated by Opus 5 posted on a launch page for Opus 5.

Claude has been huge for me. I understand tech well but I'm not a developer, and it's helped me take ideas that used to just live in my head and turn them into real working software I can put in front of test users. It walked me through publishing a live website and even setting up a proper dev environment with GitHub, stuff I never would have figured out on my own. Excited to use opus 5 and see how it compares to fable 5.

alright - Opus 5 is a banger. Fable 5-level reasoning at half the price, @Claude by Anthropic is killing it. it's already running on @Kilo Code, and early traction shows similar performance to GPT-5.6.

the real question now isn't whether AI can write good code, it's which model fits the task and at what cost. LFG

Congrats on the launch. The price move matters more than the benchmark bump - halving the cost of the top tier is what decides whether long-running agents stay a demo or actually go to production. Nice work to the team.

The claim I want to poke at is the 1M context holding up consistently, since that is usually where things quietly degrade. Multi-file refactors landing end to end instead of stopping halfway matters more to me day to day than benchmark numbers. How does the subagent management differ from before, is the parent model deciding when to spawn them now or is that still on the developer?

Half the price for close to the same quality is the kind of update that actually shows up on the bill. Most of what I throw at it is long messy work that runs for an hour, so the end to end feature work angle is what I care about. Does the lower pricing hold across the whole 1M context window, or does the cost step up once you go past a certain point?

That's nice. How does it handle long-context reasoning compared to Fable 5?

Hi Product Hunt, I’m Moe, the founder of Webhound. I’ve worked on AI research since 2023.

I started Webhound because research agents have a stopping problem.

A coding agent can stop when the tests pass. Research has no equivalent finish line. An agent can spend ten minutes, two hours, or twenty hours on the same question, and each answer can look complete. Research agents tend to stop once they have enough evidence to sound confident.

You still do not know which leads they skipped, where sources disagreed, or whether another hour would uncover the fact that changes your decision. The agent makes that stopping decision for you.

Our thesis is that budget should be a research primitive. In plain English, your prompt tells Webhound what to investigate. Your dollar budget tells it how much work to put in and caps what you can spend. At our current rate, $5 funds about 75 minutes of research.

Search finds sources for the query in front of it. Research reads those sources and follows the leads they reveal. One source can change what Webhound needs to search for next.

You get a cited report or sourced dataset, along with the sources and research notes behind it. You can inspect how Webhound reached its conclusions and where the investigation still has gaps.

Extra budget must earn its cost. We judge a larger run by the useful evidence it adds and whether that evidence improves your decision. Extra length does not count.

You can run Webhound in the app or use it behind Codex, Claude Code, Cursor, Manus, or your own software. Your agent can hand off a question, continue working, and retrieve the finished research later.

New accounts include one $5 Report or Dataset. There is no subscription.

I want blunt feedback on the core idea: does a dollar budget feel like a useful way to control how much research gets done? Do the sources and research notes help you decide what to trust?

If you have a question where missing information could cost more than the research, leave it below. We’ll run a few in public today.

dollar budget as the stopping primitive is a genuinely clean idea, most research tools just stop when the output sounds confident rather than when it's actually verified. one thing I didn't see covered - when two sources genuinely disagree on a claim and the budget isn't there to fully resolve it, does the final report surface that disagreement explicitly, or does it pick a side (majority source, more authoritative domain, etc) and present one answer? for a cited research tool I'd actually want to see the unresolved conflict called out rather than smoothed over

Does it search only public websites, or can it also work with my own documents and internal knowledge base?

Hi Product Hunt, I'm Theo, the other founder of Webhound.

Since Moe has already covered a bunch about our product I wanted to give some of my (admittedly messy) thoughts on what I've learned from designing this more agent-facing version.

Recently my cofounder and I have been noticing that a lot of our most consistent users are actually agents, and that share is only growing. As our discussions started to trend towards designing for agents, my question started to become: should we be?

I don’t think we should be designing for agents just yet. And this is not (entirely) because of my tendency to favor human centered design, but rather an attribution problem / thesis.

When we talked to the users behind the agents with carte blanche to our product, we kept hearing “I don’t want to have to leave my second brain.” Memory, context, and integrations are creating the same kind of lock-in as social graphs did with social media.

Many founders I talk to are of the belief that this will change, and that we will all be using a bunch of different AI interfaces in the future. My counterpoint to that is to harken back to the myriad of apps I used to have on my iPod touch (in comparison to the 3-4 apps I use today, not even mentioning all of the hardware products that got condensed into the iPhone).

This process of divergence in tools seems to happen every time a new technology explodes, and then we converge back to the single interface model.

If we assume that people are going to be working from a single centralized interface, the model I start to think about is heavily reliant on the control surface, how human intent is injected into it, and what steps exist within that platform / tool that are beyond my view.

When I talk about designing for agents, I am not talking about agents as the end users. As an HCI major I am still thinking about the human as the end user. There are just more layers of abstraction between human will and silicon execution. This is why I think it’s important to factor the control surface into the mental model / think about where human goals / input enter the system.

With that as the precedent, here are some of the more practical things I’ve learned so far.

For open ended tasks, budget can be a better stopping primitive than subjective completion criteria + an LLM as a judge (contractor).

Memory and workspaces / file structure + integrations create lock in. Try to build something that requires users to leave their second brain as infrequently as possible.

Error messages should be verbose and imply fixes. Since a lot of failures and behaviors sit on the user agent side, it can be hard to view these failure modes. Try to capture them with error logging.

Define your rules of engagement / best practices for agents to interact with your tool / when to call.

Don’t block when you don’t need to. Make state and status visible to avoid hangs / blocking the main agent.

Onboarding should be an MCP endpoint and should be done through the user’s agent. When they create a key, give the user a text prompt to copy to set up the MCP with their agent.

In this prompt, ask the agent to state what environment it is in (Claude Code, Codex, etc), use that client’s supported global MCP configuration unless the user wants it to be project level, save the exact server configuration you give it, call the health endpoint so that you don’t have false positives for successful installs, and then call your onboarding endpoint.

We’ve found this works best when the onboarding endpoint returns steps serially and you ask the user’s agent to keep calling “next step” until it returns that there aren’t any remaining.

There is also intent monitoring, where you have tool search and can see what people are asking their agent for your tool to do. We decided against doing this because it felt too sketchy, but it is something we thought about.

These are the learnings from a couple week deep dive and I hope they are useful to some of you, though I expect them to be relatively fleeting as the interface evolves.

How does Webhound decide when to stop digging deeper versus continue researching, especially for open ended or ambiguous topics?

probe real submit pathBudget as the stopping rule makes sense, especially when another agent is waiting on the answer. The trust layer I’d want is a short handoff note: what was checked, what was intentionally skipped, and which unresolved claims could change the decision if someone spends another hour.

Interesting! That framing of budget-as-stopping-primitive is definitely worth attention. Most research tools stop when the prose sounds finished rather than when the work is actually done. Curious whether you guys plan the search tree up front and rank leads by expected payoff? Is it greedy step by step and just stops when the meter hits zero?

How well does it handle complex prompts with multiple research objectives in a single request?

Really like the "onboarding should be an MCP endpoint" framing — that serial next-step pattern is clean. One edge case I'd be curious about: if the health check passes but step 3 of 5 fails on a config mismatch, does the endpoint expect the agent to retry from that step, or restart the whole sequence? Partial-failure handling in a multi-step handshake like this seems like it'd need its own design pass separate from the happy path.

How does Webhound handle conflicting sources? If two credible references disagree on a fact does the report flag that discrepancy or try to weight one more heavily based on the research trail?

