PH热榜 | 2026-07-21

👋 Hey Product Hunters,

I'm Tony Zhang, co-founder of Lev8 , the fastest agent to find and reach your target people and companies across every corner of the internet and reach across multiple channels.

Here are a few searches our early users have thrown at Lev8:

• “Find me VPs of Sales at fast-growing voice agent startups in the Bay Area that raised funding recently.”

• “Find everyone who starred our GitHub repo, then reach out to them across multiple channels.”

• “Find coffee shops in San Francisco with a 4.5+ rating and no website.”

AI has made it easier to build products. Getting the right people to notice them is still hard.

We felt this ourselves. Our team spent hours jumping between search tools, databases, spreadsheets, and enrichment services just to answer a simple question:

Who should we talk to, and why?

So we built Lev8.

What Lev8 does

Lev8 turns the live web into people and business intelligence.

Describe who you’re looking for in plain language, and Lev8:

• Searches across public sources

• Verifies identities

• Adds relevant context

• Surfaces the signals that explain who matters right now

What happens behind the scenes

Our crawler reaches sources that static databases miss. Parallel agents explore the web, while our identity system makes sure the facts belong to the right person or company.

The goal isn’t to create another oversized lead list.

It’s to help you discover the companies and people others miss—and provide reliable intelligence that both people and AI agents can use to make better decisions.

@Lev8 is one of the most technically grounded agent systems I’ve seen for people and company intelligence, and it works remarkably well.

The first time I tried it, I felt it had already gone beyond any general-purpose search tool I had used for finding and matching the right people and companies.

Search is an extremely long-tail problem. Lev8 handles it with a multi-agent system that can move quickly across a huge amount of scattered information and identify the exact people or companies you are looking for.

Quite often, it opens up a part of the market you did not even know existed.

It also does not stop at discovery. Lev8 connects the results to the social and outreach channels you already use, so the same workflow can continue all the way to the first conversation.

If connecting with people and companies is part of your work, give Lev8 a try. It may be the most accurate and efficient tool available for this job right now!

Congrats on the launch! I searched for a pretty unusual customer profile and Lev8 understood the request better than I expected. There were a couple of results I’d remove, but the overall direction was solid.

the "who should we talk to, and why?" framing is exactly the difficult part. as a founder, finding names is usually not the bottleneck anymore, finding people who are relevant right now and having enough real context to write something that does not feel like mass outreach is

the live web search and intent signals sound especially useful for launch prep, partnerships, and early sales. Curious how Lev8 shows confidence and sources behind each result, and how you prevent the personalized outreach from becoming confidently written spam based on weak or outdated signals :)

Really like the 'pick a play, tell Lev8 to run it' framing instead of another dashboard to configure.
How much does it actually get right on the first try vs needing you to tweak the ICP/messaging after a few runs?

Tested the launch myself, and I really liked it.

Spent about 10 minutes chatting with it and got answers to a bunch of questions I had about competitor tools. I also liked the visual lists, they make the information much easier to explore.

Still getting familiar with everything it can do.

Good luck!

How transparent are the results? Can users see the original sources behind each data point before using it for outreach?

Really liking the idea behind Lev8! It helped me discover relevant potential users for my project, which was exactly what I was looking for. One thing I noticed (and this could just have been my experience) was that none of the suggested users had email addresses available. I'm not sure if I was just unlucky or if it's because of the data source, but I thought I'd mention it. Overall, it's a really promising product, and I'm excited to see how it evolves. Great work! 🚀

Can I input a list of Typeform inbound submissions and have Lev8 instantly enrich and score them before pushing to Slack?

Congrats on the launch @tony_zhang! The live web crawler approach is a huge step up from static databases that lag by 6+ months.

Because you're scraping real-time signals (like GitHub activity or forum posts), how does the identity system disambiguate someone who maintains multiple roles—e.g., a VP of Sales who is also an advisor or founder at a stealth startup? Does Lev8 tie the intent signal specifically to their primary domain/company context before generating the outreach hook?

the outreach-side deliverability question above is a good one, but I'm curious about the other direction: the crawler side. pulling live data from LinkedIn and similar platforms at agent speed is exactly the kind of activity those platforms actively try to detect and rate-limit or ban accounts for. is Lev8 hitting these sources through some kind of licensed/API access, or is it closer to scraping, and if the latter, does the risk of a flagged account sit with Lev8's infrastructure or with the user's own connected accounts?

the three-layer qualification process is the interesting part here. most lead gen tools just give you a list and let you figure out if it's any good. the fact that you're verifying reachability before send is huge — biggest time sink in outbound is chasing contacts that bounced or changed roles 6 months ago. curious how the intent signals layer works in practice though, like what's the false positive rate on "hiring activity" triggers?

Nice launch! Going beyond "find people" into personalized multi-channel outreach is a smart move.🌟

Curious how you handle deliverability and rate limits across channels as send volume scales?

the multi-channel outreach part is where I'd want more detail before turning this loose - if it's finding someone's personal email/phone/social from public sources and then messaging them across several channels automatically, what's stopping that from tripping spam filters or straight up violating CAN-SPAM/GDPR rules on cold outreach consent? is there a built-in rate limit or opt-out handling per channel, or is that left to the user to figure out themselves?

If I base outreach on the wrong information once, I start questioning every result. How do you help users build confidence in the data they're seeing?

Solid for getting past the manual scrapping grind, the parallel agents actually return fresher results than I expected. Waterfall enrichment on a messy CSV saved me a couple hours yesterday.

Finally a way to stop stitching together five different enrichment tools. One live webset beats a pile of stale exports.

The "monitor intent signals + send personalized messages automatically" combo is where these tools live or die. Intent data is only as good as it is fresh, and personalization at scale usually reads as templated the moment volume goes up. How do you keep the outreach from feeling automated once someone's sending hundreds a day? Nice launch.

This is good for research, but can you give some tips on what to do with that information?

This is exactly the kind of tool I didn't know I needed. I've spent way too many hours manually cross-referencing LinkedIn and company sites for outreach lists. Curious how it handles smaller/local businesses (like dealerships or regional service companies) where the web footprint is thinner than a typical SaaS company. Does the agent flag when confidence is lower on those searches?

Interesting product! How does Lev8 verify conflicting or outdated information from multiple web sources before automatically sending outreach?

The enrichment looks useful, but I’d want to know how the data is verified. Can users see the source for each field, especially job title and company? That would make it much easier to trust before any messages are sent.

Congrats on the launch. Since outreach channels get connected by authorizing your own account, are those credentials scoped so only your own campaigns can use them, or is there a shared pool the agents draw from across users?

Would love a way to save and reuse my best-performing search prompts as templates, maybe with variables for company size or role. Right now I rebuild the same complex queries every time I start a new campaign, and it feels like the kind of thing this could handle automatically if I could just hit run on a preset.

How does this compare to using Clay?

Pretty cool. I tested it and the search criteria based on what I asked for is pretty on point.

Congrats on the launch! This sounds like a super useful tool - one question though. How is it different from an apollo or hunter? As an avid user of both (although I've never been particularly interested in purchasing a paid subscription from either) I'm curious as to how you guys are differentiated.

Parallel AI agents for live search beats static databases — but how fresh is the data on average? And for niche B2C (music teachers, not tech execs), does waterfall enrichment actually find contacts or is it LinkedIn-only depth?

Tried Ditto on a few landing pages and the output was surprisingly clean. One thing that would make this indispensable for me is if it could also extract and map the original page’s Google Fonts or custom font stack into the generated project automatically. Having to manually re-add fonts after cloning breaks the flow a bit.

How can you defend the fact that the page was copied? I often see people then start to argue about who is copying whom.

The interesting constraint with deterministic cloning is that it only sees the rendered DOM, so anything gated behind auth, scroll triggers or JS state never gets captured. Curious how you handle client-rendered sections that only appear after interaction.

Really cool though - congrats on the launch!

Can confirm this thing works amazing. Even nailed a complicated Lottie interaction and some complicated layouts.

Back in the '90s, we used a program called Teleport to clone websites, and we were absolutely amazed by how powerful it was. Time really flies—and so do technologies.

But honestly, this is a brilliant idea. Now we can just send every client who says, "I want exactly the same thing, just with my own logo," straight to you. 😂

The componentization is what got me, not just the clone but the fact that tokens and hover states come through intact. That level of fidelity on the interaction layer usually takes weeks of cleanup work.

Pointed it at my portfolio site and got back surprisingly clean component code with the hover states actually working. The deterministic part is no joke, second run matched the first exactly. Solid open source move with the REST API included too.

pulled up a couple of sites and the generated next code was honestly way cleaner than i expected, like the component split actually made sense and not just a wall of divs

Love how fast and clean the output is, especially that it keeps hover states intact. One thing that would make this even better is letting me diff the generated code against the live site as it changes, so I can spot regressions during maintenance without having to rebuild from scratch.

the fact that hover states and interaction details actually make it into the generated components is such a thoughtful touch. so many site cloner tools strip those out and you're left with static-looking pages

A diff view between the generated code and a hand-written baseline would be super useful so you can quickly spot where Ditto makes questionable choices, especially around accessibility or semantic HTML.

honestly this looks really useful, but one thing i'd love to see is a diff view between the original site and the generated code so i can quickly understand what ditto decided to extract or skip

Honestly this is kind of wild, I pointed it at a random landing page and got back actually clean componentized code instead of the usual div soup. Hover states and tokens coming through too was a nice surprise.