If a $5 run hits a dead end early, does the user get to adjust the budget mid-stream or is it strictly set at the start?

A nice direction, especially the cited reports angle. One thing that would make this more useful for me as a user: a way to pause a long-running agent and resume later without losing the work it has already done, plus an export to plain markdown or a Notion page so I can drop the output straight into my team's knowledge base. Right now I worry about kicking off a multi-hour run and having it vanish if I close my laptop.

It would be really helpful to have a way to export the agent's reasoning trace or intermediate findings, not just the final report. Sometimes I want to dig into how the agent arrived at a particular conclusion, especially when I'm using it for due diligence work where I need to show my work.

The budget dial is the right primitive. We maintain a pricing dataset where the answers change monthly and every number needs a source you can defend - and the failure mode of most research tools is a confident summary over dead or circular citations. If the reports really keep the working documents attached, that's the part worth paying for: auditing the claim afterwards, not just reading it.

The budget dial is a smart framing. Research has no natural finish line, and looking finished is exactly how a shallow answer slips through.

The thing I would want in the output is a fetch date on every source, so a report from last month can be re-checked instead of trusted forever. Does the dataset mode keep a timestamp per row?

makes sense, flagging the lack of evidence as a limitation instead of just going quiet is the right default. thanks for the answer

The budget-based depth control is an interesting way to make research effort explicit. How does Webhound decide when a claim needs further verification, and can users inspect why the agent stopped researching a particular lead?

Moe — the disagreement-surfacing answer to Gal is solid. My research problem's different though: federal contract award data and past-performance records aren't scattered across the open web, they're stuck behind a clunky government portal with no public API or real search. Does Webhound handle sources like that, or is it built for open-web research?

The dollar budget is a clever constraint, but the stronger idea might be making research depth explicit. Most tools hide the stopping decision behind a confident-looking paragraph. Here, at least, I can decide whether a question deserves five dollars or five minutes. I like that “not enough evidence” can be a valid output.

Budget as the stopping primitive answers how much, and there's a second question sitting under it: how much does the same $5 vary? Two runs on one question at one budget follow different leads, and an agent that follows leads is path-dependent by construction — whichever source it happens to open early reshapes everything after it. The report is a function of the budget and of which door it went through first.

That lands hardest exactly where Clemente was pointing, on the MCP path. A human feels a thin answer and re-runs it. An agent takes the first report as ground truth. I do eval work on my own app's generated output, and the number that changed how I ship wasn't the average score — it was the spread across identical inputs. The mean looked healthy for weeks while the bottom of the distribution was quietly unusable.

Have you measured that spread on a fixed question and budget? And does the per-claim confidence score reflect run-to-run stability, or only the evidence inside the single run that produced it?

Depth over speed is a refreshing pitch when everything else is racing to answer in two seconds. Exposing budget as the control on research quality is smarter than hiding it behind a vague quality slider. When it builds a dataset rather than a report, how does it handle two sources that contradict each other, does the row keep both values or does the agent pick one?

This looks useful, I lose hours copying company details off websites into a spreadsheet by hand. Having the choice between a clean dataset and a fully cited report covers pretty much every research job that lands on my desk. If I ran the same query again next month, would it give me a fresh dataset I could diff against the old one to see what changed?

Congratulations

I'd love to see how Webhound handles edge cases, like sources with paywalls or outdated information. How do the research agents adapt to these challenges?

I see that this is positioned as the "research engine" behind your AI agent, but it wasn't clear to me how this is different than getting the frontier model to do deep research and continually prompting it to continue researching after the default stopping point.

Budget as the stopping rule makes sense, especially when another agent is waiting on the answer. The trust layer I’d want is a short handoff note: what was checked, what was intentionally skipped, and which unresolved claims could change the decision if someone spends another hour.

the budget-as-stopping-primitive idea makes sense to me. curious about the opposite failure mode though - what happens when a topic just doesn't have much written about it, like an internal process at a small company or something too new to have coverage. does it recognize early that more budget won't surface anything and stop, or does it keep spending trying to find sources that don't exist? that seems like the case where the dial wouldn't actually save you money.

Congrats on the launch, Moe.
Webhound is tackling a really interesting problem with AI research the stopping problem is definitely something that gets overlooked. I especially like the idea of using a dollar budget as a research primitive instead of letting agents run indefinitely. I will defiently try this out

Point at an element, describe the fix in plain English, that workflow alone would save me hours every week.

The "invisible to ChatGPT" concern is real for me right now. I run a small content site, and I've noticed referral traffic from AI search growing while my usual SEO metrics stay flat. Curious how deep Robynn's audit goes on that front.

I'm curious how it handles brand voice specifically. My site has a pretty distinct tone, and I don't want fixes that flatten it.

I like that this isn't positioned as a redesign tool. Most of my site doesn't need a redesign, it needs someone paying attention to it consistently, which I haven't had time to do myself. If the audit genuinely benchmarks against my actual competitors rather than generic best practices, that's the difference between useful and just noise.

This is the first pitch I've read that treats website upkeep as ongoing care instead of a one-time project.

he part that caught me is the rollback on regressions. I've used tools that let you publish changes but never actually track if they helped or hurt. Measuring against GA and reversing bad calls is the real differentiator here.

I run a small agency and stale pages are my biggest headache, clients ship a site, then forget it exists. What I like about it is the "no rebuild, no migration" promise. Most audit tools want you to move platforms first. Does it work across custom-built sites or mainly on standard CMS platforms like WordPress?

I've tested a handful of "AI website" tools over the past year, and most either overpromise or require a full migration before they're useful. It connecting directly to a live site is a smart call. My question is around approval, when you say "agents stage it, you approve," how granular is that review? Can I approve individual copy changes separately from structural ones, or is it all-or-nothing per page?

Broken links and slipping rankings are exactly what quietly kill a website. Glad someone's building for that specific problem.

Hey Product Hunt! 👋 Madhukar here, co-founder of Robynn.

My co-founder, Rafiq and I both worked at Interwoven, the company behind the first generation of web content management (TeamSite/LiveSite — still running inside OpenText decades later). That generation made websites possible to publish. But it froze an assumption in place: a website is a project with an end date. Launch day is the finish line, and then the site quietly decays — broken links, stale pages, rankings that slip, and now invisibility in ChatGPT answers nobody is watching for.

Big companies solve this with people: web engineers, agencies, SEO retainers. Small teams get a to-do list that never shrinks. Robynn is that engineering muscle for every marketing team. Connect your existing site (no rebuild, no migration), get a continuous diagnosis against your brand and ICP, point at anything and describe the change in plain English. Agents rewrite and stage it; nothing goes live until you approve. Then we measure what moved in GA — wins get reinforced, regressions roll back.

The design principle we won't budge on: the machine does the labor, you keep the judgment. The first diagnosis is free and takes minutes. Would love your feedback — I'll be here all day answering questions!

PS: A quick note about pricing. When you sign up you get 300 free credits. Reach out to me over DM on LinkedIn or X if you are interested in more and we are happy to add more if you need them.

Staging changes for approval before anything publishes is the detail that makes this usable on a real site, auto publishing would have been a hard no for me. Closing the loop through Google Analytics and rolling back regressions on its own is a nice touch. What actually triggers a rollback, is there a minimum traffic threshold before it decides a change hurt performance?

The "website starts decaying the day it ships" framing is painfully accurate. as someone preparing a product launch, I know how quickly pages, copy, links, SEO, and docs can fall behind once the team moves on to the next thing. I like that Robynn works on the existing site instead of forcing a rebuild, and the stage -> approve -> measure -> rollback loop feels like the right balance between automation and control. Curious how it decides which changes are actually worth proposing instead of turning every small issue into more work for the team :)

the wins-get-reinforced, regressions-roll-back loop is the interesting part to me. GA signal is noisy and slow, especially for a lower-traffic site where a real ranking or conversion shift can take weeks to separate from normal variance. how long does Robynn wait before it calls a change a win or a regression, and does that window scale with the site's traffic volume so a small site doesn't get a false rollback from noise alone

Congrats on the launch! I ran my site through the free audit, and the findings were clear and actionable.

SEO audits themselves are already a crowded space, so for me the most interesting part is the closed loop: diagnosing an issue, staging the actual change, publishing it after approval, measuring the result, and rolling it back if performance drops.

That could make Robynn much more useful than another checklist or SEO score. I’d be curious to learn how you isolate the impact of a specific change from seasonality, campaigns, or other updates happening on the website at the same time.

Great execution and a very polished experience!

The rollback is the part that got me. Most of these stop at telling you what's wrong, which just hands you another backlog. Closing the loop is the hard half - and once reverting is automatic you don't have to be right about any single change, just reversible. That's the whole trick.

One thing I'd want: a holdout. I run a metrics dashboard for my own stuff and got burned by this recently - download numbers are weekday-driven, weekends run 40-60% below midweek, so every Monday looks like a win and every Saturday looks like a regression. Nothing to do with anything I shipped. Without a control slice you end up reinforcing the calendar.

Congrats on the launch.

Nice launch. We have no web engineer and our site slowly rots, broken links and pages nobody has touched in a year, so a free audit that runs in minutes is tempting. Pointing at an element and describing the change in plain English sounds much better than filing a ticket and waiting. Does it work on a site built in a visual builder like Webflow, or does it need direct access to the code?

I believe the first free audit is a smart way to build trust. Real findings often convience people fatser than promises. What are the most common issues discovered during those first audits?

Continuous improvement feels much more realistic than treating websites as one-time projects. 👏 Do you have plans for multiple websites?

The rollback on regressions is underrated. Plenty of tools help you publish changes but almost none actually admit when a change didn't work and undo it.

the rollback part is what caught my eye. it'd be really cool if u could also explain why a change got rolled back so teams can actually learn from it instead of just seeing "it didnt work."

the email digest with a default-publish timeout is a fair middle ground honestly, at least it forces someone to actively ignore it rather than never seeing it at all. curious if the digest highlights which changes are lower-risk copy tweaks vs actual layout/pricing changes, since I'd guess most of the rubber-stamping risk comes from lumping trivial and consequential changes into one daily email

@madhukar_kumar1 How quickly do customers usually notice measureable improvements after deploment?

Congrats on the launch. Sounds like a great product. A couple of quick questions though. Does it support split testing and CRO? I know you can't "split test" for SEO but other on-page elements and UX are also very important. Also how do you connect your website, do you have a WP plugin that makes it easy?

the approve-before-publish step is the part I'm skeptical of, not the audits themselves. for a small team that's already stretched thin, isn't that approval queue going to turn into a rubber stamp after the first few weeks, the same way cookie banners and permission prompts do once people get used to clicking through them. if that happens the whole you keep the judgment pitch quietly stops being true even though the UI still shows an approve button.

The rollback feature immediately builds confidence. Nobody wants AI publishing changes without a safety net. :D

It's refreshing to see AI focused on maintaining websites instead of just helping people build them. Too many websites become outdated as soon as they go live.

How does Robynn distinguish between temporary ranking fluctuations and changes that actually need intervention?

I often see sites with outdated privacy policy and copyright pages. Robynn could now take care of these :D

As someone who is not a marketer but loves to blog and has a personal website, I would like to try Robynn to increase traffic to my blog, I am tired of Linkedin as a way to share my writings, I would like something that persists. I would like people to read my blog six months after I have posted it, and hopefully Robynn can help me with better SEO and GEO rankings, does it do that?

Hi everyone!

Superfile brings a proper file-management workspace to the terminal.

Run spf and you can keep several directories open at once, move files between panels, search quickly, handle bulk operations, and preview code or images without leaving the terminal.

The interface is what makes it work so well. The layout is clear, the keyboard flow is easy to follow, and it already feels polished before you touch the config!

looks nice. one thing id be curious about is storage management. like can it help me find huge files, duplicate folders, old downloads, or other junk i dont need anymore? thats probably a workflow id use way more often than just browsing files.

the multi-panel + live preview combo is the part I'd actually use daily over plain ls/cd. curious how it holds up in two scenarios that tend to break TUI file managers: a directory with tens of thousands of files (does listing and search stay snappy or does it need to index first), and running it over a laggy SSH connection where redrawing a rich UI can get janky compared to a bare-bones tool. those are the two things that'd decide whether this replaces my terminal workflow or just becomes a local-only nice-to-have

multi-panel plus real previews without leaving the terminal is genuinely the thing I miss from GUI file managers, nice to see it done keyboard-first instead of bolted on. question outside the performance stuff already asked - is there a shared theme/config gallery or plugin registry forming around this, the kind of thing where people post their dotfiles setup. that ecosystem is usually what decides if a TUI tool becomes a daily driver people customize for years or just a neat launch that fades

Love seeing modern TUI tools get this level of polish @zac_zuo! Breaking away from single-panel terminal navigation saves so much context-switching.

For keyboard-heavy workflows, how flexible is the keybinding engine? Can users configure Vim-like modal navigation (like relative line jumps or custom motion macros between panes), and does config reloading happen dynamically on file save without needing to restart the session?

The multi-panel workflow looks genuinely useful for moving between projects without leaving the terminal. How does superfile handle conflicts or interrupted operations when moving large batches of files between panels?

Congrats on the launch! The UI looks incredibly clean for a terminal tool. I'm always looking for ways to reduce context switching when navigating through different projects and repositories. Are there any plans to integrate Git status directly into the file views (like highlighting modified or untracked files)? That would be a massive time-saver. Upvoted! 🚀

I'm already loving the way this looks even though I haven't tried it out yet.... We need more TUI tools like this!

The multi-panel preview that actually renders images, videos, and code side by side without breaking a sweat is exactly what makes this feel like a real desktop manager in the terminal. Nicely done.

This is awesome maybe I can push more friends to linux if they had this in the terminal xD

Love that superfile brings a true multi-panel workspace to the terminal, keeping everything keyboard-driven means I never break flow to reach for the mouse.

been wanting a proper file manager that lives in my terminal and this actually delivers, the multi panel setup feels really natural to navigate

Love it. I still remember using Norton Commander back in the day. Or actually xtree, which I loved, and none of my friends could comprehend :D

Grok 4.5 is the first "model trained alongside @Cursor".

Grok Build is also now open source.

The real-time search integration feels genuinely useful, especially how it pulls in trend context without making you dig through tabs.

been using Grok for a few days and the real-time search actually feels live, like it pulls in stuff minutes old. the tone is snarkier than I expected, which is fun for casual questions.

Honestly surprised to see only one comment on Grok 4.5 release as of now. I’ve tried Grok several times but never managed to get results good enough to keep using it, especially compared with other models. I’m interested to see whether 4.5 genuinely changes that.

Introducing Estera - the AI receptionist that answers every phone call and WhatsApp message for your business, 24/7.

Service and hospitality businesses - salons, clinics, restaurants, real estate, hotels, dealerships - live and die by response time. But most calls come in while staff are with a customer, and most WhatsApp messages arrive after hours. Every missed one is a booking that quietly walks to a competitor.