Love how fast and consistent Ditto is. One thing that would level it up for me: a visual diff mode where I can paste my own hand-written component next to the generated one and see them side by side. Perfect for cleaning up legacy markup and spotting what the AI missed or got wrong.

ran it on a cluttered landing page i had bookmarked and the component split came out way cleaner than i expected, hover states and all. the free rest api is a nice bonus for quick experiments.

Would love a "rebrand swap" mode where I paste a palette and font choice and ditto restyles every component in the generated output, so I can skip the tedious find-and-replace pass after cloning a competitor's layout. Would be a killer add for solo builders migrating inspiration into their own brand.

Nice one, congrats on the launch. This looks genuinely useful for getting past the blank-canvas stage, especially when rebuilding an existing customer site. How about the quality of the output once you start editing it. Does Ditto identify sensible reusable components, or is the main goal visual accuracy with some cleanup still needed afterward? Also, how does it handle sites with heavy client-side rendering, forms, or third-party scripts? Those are usually the parts where a clean clone gets difficult.

the determinism angle is such an underrated move, makes the whole thing actually predictable instead of gambling on a vibe-based clone every run

@samraaj_bath1 Congrats on open sourcing this! For a marketing site with a few dozen pages, does Ditto crawl and clone the whole site graph in one pass, or is it strictly one URL at a time and I'd need to loop it myself? That can change the real time/API cost of an actual migration job!

The deterministic angle is the smart differentiator here, most "clone a site to code" tools lean on an LLM and you get a slightly different result every run, which makes them impossible to trust in a real workflow. Fast, cheap, and consistent is a much better pitch. Capturing tokens and hover/interaction states rather than just static markup is also the part people usually skip and then regret. The free REST API + MCP server plus open source is a generous combo. Curious how it handles sites that lean heavily on runtime JS rendering. Congrats on shipping!

Hello!

That's a very interesting idea. How do you handle the issue of accessibility?

deterministic output is the part that sells me. every other site-to-code tool spits out something different each run, so you can never trust it in a real workflow. clean componentized next/vite plus open source is a genuinely useful combo, not just a demo.

the deterministic angle is the part that stands out to me, most "clone a site" tools I've tried lean on an AI guess and you get slightly different output every run. one thing I didn't see covered in the thread yet: if the source site gets updated after I've cloned it, is there a way to re-sync or diff against the new version, or is this meant to be a one-time snapshot you then own and diverge from?

Seriously amazing that it can extract all sorts of states including hover + focus. We could only dream of such a thing a year ago! And it's open-source?! wild!!

the deterministic angle is what caught me. most of these tools hand the page to a model and you get something different every run. curious how it holds up on messy sites where the css is basically soup. also glad you shipped an mcp server, that's honestly how i'd use this

The deterministic output is such a smart call, makes the whole thing feel reliable instead of just another flaky scraper. Love that the components and tokens come through intact.

HELL YES!!!

Pulled a few landing pages through it and the component breakdown was surprisingly clean, hover states and all. Way less guesswork than I expected for a from-scratch rebuild.

Hey PH 👋
I'm Manil, founder of CartAI. We have been building CartAI for more than a year now and today we're shipping the developer release.
The problem we kept running into: every AI agent demo ends right before the part that matters. Browser automation can navigate the web but can't pay. Payment APIs can move money but can't navigate. Your agent gets to the checkout button — and stops.

CartAI is the API that closes that gap. Integrate, point an agent at any web property, and it completes the transaction — checkouts, subscriptions, invoices, orders.

It's four products covering the full transaction lifecycle, with checkout as the wedge everything sits on:

→ Catalog — find the product: search across merchants, live variants + pricing, and checkout estimates before a cart even exists
→ Checkout — clear the order: complete on the live merchant surface, tracked to confirmation with normalized orders + webhooks
→ Payments — move the money: hosted, PCI-compliant sessions on Visa Intelligent Commerce and Mastercard Agent Pay
→ Monetization — share the upside: affiliate commission captured automatically across 70,000+ brands, attribution preserved through to the sale

Three ways to use it:
→ Automate — your agent runs the full flow autonomously
→ Embed — drop our agent into your app workflow
→ Enable — add checkouts to surfaces that never had one. Make any surface commerce enabled.

What's actually hard here isn't the navigation. It's PCI-compliant payment handling, cooperative bot-mitigation so you clear checkout instead of getting blocked (we cooperate with Cloudflare/HUMAN/Akamai/Fingerprint via Web Bot Auth and signed agent identity using Skyfire — we don't evade them), and workflows that always terminate in a known transactional state. That's the part we built.

CartAI clears real orders on production merchants: start to finish, from one API call. There's also an open-source MCP server if you'd rather drop it straight into Claude, Cursor, or your own agent.

Would love your feedback, especially from anyone building agents that need to actually transact. What would you point CartAI at first?

Docs → docs.cartai.ai

I like that checkout works without merchant integration. How do you handle unexpected checkout changes on different stores? Publishing success rates by merchant could build more confidence.

Me value the simple API approach for developers. What monitoring tools are available after checkout starts? Real time alerts could help teams respond faster.

It would help a lot if the Catalog API exposed a confidence score or freshness timestamp with each price result, so we can decide when to refetch before sending the user to checkout. Right now we have to guess whether a cached price is still good enough to trust.

The 'checkout as the wedge, no merchant-side integration' framing is the interesting part for me, since that's the piece every agent demo skips. The first thing I'd test is failure semantics on Checkout: if the agent clears a cart but the confirmation webhook never lands — merchant surface changes mid-flow, or the session times out — does the API give me an idempotency key so a retry can't double-charge, and do I get a clean failed state with no partial charge? That reconciliation is where agent payments usually fall apart in support.

The buyer-side and merchant-side approaches to this are converging fast, and the tradeoff you picked is the interesting one. Protocol-based integration only works where the merchant has already done the work, which is a thin slice of the web today, so completing on the live surface covers far more ground immediately.

The part I would watch is order state after confirmation. Refunds, partial cancellations and price changes get messy to normalize when you do not own the merchant's system of record.

Do the webhooks cover post-purchase events, or is confirmation the end of the tracked lifecycle?

completing checkout on any merchant surface with zero integration work is the hard part everyone else punts on, so props for actually shipping that instead of just a demo video. curious how disputes/refunds work in practice - if the agent picks the wrong variant or size because the merchant's page was ambiguous, does the normalized order data make it easy for the end user to catch that before it ships, or does it only surface after the fact like a normal order confirmation email would?

I'm curious how merchants feel about this. Do they see agent-driven checkouts as another sales channel, or do some still treat them as bots they'd rather block?

the cooperating with bot detection instead of evading it line is the most interesting part of this. i sell software online and my first thought was whether i'd even know an agent checked out on my site. what does the merchant actually see when cartai comes through?

The agent-rails angle is genuinely interesting, especially how you're framing cooperation with bot detection as a feature rather than a workaround. One thing that would help me trust this more: add a public status page or webhook for failed checkouts, like when a merchant site changes its DOM and Catalog pricing or selection silently breaks. Right now I'd have no visibility into why an order stalled, which makes debugging painful. A simple event stream for "checkout abandoned at step X" would close that loop nicely.

Right, first-party contact means the shopper still gets the merchant's email even when your side logged a fail. What I keep circling is the fully-autonomous case where nobody reads that email: CartAI's own record stays 'failed' while the money actually moved, so the calling agent could retry and pay twice. Do you expose a reconciliation signal, a webhook or a status I can poll later, so the agent can catch a settled-after-failed order before it re-runs checkout?

Interesting because that is a pretty hard problem to solve without being considered spam.

I am curious how it'd work in Turkey as all of the payments are 2FA message guarded.

Congrats on the launch! Checkout is the part most agent demos quietly skip, so tackling it head-on is welcome. The interesting bit for me is the confirmation loop (keeping the human in control of what actually gets charged). Good luck today!

Fail-closed is the right default, and I like that you'd rather under-claim than over-claim a success. The case that still bit us was the inverse: the charge actually settled on the merchant side while our side had already marked it failed, so the user retried and paid twice. Does the per-agent token give you a way to reconcile a 'we called it failed but it settled' mismatch after the fact, or does that fall to a manual dispute?

"Cooperating with bot detection instead of evading it" is the smart call — the evasion arms race is unwinnable long-term. The part I'd worry about: an agent completing checkout on any live surface with no merchant integration is one wrong parse away from buying the wrong thing or the wrong quantity. What's the confirmation/guardrail before it actually pays? That's the line between useful and scary. Congrats on the launch.

The charged-but-no-confirmation case Waqas raised is the one I'd lose sleep over building this. When we ran agents against checkout flows we didn't control, the timeout was never clean: the backend committed but the confirmation render died, and our side couldn't tell 'it failed' from 'it succeeded and we didn't hear back'. Since the Visa and Mastercard tokens are scoped per agent, can you reconcile against the payment network's authorization record instead of the merchant's confirmation page, so a dropped render doesn't strand a real charge?

The no-merchant-integration approach is impressive! How does CartAI handle last-minute price, shipping, or product changes before payment—and when does it ask for explicit user approval?

I like the idea, but checkout is one place where I’d still want a clear approval step. What happens if the price changes, an item is substituted, or the final total is higher than expected? Does CartAI pause and ask before placing the order?

Interesting. What framework and apps do you support today?

Nice Lanuch! How do you handle the nasty middle state where the card gets charged but the confirmation page times out or never comes back? Curious how you guarantee idempotency and reconciliation there, and how the final state a caller receives distinguishes "order actually placed" from "charged but the merchant has no record of it.

Congrats on the launch, excited to try this out

Hi @CartAI Team,

First of all, Many many congratulations for such a great launch, 🎉

Really like the approach of working with bot detection instead of trying to sneak past it. Getting listed in Cloudflare’s verified bot directory is not something most teams can pull off.

One question though. Since you’re checking out on live merchant sites, what happens if a payment goes through but the site glitches before the confirmation shows up? Does the system know not to retry and charge the customer twice? Curious how you handle those messy edge cases, because that feels like the hardest part of making this reliable.

Rooting for you guys, the idea makes a lot of sense. Great job 👏

Checkout is exactly where agent demos need hard state boundaries. The valuable bit is not just completing a flow; it is knowing when to stop, what was authorized, what changed, and how a human can audit it later.

Interesting, how are you handling payment failures ? Also, how do you handle websites blocking such AI bots (ex: cloudflare turnstile) ?

Congratulations on the launch, excited to try it out!

Congrats on the launch @maniluppal! Solving the execution wedge with cooperative Web Bot Auth rather than scraper evasion is definitely the right long-term move.

On the payments side, when an agent executes a purchase via Visa Intelligent Commerce or Mastercard Agent Pay, how is chargeback liability and dispute resolution structured between the agent developer, CartAI, and the issuing bank? If an end-user claims their agent made an unauthorized purchase within its spending grant, does the tokenized agent identity serve as cryptographic proof of authorization during a dispute?

Coming at this from the merchant side, we build support tooling there, so my head goes straight to what happens after a wrong order clears.

The thread's covered the price/stock case at checkout. The one I don't see: the order completes fine, and then it's wrong. Wrong variant, wrong address, size M when the person wanted L. On a human order the merchant's support desk just emails the buyer, confirms, sends a return label. The whole post-purchase flow assumes a human on the other end who can answer "which one did you mean".

When the buyer was an agent, who does the merchant's support team reach? The agent is an API, not an inbox. The end user is behind someone else's app. So a returns or "wrong item" ticket lands on a merchant with nobody reachable to resolve it.

Does CartAI keep a channel open back to the buyer after confirmation, or does the agent vanish at the confirmation webhook and leave the merchant holding a ticket they can't close…

the smart bit is cooperating with bot detection instead of evading it, everyone else plays cat and mouse and eventually loses. the thing i'd worry about is merchant checkout flows changing under you, but if you've genuinely solved that on any live surface, that's a real moat.

Wow, this looks extremely powerful! Definitely have to check this out!

"Watch every step in real time" is the part I care about most, because the black-box complaint is real. When a step fails halfway through a run, can you fix it and resume from that step, or does the whole agent re-run from the top? Resuming from the failed step is what makes long runs actually usable, and it is the hardest part to get right.

IMO the straight 'there's no rollback' answer is worth more than a polished one, most tools in this lane would have claimed replay works and hoped nobody tested it. QQ: when the agent cancels changes on request, is it walking back through the tool calls it made (voiding the invoice it created, deleting the record it wrote)? Congrats on the launch!

The hard part with always-on agents isn't watching them work, it's what happens after an approval. If a human approves step 4 and step 7 goes wrong, can you replay from the pause point with corrected context, or does the whole run start over?

one feedback, finding tools is slow, which i feel is not supposed to be, i feel you can fix that

Real-time step logs, approval gates, and a private server per workspace make always-on agents much easier to trust. Could teams set per-agent token budgets or automatic stop thresholds so a looping workflow cannot burn through connected model credits?

the live dashboard showing every run/token/decision is the part I'd actually use daily, most agent tools just give you a final log line and nothing in between. how configurable is the 'pause for approval on anything sensitive' rule - is that a fixed list of action types you ship with, or can you define per-workflow what counts as sensitive for your own use case?

I feel like debugging AI agents is becoming as important as building them. Have you found that users spend more time creating agents or understanding why they failed?

the pause for approval before anything sensitive part is the right call. i live in coding agents all day and the failure mode is never the work, it's the thing it does confidently that you didn't want. how do you decide what counts as sensitive, is that configurable per agent?

Real-time visibility is a huge plus here, love that. One thing I'd love is the ability to set custom approval thresholds per agent, so something like "auto-approve invoices under $500 but ping me for anything above that." Would save a lot of clicks on the routine stuff while still keeping the safety net for the bigger decisions.

One thing that would really help me trust the agents even more is a simple replay button for past runs, so I can scrub through and see exactly what decisions were made when something went sideways. Live monitoring is great, but most of my debugging happens after the fact and right now I'd have to dig through logs to piece it together. A timeline-style playback view per run would make audits and post-mortems so much easier.

honestly the real-time step view is genuinely cool, but i think a simple way to export the full run log as csv or json would be super useful for auditing later on

honestly the live step viewer sounds super useful for debugging agents. one thing i'd love is the ability to rewind and replay a specific portion of an agent run instead of just watching it forward, especially when something weird happens and you wanna trace back what triggered it

The live step viewing is genuinely useful, way more transparent than other agent builders I've tried. One thing that would seal the deal for me is a simple test mode where I can replay a past run with tweaked inputs, so I can fine tune prompts without spinning up real tasks each time.

Love that I can actually see what the agent is doing in real time, that's the missing piece in most no-code tools. One idea: add a way to export the full execution log as a shareable link so I can send a client a recap of what the agent did without screen recording. Would save a ton of back and forth.

Love how every workspace gets its own private server, that kind of isolation is honestly rare for no-code tools and shows you clearly thought through real business use cases, not just the demo.

I follow the founders on X and they’re both very inspiring and creative

I already created my account on their platform and it’s a gem

Go go guys

Whats the top 3 most popular use cases for this system?

honestly the step-by-step visibility is what sold me, like finally i can see what the agent is actually doing instead of just hoping it works. the approval pause before sensitive stuff feels solid too

the live step-by-step view is genuinely useful, you can actually see what the agent is doing instead of guessing. pausing for approval before sensitive actions feels like the right default.

One thing I'd love is the ability to set custom rules for when an agent should pause for approval, beyond just "sensitive" actions. For example, letting me define thresholds like "pause if this invoice is over $500" or "ask before sending to a contact I haven't emailed before" would make the approval flow way more useful for my actual workflows.

The live step-by-step view is genuinely useful, but it would be great to have a simple way to roll back or replay a specific run if something goes wrong mid-task. Right now if an agent gets stuck or makes a bad call early on, I have to start the whole thing over. A "restart from step 3" button or a rewind slider would save a lot of time when debugging automations.

honestly the live step-by-step view is kind of addictive, like watching a little worker do my tasks in real time. the approval pause thing also feels genuinely thoughtful rather than just bolted on.

The pause-for-approval step before sensitive actions is such a smart call, honestly. A lot of no-code agent builders skip that and just let things rip, so it feels like the team actually thought through what people would be nervous about handing over to automation.

"Watch every step in real time" and "pause for approval" are the two things keeping me from trusting AI agents with actual business tasks. Most tools just give you a success/fail flag and hope you don't notice what happened in between.

Love the "pause before sensitive actions" approach. Curious how it plays with the 24/7 long-running tasks though, when an agent pauses for human confirmation, does it just hang and hold resources, or is there a timeout / fallback if no one responds in time?

Nice approach. Does the approval gate still work if an agent runs while you're asleep?

Rerun sounds aimed at a pretty wide set of use cases, from Marketing automation to Engineering & Development and Productivity. How do you think about helping new users choose the right starting point for their first AI agent? Is the product more template-driven by category, or does it guide people from a task description into an agent setup?

Voice control for work is powerful when it treats speech as intent, not just text input. The useful boundary is letting the agent capture messy instructions, then make the next action reviewable before it mutates calendars, tasks, or customer-facing work.

The combination of tasks, docs, and calendar in one place actually feels useful rather than bloated, and I liked how quickly I could switch from a project view to my schedule without losing context.

The new vocal assistant has seriously upgraded my work as a freelance since @jmq gave me early access. I've always dreamed of handling more work simply with my phone (I use Routine to coordinate my work with all my clients).

I would love to customize some parts of the interface to make it fit even better with my work!

Keep up the good work and excited to see what's next!