❓ The obvious fix is "just add a chatbot." But anyone who's tried it knows how that feels: a robotic bot that answers an FAQ, doesn't know your prices, can't actually book anything, and makes customers ask for a human by the second message.

🤔 Turns out a receptionist is the hard part, not the chat. It has to know your services, policies and pricing, speak your customer's language, handle a real back-and-forth, and write the appointment straight into your calendar - not just talk.

So we built Estera to be exactly that, live in under 3 minutes with zero technical setup:

✅ Answers inbound calls and WhatsApp instantly, 24/7
✅ Trained on your business - services, pricing, policies, FAQs
✅ Books appointments directly into your existing calendar / PMS / CRM
✅ Qualifies leads and follows up automatically on no-shows
✅ Uses your existing phone number & WhatsApp Business number - no new hardware
✅ Speaks 50+ languages

Estera ships as three connected tools: Concierge (inbound WhatsApp), Voice (inbound calls), and Messages (outbound WhatsApp campaigns) - one AI assistant across all of them.

Try the live demo (no signup) or spin up your own agent in a minute at tryestera.com.

We'd love to hear your feedback 🙌

Congrats on the launch @fredyandrei @toma_rares ! I've tested it and it was very simple to setup and use. All the flows and visual look were great! Love it!

Fredy, your answer to Gal about the agent knowing when not to decide is the strongest thing in this thread, so I will push on the vertical list instead.

Clinics do not behave like salons or restaurants. A patient calling to book usually says why, and that reason becomes regulated health data the moment it lands in a transcript, a CRM field or a follow up message. In the US that also rules out the channel you lead with, because Meta does not sign a business associate agreement for WhatsApp, so a clinic running patient conversations there is exposed no matter how careful your agent is.

Is a clinic a separate configuration for you, or the same agent with a different script?

Congrats on the launch, Fredy!

I saw you how you were sharing updates the past few months and I respect the consistency.

In a world where businesses can improve their services with products like Estera, it's a no brainer.

The live pricing and human handoff approach is practical. How does Estera prevent double-bookings when two customers try to claim the same slot at nearly the same time across calls and WhatsApp?

the trust question that'd matter most to me isn't accents or background noise, it's what happens when Estera gets something wrong mid-booking - quotes a price that's since changed, or double-books a slot your calendar already had held. does the business get a heads up to catch that before the customer shows up, or is the first anyone hears of it when the customer's standing at the counter confused? for something handling money and scheduling unsupervised, the failure path matters more than the happy path demo

I'm curious what makes businesses switch from a traditional receptionist to Estera. When does a human intervene ?

Set it up for our front desk in like three minutes and it actually picked up on the first ring, which already beats our old system. Qualifying leads part feels solid so far, going to let it run for a week.

Timely for us — we run a healthcare directory where most patient–clinic contact happens over WhatsApp, and the hard part isn't answering, it's knowing when to hand off to a human. How does Estera decide when to escalate? Also curious how you handle regulated verticals where you can't store health details in the transcript.

The real estate use case is interesting — I manage rental properties, and the after-hours call problem is real. Tenants call at 10 PM about a leaking pipe, and you're choosing between answering every call or missing something urgent. How does Estera handle triage? Can it distinguish between "my faucet drips a little" and "water is coming through the ceiling" and route those differently? That escalation logic is where most AI receptionists fall flat.

Setup was genuinely fast, had my agent handling test calls within minutes. The natural-sounding responses caught me off guard, way better than the robotic assistants I've tried before.

Your website link is not working at the moment

It looks good. But I can imagine that Estera will need to handle some edge cases when it won't be able to answer.

Then what? Would it be able to create some kind of ticket, ready to be answered by a real person?

Estera the website is cool and easily understandable i like that and would love to test this agent how perfectly it can handle calls and disputes and how much data i have to give him and does it remember things and learns from the mistakes it did?

This is nice. How does it handle callers who specifically demand to speak to a human?

the not letting the agent commit when it's unsure approach is a smart way to frame the risk. curious about the other trigger for handoff though - what if the customer just isn't happy talking to a bot and asks for a real person directly, nothing to do with pricing or a gray area. does Estera hand off immediately on that kind of request, or does it try to keep resolving things itself for a bit first? for hospitality especially that could matter as much as the pricing accuracy does.

I build on the support side so take the bias into account. Calls and WhatsApp look like one inbox and they are not, and the follow-up promise is where that shows.

On a call the conversation has a beginning and an end. On WhatsApp it has neither. The customer replies four hours later in the same thread expecting you to remember, so it is really a record that never closes. That is a different data model to a call transcript, not a variation on one.

The specific thing I would check is the follow-up. WhatsApp only allows free-form messages inside the service window after the customer's last message. Outside it you are on approved templates. So automatic follow-up either lands inside that window, or it goes out as a template and reads like one, which for a dental clinic chasing a booking is the wrong tone.

The other one is handover. Estera answers at 2am, a human picks the thread up at 9am, and the customer has no idea a shift changed. Whatever was promised overnight has to be sitting in front of that human, or they contradict their own receptionist and the customer just sees a business changing its mind.

Neither is a reason not to build it. They are the two places where one feature behaves differently per channel, and both are cheaper to design for now than to retrofit..

The under-5-second answer time is the spec I'd zero in on, because with voice the delay before the first word matters more to callers than the raw quality of the response. I build voice AI that calls aging parents daily, and getting reliable pickup plus natural turn-taking over real phone lines turned out to be harder than the model work itself. How are you handling barge-in when a caller talks over the agent mid-sentence? That one detail tends to decide whether a voice agent feels human or not.

Editing by transcript is such a clever idea, and running it fully offline is a huge plus. One thing I'd love to see is a way to export with burned-in subtitles directly from the editor so the captions stay synced without needing a separate tool.

honestly the transcript-based editing is such a clever idea, but it would be really helpful if you could split a single speaker's turn into two clips without having to insert weird text. like a hotkey to cut mid-sentence right where the cursor is, kind of like hitting enter in a doc.

Editing video by tweaking the transcript sounds genuinely clever, especially the local-only angle. One thing that would make it way more useful for me though is multi-track support, since most real projects have narration, music and B-roll that need separate timing rather than living on one timeline.

Editing video by just fixing typos in the transcript feels like cheating in the best way. Loved that it runs offline with zero setup.

the transcript editing idea is honestly so smart, cut a sentence and the video just trims to match. loaded a clip and it ran fine without any setup.

Editing a transcript instead of a timeline is one of those ideas that feels obvious only after someone builds it. Keeping everything local makes the pitch much stronger for interviews and personal footage. And a weekend project that puts a real workflow on the table is more interesting than another polished roadmap.

How does it handle languages other than English?

Running fully in-browser is a strong differentiator for this category, since the thing that makes people hesitate about editing tools is uploading raw footage of themselves to someone else's server.

The tradeoff I'd want measured is where the tab gives up. Browser memory limits are real, and a 90-minute recording is a very different ask than a 5-minute one. Do you have a rough ceiling in practice? And does transcription run locally too, or does that step reach out — because if it does, the privacy story quietly changes for the part people care about most.

How do you handle longer videos or really dense transcripts in the browser? I'm curious if there's a practical file size limit before things start getting sluggish, since everything's running locally.

Transcript-based editing in the browser is a strong idea, especially with everything staying local.

How does Rescript handle edits beyond cuts, such as removing filler words or tightening pauses without making the final video feel unnatural?

love that this is a weekend project and fully local, that's a rare combo for video editing tools. curious how it holds up on longer source files though, like an hour long podcast recording. is the transcription and scrubbing still snappy on a mid-range laptop at that length, or does the in-browser part start to strain once the timeline gets long. either way nice work shipping this solo.

Some editing products reduce the video quality when we export the edited clip. Does this retain it as is?

So for example, if the video is in 4K, it exports back 4K, right?

Does it have all features of DeScript or only the transcript based editing feature?

Cool stuff 🚀, Don't know how you pulled this off but getting gstack agents to join meetings is like a major upgrade to @garrytan 's gstack. Even though i was stuck in the shared queue for some moment, i still had it running using my own claude code as brain, Had a call with the gstack CEO and I'm already loving it. 😄

Hey all 👋

Quick update: we’re seeing heavy usage and the shared brain pool is currently at capacity.

For now, please clone the repo, create your own brain, and run it locally.

Also, this is a self-learning agent. If you run into problems, just ask it to resolve them and it will work on fixing them.

Having agents critique a shared screen in real time is a strong idea. Can each persona retain context across multiple meetings, or does every call start fresh?

the local-brain / no audio leaving the laptop detail is what would actually get this past a company's security review, most tools in this space cannot say that honestly. curious about the live-meeting etiquette side though - when a persona jumps in to critique the shared screen out loud, how does it know not to talk over an actual human who started speaking at the same moment. is there a turn-taking signal or does it just wait for a pause

The local brain running on your own machine is the detail that actually matters here — most meeting AI tools quietly route everything through their infrastructure and call it 'private.' The open MIT license on the personas is a nice architectural choice too. The part I'd dig into as a developer is the AgentCall API: when I bring my own agent into a call, does it receive raw transcript chunks in real time, or does gstack buffer and parse before the agent sees anything? The latency and chunk boundary design determines whether a custom agent can stay coherent during a fast conversation.

Hey PH 👋 Matthew here.

Everyone is writing skills and rules for their coding agents, and almost all of them die in someone's dotfiles. localskills.sh is a registry that fixes that.

Publish a skill once, install it into Claude Code, Cursor, Windsurf, Codex, or Copilot with one command. Teams get a private namespace with permissions, SSO, and two-way GitHub sync. There's an MCP server, so your agent can pull skills on its own mid-task.

What's the one skill you'd want your whole team using?

The version question upthread has a quieter cousin, and I think it is the one that actually bites.

Drift is at least detectable. Two people compare notes and find different versions. A skill that fails to load at all, or loads a version nobody expected, produces no signal whatsoever. A missing dependency throws. A missing skill just means the agent answers without the rules that were meant to constrain it, and that answer looks completely normal. I keep my own agent rules in one global file with a stack of skills beside it, and the runs that cost me were never the ones that errored. They were the ones where a rule sitting right there on disk plainly did not reach the model, and nothing in the output said so.

So underneath the permissions and pinning questions: can a run prove which skills were in its context, after the fact? Installed and loaded are different claims, and only one of them is checkable right now.

Mine would be a pre-deploy checklist skill, the boring one nobody keeps updated. The mid-task pull is the part I would want to understand though. If a skill carries rules that shape what the agent does, whoever can publish to the namespace can change agent behaviour without a code review. Is publishing gated by the same reviewers as a repo merge, or by namespace permission alone?

The "dies in someone's dotfiles" framing is exactly right — I've seen the same happen with .cursorrules and Cursor rules files that only the senior who wrote them actually knows about. One thing I'd want to understand before rolling this out: is the skills registry self-hosted by default, or does it live in your cloud? If team-specific conventions end up baked into a skill definition — internal API patterns, data handling rules — that changes the trust model depending on where the registry lives.

the "dies in someone's dotfiles" framing is exactly right, every team I've seen has one senior engineer's excellent cursor rules file that nobody else knows exists. to answer your question - probably a PR-description skill that pulls from the actual diff instead of the commit messages, ours are always out of sync with what actually changed.

question on the mid-task pull via MCP - if a skill gets updated mid-sprint, does an agent already partway through a task on the old version get bumped automatically, or does it keep running on whatever it pulled at task start? version drift across a team seems like the thing that'd quietly cause the most confusion if two people's agents are working off different skill versions without realizing it

Would love to see a way to version-pin specific skills when installing, so a team can lock to a known-good build instead of always pulling latest. Something like a lockfile or a `--version` flag per skill would make shared agent rules way more reliable across projects.

The single install command across multiple editors is genuinely useful. Saved me from copy-pasting the same skill definitions into Cursor and Claude Code for the third time.

A nice take on making agent skills portable across editors. One thing that would help me get started: a built-in preview or diff view that shows exactly what each skill will inject into my project before I run the install command, so I can trust what is being added to my rules and configs.

A versioning system for skills would be huge, something like pinning specific versions when installing so updates don't silently break workflows. Maybe a `localskills install owner/skill@1.2.0` syntax with a lockfile would make it way more reliable for team setups.

Centralized skill and MCP management feels increasingly necessary for teams. Does localskills support signed provenance, version pinning, approval policies, and rapid rollback when a skill update causes unexpected behavior?

One install command across every tool is the obviously useful part, so I will ask about the part that tends to get hard later. Once a skill is installed across a team, how does anyone know it still does what it did last month? The underlying tools move, prompts rot quietly, and the failure mode is not an error, it is slightly worse output that nobody attributes to the skill. Do you give a team any way to see that a shared skill still behaves, or is that on the author to keep an eye on? Asking because reusable prompt assets are easy to share and genuinely hard to keep honest. Congrats on shipping this.

The thing I'd want from a registry that a dotfile never had to solve is selection. I keep 16 skills in one project, and the failure mode isn't that they die unread — it's that two of them describe overlapping territory and the agent picks the wrong one, or reaches for none and answers from general knowledge instead. Right now that's self-limiting, because installing a skill costs me the work of writing it. One install command removes that cost, and a team namespace with 200 skills in it turns the description string into the real interface: whatever sits in that field is competing against 199 others for a decision the model makes in a fraction of a second.

Do you do anything about that at publish time — collision detection against descriptions already in the namespace, or a lint that flags a new skill whose trigger surface overlaps an existing one? Abdullah's point about never being able to prove what loaded gets sharper at that size, because "it never fired" and "the wrong one fired" look identical from outside the run.

@matthew_zhao3 The governance and versioning answers above cover a lot of ground. One thing I didn't see addressed: once a skill is published and adopted, can an admin see usage data, like which skills are actually being pulled across the org versus sitting unused?

Would love to see a simple way to preview what a skill does before installing it, like a short description or example output. Right now I have to trust the repo blindly, and a quick "what you'll get" snippet would make it way easier to decide which skills to grab.

The dotfiles problem is real. The cross-tool support is what makes this genuinely useful though: Claude Code, Cursor, and Windsurf each interpret rule directives differently enough that a shared skill definition needs to either normalize across them or emit tool-specific outputs. How does localskills handle skills that need slightly different behavior per agent runtime without requiring separate skill definitions for each tool?

one thing not covered above - what happens when two installed skills actively contradict each other. say one teammate published a skill that says always squash commits and another published one that says never rewrite history. does localskills flag that kind of conflict when both get pulled into the same session, or does the agent just silently pick whichever loaded last and nobody notices until someone's history gets rewritten unexpectedly.

Seconds-to-launch is a good promise, but iMessage is the one channel where the platform has no official path for this, so the interesting question is what's underneath.