The calendar view feels a bit basic compared to the rest of the platform. Adding a timeline or Gantt-style view would make it way easier to see how tasks and projects overlap, especially when juggling multiple deadlines. Would love that as an option alongside the existing layout.

Cool app, congratulations on the launch, it this a completely new app or does this integrate with existing google/apple cal, tasks etc ?

the sidebar layout is genuinely satisfying, everything lives exactly where you'd expect it to. rare to see this many features packed in without the UI feeling cluttered.

the example command in the launch bundles three separate actions in one utterance (move meetings, block time, set a reminder). if one of those three fails to resolve - say the Friday slot is already double-booked - does it still summarize and confirm the two that worked, or does the whole chain get held up waiting on the one that's ambiguous?

Love that a voice command can actually update your tasks and calendar

voice control for calendar and tasks sounds great until you say something ambiguous and it moves the wrong meeting or deletes the wrong reminder. with something like 'move my afternoon meetings to Friday' - does it show a confirmation of what it's about to change before committing, or does it just execute and you find out after the fact if it misheard something?

I wonder if the biggest challenge is changing habits. Most people already have a way of managing tasks, even if it's inefficient. What has convinced users to switch?

honestly looks solid, the calendar + tasks combo is what i usually end up juggling across like three apps. one thing that would be really useful is a quick "focus mode" that hides everything except your current task and its linked docs, basically a stripped down view when you actually need to get stuff done without distractions.

The way you bundle tasks, docs, and calendars into a single clean interface without it feeling cluttered is really impressive. Took me about ten seconds to figure out where everything lives.

voice for capturing tasks makes total sense, i lose half my ideas walking between meetings. the part i always doubt is corrections. when it mishears a name or a date, can i fix it by voice or am i back to typing? that flow is what would make or break it for me

honestly looks solid, the all in one angle is appealing. one thing i'd want is a built in pomodoro or focus timer that syncs with the task list, kind of a small quality of life thing but it would make routine feel more like the hub for actual work rather than just organization.

One thing that would make Routine a lot more useful for me is a built-in time-tracking view that lives inside tasks, so I can see how long things actually take without bouncing to a separate timer app. A simple weekly breakdown by project would be enough.

honestly looks solid, the all-in-one angle is super appealing. one thing that would push this over the edge for me is a proper meeting notes template that auto-links to the relevant task or project page, basically turning a random note into actual traceable work without extra clicking.

Honestly kind of surprised how well the calendar and tasks fit together, like everything just lives in one place without feeling cluttered. The wiki is pretty barebones though.

Does Routine AI ask for confirmation before making changes like moving meetings or updating tasks, or can it do everything automatically?

Congrats on the launch, Routine looks like a really clean way to keep everything in one place. One thing I'd love to see is native time blocking inside the calendar view, like dragging a task from a project onto a time slot and having it actually reserve that time. Would make planning my week way less of a juggling act between tools.

The all-in-one approach is appealing, but a native time-tracking feature built into tasks would really set it apart. Being able to start a timer on a task and see aggregated reports by project or client without needing a separate tool would make Routine a true one-stop shop for freelancers and small teams like mine.

Love how Routine pulls everything into one place, the calendar + tasks combo has already replaced three tools for me. One thing I'd love to see is a lightweight inbox-style view that surfaces comments, mentions and assigned items across projects in a single feed, so nothing slips through when work moves fast.

Love how you can keep tasks, docs, and calendars in one place. One thing I'd love to see is a quick "weekly review" view that pulls together completed tasks, updated docs, and upcoming meetings so I can actually reflect on the week without jumping around.

Finally checked Routine out and the calendar hooked up nicely with the task list, which is more than I can say for most tools. Wish the wiki formatting was a touch richer.

Love that Routine combines tasks, docs, and calendars in one place, it would be great if you could add a "focus mode" that hides everything except your current task and a simple timer to track deep work sessions without leaving the app.

This video looks so good, but does it support any other language, or how does it remind us by notification or something?

One thing that would really help me as a user is a built-in time tracking feature inside each task, so I can see how much time I actually spend on a project versus what I estimated. Right now I have to jump to another app for that.

voice for capturing tasks is the rare spot it genuinely beats typing, cause the friction of opening an app is exactly when you forget the thing. 5th launch, clearly iterating hard. nice one.

Hey Product Hunt! 👋

I'm Mathew, founder of The Letter Company. Today we're launching Phantomstory: our coolest product yet.

I've been a content marketer for the last four years, and I kept seeing the same pattern: one of the most effective strategies for winning AEO is third-party content. ChatGPT and Claude are smart cookies. They know a brand will always rank itself #1, so they trust third-party sources instead.

But here's the uncomfortable part: most "third-party" coverage today is a paid product placement, sold to the highest bidder.

What if third-party content could actually be honest?

That's why we built Phantomstory. With just two clicks, you can launch a complete third-party publication: a real media site that produces genuinely helpful content for readers, with honest placements of your brand on the articles where it actually belongs. It uses incredible technology: adversarial writing kernels, a smart KD-sensitive priority queue, and an amazing canvas tools to create stunning visuals.

Here's what you can do with Phantomstory:

🚀 Launch a publication in two clicks
Spin up a full third-party media site, complete with its own editorial identity. No agencies, no writers' rooms, no six-month content calendars.

✍️ Write with editorial voice, not marketing voice
Our writing kernels produce copy that reads like a practitioner wrote it for other practitioners. No fluff and no sales language.

🔍 Verify everything before it ships
Research agents fact-check claims and statistics so your publication earns trust instead of burning it.

🎨 Enhance every article automatically
Enhancement agents add visuals, citations, comparison tables, and the metadata and structured data that LLMs actually parse.

🤝 Place your brand honestly
Your product gets mentioned on the articles where it genuinely deserves a mention, next to fair coverage of alternatives. Readers get real value, and AI assistants get a source worth citing.

Phantomstory is built for devtools comparing themselves against alternatives, platform companies explaining a category they're creating, and any B2B brand whose buyers now ask ChatGPT instead of just Google.

Dozens and dozens of companies are already using it since launch last week, and we'd love for you to be next.

👉 Try Phantomstory: https://phantomstory.com

Drop your questions in the comments. I'll be here all day and would love to hear how you're thinking about AEO. 🙌

Mathew

I like the idea of building third party blogs so quickly. How do you protect domain trust over time instead of creating content too fast? Publishing steadily with clear expertise could make recommendations more reliable.

I've always believed consistency beats volume, so the incremental publishing angle caught my attention. It sounds like a smart way to build credibility over time instead of chasing quick wins.

Mathew, I run marketing for a healthcare AI company, so I read this with equal parts curiosity and caution.

In our category the sources LLMs cite are held to a higher bar: health content gets extra scrutiny from both models and regulators, and a fresh domain with no named authors rarely makes the cut.

The launch copy leads with honesty, which I respect, so here is the question that decides it for me: is the relationship between the publication and the brand disclosed to readers, and how do you think about that for regulated industries like healthcare?

I'm seeing more marketers realize that AI visibility depends on discoverable, trustworthy content. What stood out for me is the focus on incremental publishing rather than dropping everything at once. That feels more natural.

I really like the idea that a brand is only included where it genuinely belongs, rather than being forced into every article. Especially in a space where credibility is so important, that feels like a thoughtful approach and much closer to a real publication than a typical branded content operation.

What signals does Phantomstory use to decide whether a customer’s product should be mentioned in a specific article, and when does it deliberately leave the brand out?

Definitely adding this to my list of marketing tools to explore. Curious to see how early users leverage it for AI search visibility.

Congrats on the launch, this is a clever take on AEO. One thing I'd love to see is a way to pick the niche or topic cluster before the domain is provisioned, so the auto-generated content actually matches the voice and positioning of my product instead of feeling generic. Right now I worry about getting a credible-looking blog that still drifts off-brand.

Neat product! The 1 to 2 articles a day pacing is the detail I'd personally have gotten wrong if I built this:), every instinct says flood the fresh domain on day one! How's the configurable disclosure? Have you guys measured whether the disclosed publications get cited less by ChatGPT and Claude than the undisclosed ones?

clever tactic, but the thing i'd watch is whether AI answer engines eventually treat coordinated blog networks the way google learned to treat PBNs. search spent 15 years getting allergic to exactly this shape. if the answer engines stay citation-hungry and don't care, real edge for a while. if they catch the same allergy, fresh-domain hubs age fast. genuinely curious which way it breaks.

i think the simple workflow makes adoption eaiser. how do you monitor domain performance after launch? a dashboard with growth suggestions could help teams improve results over time.

There is definitely something interesting in the idea of building useful third-party content rather than relying only on a company blog. The part I’m curious about is independence. If the brand creates and funds the publication, how do you stop it from becoming branded content wearing a media hat? Do publications clearly disclose who owns them, and can the system ever produce a conclusion that does not favour the company behind it? That feels important if the goal is genuine trust rather than just another AEO tactic. Congrats on the launch!

Interesting angle. We run a content-heavy healthcare directory and invest a lot in original medical content written by physicians — curious how you handle E-E-A-T signals on a third-party blog? For YMYL niches like healthcare, author credibility seems like the hard part, not publishing speed. Congrats on the launch!

Phantomstory sounds aimed at teams trying to win AEO without building a whole publishing setup. When you say “launch a third-party blog,” how much control does the user get over branding, topics, and the structure of posts? Curious where the two-click setup ends and where ongoing editorial choices begin.

This is an interesting concept in the world of AEO, especially since you can do it fairly quickly. I don't see it on your website, but how does pricing work for this? Also, how do you handle protecting domain health and authority? Or is this is a "quantity of AI blogs" as a strategy type of approach?

Wow but i would love to see website in light theme also

I like the positioning around honest brand placement instead of forcing a product mention into every article. If AI search continues rewarding credibility over promotion, that philosophy could become a real advantage.

The combination of automated publishing, editorial quality, and AI-search optimization makes Phantomstory stand out in a crowded marketing space. Wishing the team a successful launch! 🚀

the Hearst comparison undersells the difference though - Hearst's ownership of a magazine is public record even if the logo isn't on every page. with a fresh domain and no named authors, is there actually anything a skeptical reader or a journalist could find that traces the site back to the brand, or is the undisclosed version designed so that trail genuinely doesn't exist anywhere?

Honestly the speed surprised me, I had a full blog up on a fresh domain in just a few minutes and the AI-search angle feels genuinely useful rather than gimmicky.

honestly this looks really useful for anyone tired of just tracking their brand mentions in AI search. one thing that would make it way more valuable for me is a built-in dashboard showing which articles are actually being cited or pulled into AI responses over time, basically closing the loop between publishing and real LLM pickup.

Setting up a fresh blog on a new domain took literally a few minutes, which still surprises me. The articles read naturally and weren't obviously stuffed with brand mentions.

Spun up two blogs on new domains in under five minutes and the articles actually read like they were written by someone who knows the space, not generic AI filler. Curious to see how it holds up over a few weeks of incremental publishing.

The disclosed-vs-undisclosed citation gap you shared is the most interesting number in this thread. One thing I'd watch: a fresh domain isn't in any model's training data, so every citation it earns comes purely through the live retrieval and grounding layer, not baked-in knowledge. That makes it far more exposed to a model swapping its search backend than an established domain is. Have you seen citation rates jump around when ChatGPT or Perplexity change how they ground their answers?

honestly the speed kind of surprised me, had a blog up on a fresh domain in like three minutes flat. still wrapping my head around the aeo angle but it feels more actionable than just tracking metrics.

the "third-party blog on a fresh domain" part is what gives me pause, if the whole point is that it reads as an independent voice recommending your product, isn't that an undisclosed conflict of interest by design? curious how you're thinking about disclosure here, or whether disclosing kind of defeats the purpose since the credibility comes from the blog not looking affiliated in the first place.

any real examples of this you can share? how is phantomstory using phantomstory to win AEO?

I'm Aiden, co-founder and CTO at TwelveLabs. Super excited to share Jockey with you today.

Most media search is still metadata search: filenames, timestamps, maybe some object tags from an older computer vision model. None of that captures what's actually happening in your footage: who's in a scene, what they're doing, the context, dialogue, on-screen text. For that you need models that natively understand time and space in video, not a bag of sampled frames.

That's our stack. Marengo, our embedding model, resolves a query like "the moment we almost missed the flight" to real retrieval across video and images, not keyword matching. Pegasus, our video-language model, segments an entire video on a schema you define and returns structured, timestamped moments. Jockey is a unified agentic system that reasons across your videos and images: a reasoning model plus a memory layer that builds a knowledge store from your corpus, so it can decompose a query, retrieve, segment, and reason across the whole thing.

The point is a corpus-level understanding you can act on. Point Jockey at thousands of videos and images, say "cut me a highlight reel" or "pull the best viral moments," and it comes back with timestamped cuts you can use. The model-only approach can't do that as dumping one video into a context window is bound to a single file, and a single forward pass runs out of room fast. It can tell you about one video; it can't reason across your catalog or build a reel from thousands.

Because the models are what we ship and improve continuously, Jockey's reasoning and retrieval quality improve as we push new versions, meaning no re-integration on your end.

Two ways in:

• MCP server: connect Jockey as a tool in Claude and query your library directly. ChatGPT coming soon.

• API: full programmatic access to build custom retrieval or agent workflows on your own library.

This is a research preview, so if you hit edge cases (ambiguous queries, retrieval misses, latency) I want to hear about them. Let us know anytime! 

Best,

Aiden 

Great name Aiden! The "no tags needed" pitch is the part I'd want to/ am going to stress test. Natural language search over video is great for discovery, but for compliance/legal use cases you usually want deterministic, auditable categories, not a model's best guess at a scene. Curious how consistent Pegasus's segmentation is run to run on the same clip.

The compositional queries are where library-scale video search tends to break. Single-concept stuff like 'red car' works fine off embeddings, but 'the moment right after the door opens' needs temporal grounding that flat similarity search can't reach. When we built multimodal search over video the causal and ordering queries were exactly where recall fell off a cliff. Does Jockey's agent decompose those multi-step queries and reason over ordering, or is retrieval a single embedding lookup under the hood?

honestly the search is way sharper than i expected, like i threw in a random cooking video and it pulled out the exact moment they mentioned "fold in the eggs" without any tagging. pretty cool to actually see video understanding feel useful

would be cool if you added a built-in timeline view so you can jump straight to the exact moment a search result happens in the video. right now i think you only get timestamps in text, which is helpful but still requires manual scrubbing. basically a visual scrubber with the matched segments highlighted would save a lot of clicks.

the corpus-level reasoning is the interesting jump here, most video search tools stop at "find the clip" and Jockey is going further to "reason across everything and build the reel." one thing I'm curious about that's different from the privacy/consistency questions above: as you keep adding new footage to a library over time, does Jockey only need to embed the new additions, or does growing the corpus mean periodically reprocessing the whole thing to keep the reasoning layer coherent?

The "understands your whole media library" angle is really compelling. I already have product screenshots, launch videos, screen recordings, demos, and random clips scattered everywhere, and finding one specific moment usually depends on remembering the filename or roughly when it was created.

Searching by person, moment, or context feels much closer to how people actually remember media. connecting that through MCP is interesting too, because an agent could finally find the exact clip or screenshot needed for a task instead of asking me to dig through folders.. :) Curious how Jockey handles privacy and indexing for very large personal or company libraries.

Was lucky enough to get access to Jockey a week or so ago (thank you team!), and have been blown away at the use cases we've already uncovered. It's changing the way we look at creative strategy across our roster of clients, and we've found some novel ways to extrapolate learnings that are informing some of our performance marketing campaigns that are already showing positive uplift.

Go TwelveLabs!

Finally gave this a spin on a few clips and the semantic search actually nails what's happening in the scene, not just matching keywords. Impressed it picked up on subtle actions without me tagging anything.

A live collaboration mode would be huge, letting a team tag and comment on different timestamps together while the AI pulls those notes into a shared summary report.

Honestly impressed by how well it picks up on subtle visual cues in long videos, not just obvious keywords. Searched a 40 minute documentary for a specific gesture and it nailed it in seconds.

One thing that would make this way more useful for me: a timeline-based search view where I can scrub to the exact moment a concept appears, instead of just getting text hits. Right now it sounds like the results are summaries, but for video editing workflows I really need frame-accurate jumps. Would love to see that built in.

honestly the search across hours of footage feels almost scary good, like you throw in a rough idea and it actually pulls the right moments out without much fuss.

honestly the search is way better than i expected, threw in some random clips and it actually pulled out the exact scenes i was thinking of. pretty wild that you can basically ask it questions about what's happening in the video.

A browser extension that lets you right-click any video and instantly get a chapter breakdown or summary using your Pegasus model would be huge, especially for longer YouTube content or lectures where I don't always want to scrub manually.

finally a video ai that actually finds the moment i describe instead of just dumping timestamps. tried it on some old footage and the semantic search picked out exactly the scene i was thinking of.

Searched a bunch of old vacation clips by typing "sunset over water" and it actually pulled the right moments, which kind of startled me. The natural language search feels way more useful than tagging everything manually.

The way the search results show exact timestamps with preview thumbnails makes me feel like I'm scanning a real video library, not just a text index. That attention to temporal precision shows serious craft.

honestly the search looks solid, but it would be super helpful if you could save and share specific video clips or search results with timestamps baked in. basically a way to send someone straight to the exact moment in the video rather than just linking the whole thing and saying "go to 4:32". that would make it way more useful for team collaboration

A live collaboration mode where teams can annotate and tag specific video segments together in real time would be huge. Right now analysis feels like a solo task, but most of our video review happens in group settings where marketers, editors, and strategists need to align on what they see.

Would love a timeline-based annotation view where I can click any moment in a search result and instantly see the surrounding visual and audio context. Right now results feel like a black box of timestamps, but seeing a quick visual snapshot before clicking through would make reviewing long footage way faster.

Hey Product Hunt 👋

I'm Pratik, CTO and co-founder at CreateOS.