Specifically: whose number is sending? If agents share infrastructure, one badly behaved deployment getting flagged seems like it would affect everyone downstream, and Apple's enforcement isn't something you can appeal your way out of quickly. Is there isolation per customer, and have you hit any rate ceilings yet in practice?

iMessage as an agent channel skips the app-install friction entirely, which is the right wedge for consumer use cases. The Apple Messages framework has always been tricky around delivery receipts and the SMS fallback. How do you handle agent session state when a thread silently switches from iMessage to SMS mid-conversation and the agent needs to stay coherent?

pricing comparison is big plus , most of those competitors charge just for the line, so the $50 "agents and inbox included" framing is a genuinely fairer comparison than it first looks

the demo-is-the-product approach is the right call for proving this actually works. one thing I keep coming back to though: what happens if someone adds this agent's number into a group iMessage thread instead of a 1:1 support convo? group consent and expectations are pretty different from a private support DM, and that feels like something you can't really control from the API side once a customer decides to do it.

Finally got a real iMessage line for my side project without dealing with carriers, set it up in like two minutes and the API docs were actually readable. Flat pricing beats the per-message surprises I kept getting elsewhere.

the setup speed is genuinely impressive but the line in the description that catches my eye is "payments inside the thread" - collecting card details over iMessage is a different compliance bar than support/bookings. is that handled through a tokenized link/handoff so raw card numbers never actually land in the message thread itself, or is the agent parsing payment info directly out of chat text

Signed up and got a working iMessage line in under a minute, which still feels almost illegal. The plain-English setup actually understood what I threw at it without me rewriting the prompt three times.

The 30 second setup is honestly kind of wild compared to the carrier paperwork nightmare I went through last year for my business line. Really cool that it works through real iMessage too, not some random SMS gateway pretending to be Apple.

@cale_lane Getting this error message https://snap.wrld.tech/RLThWcL7NpmYBTJq7grl - would love to try this out!

Also, emailed

Love the freeze-on-hover idea, that's been a pain point forever. One thing I'd really want: a way to tie comments to a specific element rather than just coordinates. When the layout shifts between draft rounds, my pinned feedback ends up pointing at the wrong button. Element-anchored notes would solve it.

The capture half is solved here in a way I haven't seen elsewhere, so my question is about the other end of the loop. I annotate the frame where the easing breaks, the agent changes the code — and now I have to get back to that exact state to see whether it's fixed. Same hover, same interruption point in the animation, same window size. That reproduction step is manual, and it's the reason I stopped filing my own visual bugs properly: the before was cheap, the after was a chore, so I'd eyeball it and move on.

Does a .notate session carry enough to make the re-capture repeatable — app, window geometry, frame timestamp — so a second capture is comparable to the first frame by frame? A diffable before/after pair is what would turn this from a feedback tool into a regression check, and the typed manifest you described to Anand sounds like it's already most of the way there.

the freeze-frame capture for hover states is genuinely useful, that pain point always kills me with regular screenshot tools. one thing i'd love to see is a quick way to turn a pinned comment into a short loom-style video clip instead of just a still image, basically so i can show the broken animation in motion when a teammate opens the session. would make the handoff feel way more clear honestly.

The hover-state capture is genuinely useful. I was reviewing a dropdown animation that kept closing before I could comment, and being able to freeze it mid-hover saved me a ton of back-and-forth with the dev.

the honesty in this thread is what's selling me on it tbh. one thing I didn't see covered - since it's a full screen recording tool, what happens when a password field or some other sensitive text is visible in a frame that gets captured. is there any blur/redact step before the .notate file gets handed off to a teammate or an agent, or is it on the user to notice and re-record

@victor_aremu Hey! Just checked out Notate. The gesture-based approach (closing a loop to snap a shape) is super clever. Since it exports markdown for agents, I'm curious how you see the UI evolving for non-devs like designers who might use this for handoffs? The current dark minimal vibe is great for code review, but wondering if a lighter mode or web version is on the roadmap. Congrats on the launch!

the 'no replay guarantees, I will never pretend to reproduce app state' line further up is a rare kind of honesty for a launch post, most tools would just quietly overclaim there. practical question on the recording side - a long capture with per-frame data plus the pixel history for the perceptual diff sounds like it could get heavy fast. is there a size ceiling or compression pass on longer sessions, or does a 10 minute recording just produce a genuinely large .notate file you're expected to live with

Shared annotations for humans and agents could provide valuable execution context. How does Notate keep annotations attached when the underlying webpage, design, or code changes substantially?

"For humans and their agents" is the phrase doing the most work here. A human annotation is fundamentally positional — this thing, right here — and an agent has no pixels to point at, so the same note has to arrive as something structural instead.

What does the agent actually receive? Just the note text, or the note plus enough surrounding context to know what it was attached to? I'd guess "make this button smaller" is close to useless without the second part, and that's the piece that seems genuinely hard to get right.

Freezing transient UI states is the part that has always been missing. Regular screenshots lose hover states and animation frames at exactly the moment you need to point at them. The agent-readiness angle is the interesting differentiator here. How does Notate structure annotation metadata for agents? Is there a typed schema alongside the visual, or does the agent infer context from comments and pin positions?

Freezing hover states and pinning feedback to the exact frame solves a real gap in UI review. Can a teammate reopen a shared session and add or resolve annotations without needing the original capture environment?

Hi ProductHunt! 👋 I created Rivault to solve for the friction and worry of typing in sensitive details and data (SSN, payment details, phone number, emails) for CUA to automate tasks. This allowed me to store, manage and own those data without having any sensitive data stored in session logs and memory.

Please let me know what do you think about Rivault and if you have any questions! 🙇‍♂️
P.s. If you are interested in this space and would like to build together, send me a DM or reach out to me on LinkedIn!

Hyu — most of the fields I saw were consumer-shaped (passport, credit card, SSN). WinBidIQ's agent handles company UEI numbers and CAGE codes, not personal PII. Can you define custom field types in the vault, or is it built around a fixed set of common credential types?

Congrats on the launch, Hyu. The redaction covering screenshots and every trail across platforms answers the exact question I'd have asked first, most tools stop at "we don't log it" and never touch what the agent saw on screen. What I'd still want to know about the approval side: when a task needs more than one entry in sequence, say a card number for checkout, then an email for the receipt, does each pull need its own Face ID approval, or does unlocking once open a window the agent can draw from until the task finishes?

finally a clean way to hand sensitive info to AI without trusting it. the Face ID approval flow actually makes sense for booking tasks, and the zero-knowledge part gives real peace of mind.

this thread already covers the interesting security architecture questions, so here's a boring but real one - what happens on the recovery side if someone loses the device their Face ID/Passkey is tied to. is there an account recovery path back into the vault, and if so what stops that recovery flow itself from becoming the weak point that all the redaction and zero-knowledge design was supposed to avoid

@hyu_lim the auth-request + deterministic-redaction model is a real step up from pasting a PAN into a prompt — and +1 to Yuki's proxy-vs-hold point, since whether the agent ever sees the raw value or just a reference is the whole game under prompt injection.

Pushing into the payments case specifically: for a card at checkout, could Rivault hand the merchant a scoped single-use / network token instead of the real PAN? That would shrink both the PCI surface and the blast radius if the destination page turns out hostile.

And is a release bound to a destination — "this card unlocks only for delta.com checkout" — so a redirected or injected form can't draw it? Curious how you're thinking about the agentic-payment rails 👌

A dedicated context vault addresses a real problem for long-running agents. How granular are the permissions, and can access be revoked immediately with a complete audit trail of what each agent retrieved?

the whole thread is asking where the value goes after unlock, but I'm stuck a step earlier - the vault, the Face ID prompt, and the agent process are all running on the same device. if that device has anything actually malicious on it, or the CUA framework itself has a bug that lets it read process memory, does zero-knowledge storage even matter at that point, or is the entire security model resting on "trust the local machine" and everything after that is just making the honest-agent case safer

The credential injection problem in agentic workflows is real — I've seen teams paste sensitive values directly into agent prompts or env vars just to get things working, which creates a different kind of risk. Before integrating: is the vault configuration declarative, where you define which entries each task is allowed to access upfront, or does the agent request credentials at runtime dynamically? For a multi-step pipeline where different tasks need different secrets, I want to understand whether I can pre-authorize the full workflow rather than getting a Face ID gate on each individual lookup.