CreateOS Sandbox gives every AI agent its own isolated environment with its own guest kernel. Egress is enforced in-kernel via eBPF, from outside the sandbox. Code running inside cannot route around the policy, even when fully compromised.

What's in it:

eBPF egress control — allowlist by host, IP, or CIDR, enforced in kernel

Fork — snapshot a running sandbox, clone it in milliseconds

Encrypted p2p mesh networking, plus VPN back to your local machine

BYO-S3 — mount any S3 service as a shared disk across sandboxes

CLI integration — `stripe projects add createos/project` provisions a project with credentials already in your .env

BYO-Infra — run sandboxes on your own hardware

We shipped 51 real-world examples: Claude-managed agents, multi-node clusters, batch inference, ffmpeg transcoding, and more.

We also shipped a Sandbox plugin for Claude this week. Claude can run untrusted code in a box that self-destructs when idle.

Get started:

CLI: `createos sandbox create`

Dashboard: https://createos.sh/app/sandbox

Quickstart: https://nodeops.network/createos/docs/Sandbox/Quickstart

500 free alpha credits, no card.

One thing I'd like feedback on: which of your production egress rules would this sandbox break? That tells us more than a star rating.

Links:

Docs: https://nodeops.network/createos/docs

Dashboard: https://createos.sh/app/sandbox

CLI: https://github.com/nodeOps-app/c...

SDK: https://github.com/NodeOps-app/c...

SDK examples: https://github.com/NodeOps-app/c...

Claude plugin: https://github.com/NodeOps-app/c...

Sandbox as self-hosted GitHub Actions: https://github.com/NodeOps-app/c...

Super proud of the entire team for shipping this 🤩🎉

As the marketing lead here at CreateOS, seeing all the hard work go into building CreateOS Sandbox for secure, high-performance AI agent isolation has been incredible.

Can't wait to see what amazing agents developers build with sub-30ms provisioning and bulletproof eBPF egress control. Let us know what you think.

Been on the infra side of this build, and it's honestly wild how much our CI pipelines sped up once we moved ephemeral test runners over to these sandboxes. The self hosted GitHub Actions integration saved us a decent chunk on runner costs, and not having to hand write egress rules anymore because eBPF handles it at the kernel level has been a relief for our on call rotation. Curious to see how the BYO infra option holds up once more teams start running their own hardware behind it, that was one of the more requested features from folks on our side

Been using CreateOS Sandbox, some of the things that stood out to me
1. Seamless integration with my existing managed agents to have an isolated execution environment
2. I can run these sandboxes on my own Cloud, existing compute layer
3. The examples repository is really good with variety of use-cases -> https://github.com/NodeOps-app/createos-sandbox-sdk/tree/main/examples

Great work by Pratik & team.

If your AI agents run code you don't fully trust, this is worth a look. Each agent gets its own isolated sandbox with a real guest kernel, ready in about 30ms, so isolation stops being a tradeoff against speed.

The smarter part is the egress control. Rules are enforced in kernel via eBPF, from outside the sandbox, so even a fully compromised agent can't get around your allowlist. Add in millisecond sandbox forking, your own S3 as shared storage, and the option to run it on your own hardware, and it covers most of what actually breaks in production agent setups.

There's already a Claude plugin so Claude can run untrusted code in a sandbox that wipes itself when idle, plus 50+ real examples to build from. 500 free alpha credits, no card needed, so it costs nothing to test against your own workload.

The in-kernel eBPF egress enforced from outside the sandbox is what makes this actually trustworthy to me — most sandboxes run policy in a userspace proxy that compromised code can just route around. Since eBPF sees packets at L3/L4, how does the host allowlist resolve names: is DNS resolved outside the box and the resulting IPs pinned to the rule, or can a compromised process inside do its own resolution and point an allowlisted hostname at an attacker-controlled IP? And on Fork, does the clone inherit the parent egress policy the instant it snapshots, or is there a window before the eBPF rules attach to the new sandbox?

Hey everyone, Mrudul here from CreateOS. I spend most of my time talking to companies that want to run AI‑generated, untrusted code in production, and the same obstacle keeps appearing. It’s rarely about whether the code runs; it’s about what the code can reach, where the data resides, and who can see it.

That’s the problem I care about with CreateOS Sandbox. Each sandbox is an isolated environment, and the most useful feature is the ability to create a private network across many of them. You can stand up a cluster of agents that communicate over private DNS while remaining sealed from everything else, something most teams assume requires a full VPC and weeks of setup.

The other key advantage is control. You can run the entire stack on your own hardware, including the control plane and storage, and we are SOC 2 Type II and ISO 27001 certified. For teams undergoing a serious security review, that distinction can be the difference between a trial and a polite “no.”

As AI agents become more autonomous, we kept running into the same problem: how do you let them execute arbitrary code without putting the rest of your infrastructure at risk? We built CreateOS Sandbox to answer that. Every agent gets its own isolated environment with a dedicated guest kernel, and network access is enforced externally through eBPF, so the policy still holds even if the code inside the sandbox is fully compromised.

Beyond isolation, we focused on making it practical for real workloads. You can fork running sandboxes in milliseconds, mount your own S3-compatible storage across environments, connect securely through encrypted networking, or even run everything on your own infrastructure if that's what your deployment requires.

On the product side, the whole bet here was refusing the usual tradeoff.

Every sandbox tool makes you pick: fast provisioning or real isolation. We wanted both ~30ms to spin up, and a real guest kernel per agent with egress locked down in-kernel via eBPF, enforced from outside so a compromised agent can't route around it.

If your infra can't move at the speed your agents think, the agents aren't fast, they're just waiting. That's the problem we set out to kill.

500 free sign-up credits, no card. Genuinely want to hear what breaks against your workloads.

been using this from last couple of weeks and the best thing about this is that its insanely fast, agents can spin up the sandbox in sub-milliseconds
great work by team

Congrats on your product and the launch. This sounds interesting. But how does it compare to Docker or Firecracker?

Hardware-isolated sandboxes are a strong direction for agent work. The practical question I always look for is whether the environment makes rollback, artifacts, network access, and final-state evidence obvious enough that a small team can trust the result without babysitting it.

30ms startup is great, but for agent workloads I'd also want a teardown receipt: which egress rules fired, what credentials were present, and whether the writable layer was actually destroyed. Is that available per sandbox today?

The 30ms cold start is genuinely impressive for hardware isolation, basically unheard of in that category. Loving that you threw in the Claude plugins and SDK examples right out of the gate too, makes it way less of a chore to actually try it.

The 30ms cold start actually held up when I spun up a few sandboxes back to back, no weird warm-up lag. Loved having the SDK examples handy instead of digging through docs.

The 30ms cold start actually held up in my testing, which is wild for hardware isolated sandboxes. The claude plugin integration made spinning up agents feel almost frictionless.

The 30ms cold start is genuinely impressive for hardware isolated sandboxes. Whoever tuned that cold boot path clearly obsessed over it.

The ~30ms cold start is genuinely impressive, especially for agent loops where latency compounds. One thing that would save me a lot of time: a local dev mode that spins up a fake sandbox emulator so I can iterate on agent logic and exception handling without burning real compute credits during testing.

The 30ms cold start is genuinely impressive, especially the p90 timing. One thing that would help me as a builder is a built-in state diff or snapshot viewer in the CLI. Being able to compare filesystem and env changes between two sandbox runs without having to write custom diff scripts would make iterating on agent behavior much faster. Maybe something like `sandbox diff run-123 run-124` that outputs a clean summary.

Super useful for a whole host of scenarios!

Curious if you plan to give more integration options for the plug-n-play style agents? Looks like currently we just would connect / control with Telegram?

Spun up a sandbox in around 30ms like they claim, the CLI felt snappy and the SDK examples actually made sense for once.

A built-in cost dashboard would be super helpful, especially showing compute time per sandbox session and monthly burn. With cold starts at 30ms there's a real risk of accidentally spinning up thousands of sandboxes during testing, and right now there's no easy way to see what's running or set spend alerts.

Spun up a sandbox in about 30 seconds and the hardware isolation gave me real peace of mind for running untrusted agent code. The CLI felt snappy and the SDK examples made it easy to wire into my existing flow without much fuss.

30ms p90 for hardware-isolated spin-up is a serious number — most sandbox stacks I've tried are an order of magnitude slower, and it changes what you can do per-task vs per-session. Curious about the lifecycle model: when an agent needs state to persist across runs (a working directory it comes back to tomorrow), do you snapshot/restore, or is the intended pattern ephemeral-always with external storage? Building long-running agents, that's the decision that shapes everything downstream.

i run coding agents most of the day and the thing that always makes me nervous is what they can touch. hardware isolation with egress enforced inside the sandbox feels like the right way round. does the 30ms spin up hold once you're pulling real dependencies in, or is that a bare box number?

A native MCP server for the sandbox so agents can spin up isolated environments on the fly without bolting on extra glue code. Would make the CLI and SDK even more plug and play for Claude and other MCP clients.

I will take a look because anything has the be better than MS Outlook these days! Does it handle multiple email accounts?

love that it's local-first and tiny, but honestly a unified inbox view would be a game changer for me. like, one screen where i can see all my accounts side by side instead of toggling between them. would make it feel even more minimal and way faster to triage.