Everything in the thread so far treats sensitive data as fields — passport, card, SSN — and those are the tractable class precisely because they have a shape a redactor can match. The category I'd want to hear about is the one with no shape: the paragraph a person types describing their situation. A health detail, a legal problem, something about their kid. More damaging than a card number, can't be reissued, and no pattern finds it.

I build a consumer app where the most sensitive thing in the product is free text someone wrote in a bad moment, and the only answer I ended up trusting was architectural rather than filtering — certain content is contractually never allowed to leave the device, so there's no store, no log, and nothing to redact afterward. Deterministic redaction after the fact is a strictly weaker guarantee than never having transmitted it.

So: does Rivault hold context that isn't a typed field? If the vault carries a freeform note for the agent to use, what is the redaction matching against?

The hard part with secrets and agents isn't storage, it's that the moment a value lands in a context window it's effectively been disclosed. Anything that can steer the model afterwards can usually get it back out.

So the question I'd have is where Rivault sits in that sequence. Is it a vault the agent authenticates against and then holds the value, or does it stay in the path — proxying the call so the agent works with a reference and never sees the raw credential? Those are very different security properties under prompt injection, and only the second one really survives it.

The zero-knowledge vault with Face ID unlock is a clean way to solve CUA credential injection. We've thought a lot about the tension between automation speed and human-in-the-loop gates when building agentic workflows. The auth request pattern is smart but introduces a latency dependency. How do you handle timeouts or retry logic when the user isn't available to unlock and the agent is already mid-task?

That's interesting. How do you handle conflicts when multiple agents try to access the same vault entry?

Hyu, the vault is the easy half of this problem, and I suspect you already know that.

Once the agent unlocks the passport number and types it into a page, that value exists in the DOM, in whatever screenshot the agent took to see the field, and in the model context that reasoned about the page. Deterministic redaction after the task covers your own store, not the trail the agent left getting there. Doing PHI redaction in healthcare, the downstream copies were the hard part every single time, never the storage.

Does the redaction reach the agent's screenshots and page context, or does that stay the responsibility of whoever runs the agent?

Hey 👋 I'm Jonathan. I struggle with ADHD on a day-to-day (I'm sure a lot of founders do 🙃), and conversations disappear the moment they end - same with a lot of my thinking, and I hate traditional notes/the transcripts we get from traditional AI noteakers. I wanted something that organised my ideas the moment I say them out loud.

So I built Tackly - the AI notetaker that maps your thinking in real time, whether you're in a meeting, so just thinking on your own.

Dual Ingestion & Tri-Tier Real-Time Pipeline: To maintain sub-second (<80ms) visual responsiveness across both solo rants and group meetings, Tackly uses a dual-ingestion architecture covered by initial fast-path rendering (powered by Gemini 3.5 Flash Lite).

Talk Solo 💅
Hold a key, think out loud and watch a map of your thoughts (summarised, to only keep the important stuff) get built out in front of you.

Join a meeting 📆
No need to connect your calendar as soon as you sign up - invite a bot instantly into your meeting by just pasting a link. As soon as Tackly joins your meeting a board is created, and a whole teams conversation gets mapped out in real time (you can even share the board via a live link for others to watch).

Upload a transcript 🙊
Hate boring summaries from other notetakers, just upload a transcript - it gets mapped out instantly.

During the build, and once core functionality was viable enough for me to use it - I created a thread on Tackly, spoke - and I followed along with my thoughts so easily. Exporting every session as an MD to Claude, which helped shipped Tackly in less than a week. I truly believe this is the new way of thinking.

I truly believe this is the new way of thinking.

Huge congrats on shipping @lorbes Solving note-taking for visual/ADHD thinkers is a massive problem.. qq How does Tackly decide how to cluster scattered thoughts into nodes when a solo rant jumps randomly between three different topics?

The 'conversations disappear the moment they end' problem is real — I end up with Notion pages full of half-formed thoughts I never revisit. The real-time map-building demo looks like it captures the shape of an idea before it evaporates. Practical question about sharing: once I've built out a session's map, what does handing it off look like? If I'm building on something from a voice note last week and want to loop in a collaborator, is that a shareable link, a static export, or do they need to be in Tackly to see it meaningfully?

the "transcript with a summary bolted on" line is so accurate lol, that's exactly what's wrong with most of these.

quick question, when you export to MD and feed it to Claude, is that a one-off thing or are you building context across sessions? feels like that loop could get addictive once you're a few weeks in.

Congrats on the launch! The fixed 20-node taxonomy is impressive! I was thinking... T1 assigns a provisional node type in under 80ms, then T2 can re-parent it later. When I'm mid-thought watching the map build, does that correction ever visibly reshuffle a node I was still reading?

This looks really interesting, congrats on launching! The real time rendering is remarkable. How much editing can you do afterwards the first draft is done? Do you control it by voice or there is manual way to move stuff around, add additional notes, and so on?

the sub-80ms rendering target is a smart thing to optimize for specifically, since with a thinking tool any visible lag between saying something and seeing it land breaks the whole point of thinking out loud. question about the 20 nodes - is that a fixed taxonomy of note types, or an actual cap on distinct thought clusters per session? if it's the latter, what happens on a long rambling session that would naturally produce more than 20, does it merge similar ones or just get more selective about what makes the cut

the ADHD framing resonates, that line about never losing a thought is exactly what's missing from tools built by people who don't actually have that problem themselves. curious about contradiction handling though - when you're rambling and go back on something you said 2 minutes earlier, like actually no scratch that, does Tackly catch that as an update to the existing node or does it just add a new one sitting alongside the old take?

Thanks! Yeah, traditional summaries just don't cut it. Right now it's a one-off export, but cross-session context is on the radar agree that long-term memory loop would be lethal.

Jonathan, this speaks straight to anyone whose head fills up faster than they can write it down. Talking it out and finding it sorted afterward sounds like a small weight lifted.

This is such a refreshing take on AI note-taking. Turning a conversation into a live visual map feels much easier to revisit than scrolling through a long transcript or generic summary. I especially like that it works for both meetings and solo thinking sessions. Can Tackly recognize when an earlier idea comes up again and automatically connect it to the relevant node?

This is extremely helpful. Lately, I had to restart my Mac many times. I am suspicious about one particular app that is responsible for that, let's see whether I am right. Good solution for that :) Is it accur8?

honestly this looks really useful for tracking down which agents are chewing through SSDs. one thing that would make it even better is per-app historical graphs, like being able to see disk writes over the last 24 hours or week for a specific process, so you can spot patterns and figure out which tool is slowly eating your drive over time.

Finally something that shows which AI agent is hammering my SSD during long refactors. The per-process I/O view is genuinely useful and it runs light in the background.