Honestly the 5 MB installer is kind of wild for an email client, opened it up and was reading mail before my coffee was ready. The bring your own AI key approach feels right too, no weird data stuff happening behind the scenes.

Love that you kept it under 5 MB while still shipping full-text search and AI drafting. The local-first approach with bring-your-own-key feels rare these days.

Love how lightweight this feels, especially coming from bloated desktop clients. One thing that would make it a daily driver for me: add support for PGP or at least S/MIME signing and encryption. Local-first is great, but most of the work I get requires signed replies, so I keep having to fall back to Thunderbird for that.

5mb installer and sub-second cold start on windows is honestly kind of wild, my current client takes forever to open. love that ai is opt-in too.

been using it for a few days and honestly the instant full-text search is kind of wild, way faster than outlook. the minimal vibe is actually refreshing, feels like it respects my time

Sub-second cold start from a 5 MB installer is genuinely impressive craft, especially with native search baked in. Really respect the discipline of cutting the calendar and rules engine instead of bolting them on.

Tried it on my old laptop and yeah, the cold start is genuinely instant which is wild for an email client. Love that the AI stuff is opt-in instead of shoved in my face.

Love that you kept the scope tight enough to ship a 5MB native Rust client that still feels full-featured. Bringing your own AI key is a smart move, treats the model layer like infrastructure instead of locking people in.

The 5MB Rust install is genuinely impressive, and it actually opens faster than Outlook does. Curious to see how the BYOK AI drafting holds up on longer threads.

finally a windows email client that doesn't try to be a whole operating system, love that the rust core keeps it to like 5mb

Love the minimal take, especially the Rust core keeping things snappy. One thing that would make it stickier for me: a quick "snooze until later today" or "remind me tomorrow" option right in the inbox. Most of my email backlog isn't urgent, it just needs to come back at the right moment, and that small loop would make Skim feel genuinely complete without adding much weight.

love how lean this is, the rust footprint is wild. one thing that would sell me even more is proper vim keybindings for the inbox and reading panes, j/k navigation feels like a natural fit for something this minimal.

Honestly the local-first angle is super appealing, and a tiny Rust installer sounds great. One thing that would seal the deal for me though, kind of a small ask, is per-account send aliases so I can manage work and personal replies without juggling separate logins. Would fit nicely with the minimal vibe too.

finally an email client that respects my storage and my sanity, the 5mb install is unreal on windows. love that i can just plug in my own api key and skip the data-sucking defaults.

love that this is BYOK instead of yet another subscription tier, that's rare for an email client. since the AI can answer questions across the whole mailbox, does it send full email bodies to the API per query or is there some local pre-filtering first so you're not shipping your entire inbox history every time you ask something?

Do you have plans for Mac or Linux, or is Windows the main focus? The ~5MB installer and sub-second startup would be a pretty strong selling point for anyone tired of bloated Electron apps, so I'm wondering if expanding makes sense or if you want to keep it tight and focused.

Hey Nikita, congrats!
Any plans to release a Mac version?

The Rust-backed cold start is genuinely instant, feels closer to opening a text editor than an email app. Glad the AI part is opt-in with your own key instead of some bundled subscription, keeps things lean.

Curious how you handle Gmail OAuth secrets in a public repo though! Great launch, refreshing branding, and love every MIT license.

Hey Product Hunt 👋

We are the team behind Bolna, and we are here to make the whole process of building a Voice AI agent a whole lot easier and quicker than ever before.

Let’s be honest! The starting point to build a Voice AI agent has always been intimidating. Hours of effort gathering context, foolproof guardrails, detailed conversation flows, and yet something sifts through the whole orchestration, leading the agent to embarrassingly fail at deployment.

Agent Studio on Bolna takes all the above hindrances away between you and your perfect Voice AI agent by specifically putting every piece of the matrix together with one drag-and-drop brief. All you have to do is upload a contextual document and let it fill in the blanks to deliver a production-grade voice agent which can go live the same day as you sign up.

What makes Agent Studio handy for Voice AI agents:

• No prompt engineering expertise to get started.

• A modular structure in place including Identity, Conversation and Closing blocks.

• Built on thousands of proven templates applied across sectors.

• Handles all edge cases, instead of just linear scripts.

• Auto-picks the best combination of models for each use case.

• Warm endings and fallbacks for smooth user experiences.

• Reviews every agent across quality benchmarks and fills in whatever is missing before shipping the agent.

This changes the way Voice AI agents were fundamentally built, taking weeks to curate the perfect prompt and deployment. Agent Studio makes your perfect Voice AI agent come to life with a single session, delivering a production-grade deployment that holds up through real calls for enterprise scale. The barrier to building and scaling with voice has now dropped to almost nothing.

Presenting this to the Product Hunt community is huge for us, with weeks of deliberation gone into each step of the process. We’d love for you to give Agent Studio a shot to build Voice AI agents. Go all out with your use cases on Agent Studio, and we are sure you’d adopt it as your core Voice AI stack! And if you end up loving it as much as our beta users have, we’d love your support here!

Every upvote and comment means a lot to the team.

So what's your excuse for not building a Voice AI Agent anymore?

Cheers,

Sonam from Bolna

P.S. We'll be around to answer questions throughout the day, so shoot all your questions right away!

Hey Product Hunt 👋

For the longest time, building a Voice AI agent at Bolna meant one of us sitting down and building it by hand.

I have spent tens of hours building agents, any my team, who is building agents for some of India's largest enterprises, has probably spent 100x that.

Every new agent meant hours of gathering context, writing guardrails, mapping conversation flows, and second-guessing every edge case. We "Did Things that Didn't Scale" until we just physically couldn't. There was too much of a gap between the quality and speed of agent building required and what we could provide (even as our FDE team grew from 2 to 20 in 3 months)

Then somewhere past 200K+ hours of calls, a pattern emerged. The same building blocks. The same edge cases. The same fixes we kept making by hand, over and over.

Agent Studio is us finally putting all of that back into the product itself.

You upload a context document. Drag and drop a brief. And it -

1. Fills in the blanks using thousands of patterns proven across sectors

2. Structures the agent into Identity, Conversation and Closing blocks

3. Auto-picks the best combination of models for your use case

4. Handles edge cases, warm endings and fallbacks, not just linear scripts

5. Reviews itself against quality benchmarks and fills whatever's missing before shipping

Sign up in the morning, have a production-grade agent taking real calls by evening. No more prompting frustratiosn.

The part I keep coming back to - the hardest thing about Bolna's first two years was the manual work. Agent Studio is us handing that exact superpower to anyone who signs up.

The barrier to building with Voice has dropped to almost nothing.

We'd love for you to break it, push it, throw your weirdest use case at it. Give us your thoughts!

Super proud of @xan_ps and the entire team for building and shipping this!

So - what's stopping you from shipping a Voice AI agent today?

Back on PH after a long time and there's a reason! While building out voice agents I was always looking for ways I could vibe-create an agent and just talk to it. While Claude solved for building apps for custom use cases by just explaining in English, I found an equivalent for the Voice AI sector here!

Loving it through and through! But have a few suggestions. How can I share those, team?

insane !!!

it made life easier to write agents on the platform itself , now i can easily make agents for different usecases from different industries with Agent Studio within just 5 mins of uploading few documents and context

Finally got around to testing Bolna for an outbound campaign and was honestly surprised how natural the voice sounded in Hindi and Tamil. Setup took maybe an afternoon.

Finally moving away from long iterations because chatgpt just doesn't understand how my voice agent is supposed to work! Glad to know it is trained on real data.

Honestly the multilingual support surprised me most, our test calls in Hindi and Spanish both sounded pretty natural compared to other voice AI tools I've tried. Setup was straightforward too.

The async setup was surprisingly painless for handling multilingual flows, and the latency on outbound calls felt close to real human pacing.

the "handles all edge cases, not just linear scripts" line is the part I'd want to poke at before pointing real call volume at it. since Studio can get you live the same day you sign up, is there a staging/sandbox step where you can run a batch of adversarial or unusual calls against the generated agent before it starts taking live traffic, or is the expectation that you go live and iterate on real calls from day one?

Maitreya, the 200K hours of calls turning into reusable blocks is the most convincing part of this story: that is the kind of pattern library a hand-built agent never benefits from.

My question comes from my corner of the market. I build AI for healthcare, and voice is the obvious next interface for appointment reminders and patient intake, but every call recording there contains protected health information.

Do you support healthcare use cases today: things like a BAA, redaction of recordings, or region-locked storage? That answer decides whether teams like mine can even prototype on Bolna.

Looks super easy to use! Congratulations on the launch.

Everyone ends up building similar-sounding voice agents if they're using the same templates. How do you keep each one feeling unique?

The auto-assemble-from-a-doc part is the real leap here, and it's also where I'd look hardest. When we generated agent flows from source material, the thing that never came for free was the negative space: a caller asks something the doc just doesn't cover, and the model improvises a confident policy answer instead of saying 'I don't know, let me transfer.' Does Studio insert those refusal and escalation boundaries automatically, or is that still the part you tune by hand after the agent is generated?