Finally something that shows which of my AI coding agents is hammering the SSD without permission. The live file access view is genuinely useful for catching rogue processes I didn't even know were running.

honesty about the current overhead being more than you'd like is appreciated, most launches would just dodge that question. hoping the leaner version ships soon, would love to leave it running all day without a second thought

They’ve made AI coding agents completely opaque about on-disk activity. You see either the fan spinning, an out of space warning, or a build taking ages, but you can't actually see it doing anything. I like that this tool is focused strictly on attribution and does not aspire to be another ubiquitous system monitoring tool. The fact that this is open source is also kind of important - if you actually found AI-driven processes that are eating up gigabytes, you’d want to know what precisely that process is looking at.

the AI agent angle is a good hook, Cursor and Claude Code do churn through disk more than people expect. question about the monitor itself - with live file access inspection running, does FindDiskKiller add noticeable CPU or battery overhead if you just leave it open in the background all day, or is it light enough that it's not a tradeoff. would hate to fix one problem and quietly cause a smaller version of the same one.

The live file access view is especially useful for debugging unexpected churn from AI coding agents. Does it group activity by workspace or project, so I can separate normal build and cache writes from a runaway process touching unrelated files?

the AI coding agent angle is what sold me, Cursor and Claude Code do generate way more disk churn than people realize until they actually look. one thing I'm curious about beyond the live view - does it keep any history so you can look back after the fact? like if my Mac was crawling 20 minutes ago but I only notice now, is there a timeline to check what was hammering the disk at that point, or is it strictly a right-now view

A debate we had: do we just document how to scope credentials right and leave it to you - or enforce read-only even if you run it with admin creds? We went with enforcing it, despite the tradeoff of keeping up as providers ship new APIs. Pulling permission definitions from each provider (a few MB, daily by default) instead of a static map keeps coverage current automatically. Every API call is resolved to the IAM actions it needs and authorized against the provider's own policy definitions before a credential is attached, failing closed on anything it can't classify. The allowed sets come from the providers themselves (AWS IAM action simulation, GCP role permission eval, Azure RBAC role definitions, the K8s cluster own live view RBAC role) so coverage tracks the APIs as they grow. On AWS, assumed-role credentials are additionally re-vended through STS AssumeRole scoped to SecurityAudit. Every request host is pinned to its mapped service and region and the resolved IP is verified before connecting, so the agent can only reach your own infra. The code execution has no host access and every call is enforced to go through the read-only action gate or it fails.

the read-only enforcement design is genuinely more thorough than most tools that just document best practice and hope you follow it. one thing I'd want answered before pointing this at a real environment though - the answers this thing produces are basically a complete map of what's exposed and how to get to it. where does that output live once it's generated, and is there any access control on the audit history itself, since that history becomes a pretty valuable target on its own

ran a couple of checks against our AWS setup and the read-only guard is genuinely enforced, not just a claim. the per-turn sandboxed scripts feel like the right way to do this instead of hoping an MCP tool behaves.

Honestly love that the IAM-bound read-only layer is the default, not a setting you have to remember to flip on. Most security tools treat least-privilege as homework, you basically baked it into the call path. That kind of restraint takes real thought.

The promise of investigating cloud infrastructure without risking production is compelling. How do you enforce read-only boundaries, and does every finding include the evidence and cloud resources used to reach it?

fail-closed-by-default on the classification is the right call, but it raises a timing question - you pull permission definitions daily, so what happens on the day a provider ships a brand new API action that isn't in that day's pull yet. does it just refuse to classify it and block the call until the next sync picks it up, or is there some other fallback for that gap window

Shaked — the "script per turn, not one call" design is neat. Curious what happens when one call inside that script hits something outside the read-only gate — full abort, or does it just fail that line and move on? My own agent chains multi-step calls, and one blocked step derailing the whole run is a real pain.

Hi Shaked, the reassurance that it only ever reads and never touches my systems is exactly what would make me brave enough to really poke around a live setup. That peace of mind matters more than people tend to admit.

Love the "build it with an AI over SSH" approach, that's a really clever onboarding trick. One thing I'd want before committing a Pi to this full time is proper scheduled task support with local fallback if the network drops, so things like the kitchen calendar or recipe helper keep working even when the internet flakes out.

kitchen dashboard and fridge recipe helper are exactly the kind of small home automation that never gets built because setting up a server is annoying, so having it fall out of one Pi is a nice unlock. question on the phone number side though - since the agent has its own number, is there any allowlist on who it'll actually act on, or does anyone who texts that number get to add to the family calendar / whatever gets built on top. wrong number texting in feels like the actual attack surface here, more than the iMessage bridge itself

This is cool. I have a Pi and I’m going to try setting this up — I’ll come back and say if it worked or not. Congrats on the launch. More tutorials like this would be great.

Very nice, how to test?

The interesting constraint here is that iMessage doesn't really want to run anywhere except macOS. So a Pi in the corner of your house implies something else is doing the actual sending — a Mac acting as a bridge, or one of the reverse-engineered routes.

Which path did you take, and how stable has it been across iOS updates? That's the part I'd want to understand before hanging an always-on agent off it, since the failure mode isn't a crash you'd notice — it's messages quietly not going out while you assume they did.

Matching footage to heart rate, speed, elevation, and distance at the frame level is useful. How does Edit Mind handle clock drift or missing timestamps when combining footage from multiple cameras with one Strava activity?

Hey,

I'm Ilias, an indie maker and a cyclist in my free time. I take videos while I'm riding my bike using my smartphone or my GoPro camera, and I wanna sync and index my videos to my Strava to have video playback while seeing my ride stats.

I spent a couple of days working on this feature of video playback, and you can now search your videos using your Strava activity data like speed, elevation, PR, heart rate- you name it.

honestly the local search is a game changer, pulling up clips from my timeline without uploading anything feels weirdly fast. basically cuts out the whole cloud step i never liked anyway.

honestly this looks super useful for anyone drowning in footage. one thing that would make it a no-brainer for me though is adding some kind of waveform or audio-based search so you can pull clips based on what's being said, not just what's on screen

the creation-date fallback for phone footage makes sense but phone clocks are the part I'd worry about - travel across timezones without updating settings, or a phone that's just a few minutes off, would throw the match off in a way you wouldn't notice until playback looks wrong. is there any sanity check against the Strava activity's actual start/end window to catch a phone clock that's off, or does it trust the file timestamp as-is

I work out at the gym. What would be a different prompt than just "give me videos of when I was running on the treadmill"?

Hey Product Hunt 👋

Today we’re launching Repaint Socials. It lets you build a website with AI using content from your Facebook, Instagram, or Google Business profile.

Most AI website builders claim to create custom websites, but start with generic content because they know almost nothing about your business. Then you’re left to replace every detail yourself.

Repaint pulls real content from your online profiles, so the first version is actually specific to your business.

We originally made Repaint to rebuild existing websites, but we realized that other platforms also have lots of valuable information, like reviews, photos, hours, addresses, and contact info. With Repaint Socials, it can get all your information no matter where it lives online.

To try it yourself:

1. Go to repaint.com/social-import

2. Paste the link to your profile

3. Watch it gather context, then click “Build Website”

We’d love feedback. Thanks!

the pivot to pulling from Google Business, Instagram and Facebook instead of just the existing site makes sense, most small business sites are the stalest source of truth compared to what people actually keep updated on social. one question on the review-pulling part specifically - if it's importing actual customer reviews onto the new site, do you get to pick which ones show up or does it just grab whatever's there? putting someone's review on a new page without them realizing it'll get reused as marketing copy feels like a thing worth being careful about.

It looks good on paper. But it'll need enough data from Google or Facebook to deliver reasonable results.

I guess businesses with enough data there already have websites.

Would love to see a side-by-side comparison view after Repaint generates the new version, so I can quickly spot the differences instead of squinting at two tabs. Also makes it way easier to decide what to keep or tweak before exporting.

Using real business content instead of starting with generic placeholder copy makes the first draft much more useful. How does Repaint decide which information should be most prominent on the website? And congrats on the launch of course!