A real-time analytics dashboard showing live sentiment and call drop-off rates during campaigns would help teams tweak prompts on the fly. Right now we mostly wait until after to see what worked.

Spent a few minutes poking around and the multilingual demos actually handled code-switching better than I expected. Pricing transparency was a nice surprise too, most voice AI vendors make you book a call just to see numbers.

Would love to see a built-in call quality analytics dashboard that breaks down latency, transcription accuracy, and drop-off points per language. Right now we're piecing it together from logs and it makes it hard to spot where the model is struggling in Hindi versus English versus Tamil. A single view that surfaces the metrics that matter most would save our team hours every week and help us fine-tune without guesswork.

A dashboard view showing live call metrics with filters by region, language, and intent would be super helpful when running thousands of concurrent calls. Right now I'm guessing it's hard to spot issues at scale without granular real-time visibility into performance breakdowns.

Would love to see a built-in call quality analytics dashboard showing latency, transcription accuracy, and drop-off rates per language. Right now it's hard to benchmark performance across different regional deployments without piecing together data from multiple sources.

Most of the thread is on edge cases and compliance. The linguistic complexity claim is the one I'd poke at, we run voice on the support side in a lot of languages.

On real calls people don't stay in one language. They switch mid sentence, and the worst place is digits. Someone speaking English will still say an order number or a postcode in their own language, almost every time. Same when they get annoyed, then a whole sentence flips.

If language is picked at call start and pinned there, that is where transcription falls apart, and it falls apart on the field you least want wrong.

So does the agent detect the switch inside the call and follow it, or is it locked once the call starts? And are digits handled separately from the rest..

The requires field encoding dependencies between elements is the interesting part — "select a plan before this button is clickable" is exactly the context accessibility trees drop, and it's where most browser agents burn their retries. Question: how do you handle drift? If the site ships a redesign, does the manifest regenerate per-call (so it's always fresh but you pay latency), or is there caching with some invalidation heuristic? That trade-off decides whether I'd trust it in a scheduled job.

@max_nordstrom that's a genuinely honest answer, most tools would just quietly claim it handles all cases. the disabled/aria-disabled inference makes sense as the pragmatic v1 scope. for the async-JS-validation gap, would something like a lightweight 'confidence' flag on the requires field be feasible - even a rough signal telling the agent 'this dependency was inferred from static DOM, verify before relying on it' vs 'this one's a hard disabled attribute' - or would that just push the complexity onto the caller without really helping?

i spend a lot of time watching agents guess their way around web pages and it is a mess, so this scratches a real itch. the requires field encoding dependencies is the clever bit. who publishes these first though, site owners or agent builders running it on other people's pages? feels like the chicken and egg question

the requires field for element dependencies is genuinely useful, something I've hacked together badly on past projects. excited to try it with an MCP server setup

The dependency tracking between elements is genuinely useful, saved me from manually wiring up plan selection logic in my agent. Curious how it handles shadow DOM and dynamic SPAs though, that's usually where these tools fall apart for me.

Finally a tool that captures the "why" behind clickable elements, not just the selectors. The requires dependency mapping is genuinely useful and saved me a ton of trial and error on a flaky login flow.

honestly the requires field is the part that sold me, most selectors break for me when modals depend on prior actions. gave it a quick spin on a messy form and it nailed the dependency mapping first try.

the requires field for encoding element dependencies is genuinely clever, that's the kind of context aria snapshots miss and usually breaks agents. gonna test it on a gnarly multi-step checkout flow

The action graph idea is really clever, especially encoding those prerequisite relationships between elements. One thing that would help me a lot is a built-in diff or change-detection endpoint, so when a site updates its structure I can quickly see which locators broke and get suggestions for replacements instead of manually hunting through the JSON map. Would save a ton of debugging time in production.

Really cool approach, the requires field for dependencies is a smart solve. One thing that would make this even more useful for us is a built-in way to record and replay the exact sequence of agent actions against a stored map, so we can diff when a site changes and pinpoint exactly which locators broke. Would save a ton of debugging when a flow stops working after a deploy.

the requires field is honestly the standout for me, it captures context i've been missing when agents just blow past dependencies and fail silently. one api call returning resolved locators plus that semantic info is way cleaner than juggling aria snapshots.

The MCP server integration is a nice touch for getting started quickly. One thing that would really help in production is some kind of caching layer for repeated calls to the same URL, since right now it seems like every API request would re-scan the whole page. Even just an optional ETag-style header so we can skip re-mapping when the page structure hasn't changed would save a ton of tokens and latency when agents are looping through tasks on stable sites.

Finally something that gets rid of the brittle XPath mess my scraper has been held together with. The requires field is a nice touch, saved me from another broken dropdown flow.

Really useful for browser agent work, the requires field alone is a nice touch. One thing I'd love is a built-in way to test the generated map against an actual page session, so you can catch mismatches between the JSON output and what the DOM actually looks like before your agent runs.

finally something that gets at the real pain point. the requires field encoding dependencies is genuinely clever, going to wire this into our langchain setup this week and see if it holds up on a few gnarly sites.

the requires field is genuinely useful, been burned by agents clicking submit before the form was complete way too many times. one thing id love to see is a way to mark elements that trigger async loads, like a dropdown that fetches options after you click it. right now my agents try to interact before the data is there

honestly the requires field sounds super useful, would love to see some kind of caching layer for repeat calls on the same URL so agents aren't paying the extraction cost every single time

Tried it on a gnarly multi-step checkout flow and the dependency graph actually caught a hidden "select plan before continue" rule I had been missing. The resolved CSS plus role locators side-by-side saved me a ton of selector soup.

The dependency graph is a really smart touch since most selector tools miss those relationships entirely. One thing I'd love to see is a built-in diff endpoint that compares two Manifest snapshots over time, so when a site silently changes its checkout flow the agent can detect the break before it fails mid-task.

finally something that gets the dependency thing right, you know? the requires field saving me from writing all that state-checking logic is honestly a huge time sink off my plate

The requires field for encoding element dependencies is genuinely clever. It's the kind of context that usually takes teams weeks of brittle DOM mapping to figure out, baked right into the response.

the requires field is honestly such a nice touch, most tools i've tried just hand you a flat list and you have to figure out ordering yourself. this feels way more usable out of the box.

Having a built-in way to record and replay a full user flow during integration testing would be huge. Something like a "golden path" capture from one successful Manifest call so my agent team can replay it on staging and catch drift before it bites production.

Love that the requires field captures element dependencies, that's a real pain point I keep hitting. One thing that would save me a ton of time: a built-in wait helper that knows to poll until those required dependencies are satisfied before resolving the click action. Right now I still end up writing retry logic around it.

the requires field is such a smart touch, encoding dependencies directly into the map saves so much back-and-forth between agent and page. clean SDK and MCP server setup too.

honestly this looks super useful for shipping agents, the requires field alone solves a headache ive dealt with for months. one thing id love to see is some kind of caching or snapshot diffing so you dont have to re-scan a full page every time the agent loops, basically only re-resolve the parts that actually changed.

the requires field is the clever part here, most of these tools stop at "here's a button" and skip the ordering constraints entirely. how do you actually detect those dependencies though - is it driving the page and observing which elements become clickable after which actions, or some static analysis of the DOM/JS? asking because sites that gate steps behind async validation (like a plan check that hits an API before enabling the button) seem like they'd be hard to catch without actually executing the flow.

The requires field is a really nice touch, caught me off guard how often other tools just hand you a list of clickable things without telling you what's gated. Quick to wire up with the Python SDK too.

Congrats on the launch. The action-manifest idea is useful because browser agents usually fail on hidden state, auth changes, or brittle selectors. How do you encode uncertainty or required preconditions when a page changes after the manifest is generated, and can developers inspect why an action was skipped?

Hey Product Hunt! 👋

Topolines turns any real place on Earth into a clean topographic contour map. You pick a location on the map, it pulls the actual elevation data for that spot (SRTM 30m, or 10m GeoTIFF for finer terrain), and generates the contour lines from it. So the ridge you export is a real ridge, not a shape you drew.

I built it to solve a recurring friction in my own design work: getting vector topographic maps without opening heavy 3D software or rebuilding them by hand in a vector tool.

From there you can tweak contour density, line weights, the elevation taper and colors, then export as clean SVG or HD PNG, ready for Figma, web, or print.

Try it on somewhere you know well, that's when it clicks. I'm around all day for questions and feature requests 🙌

Seeing a few comments about drawing or sketching zones, so a quick clarification for anyone landing here: there's no freehand drawing in TopoLines. You pick a real location on a map and the contours are generated from actual elevation data (SRTM 30m, or 10m GeoTIFF on paid). The terrain you export is a real place, not a shape you drew. Happy to answer anything else.

I am completely obsessed with this app. I've been hiking in the LA area for 3 years now with a really weak angle. I'm able to assess depth over distance swiftly and decide if I can make it down or up super steep mountainsides. Both me and my ankle thank you 🫡

The contour generation actually looks smooth and respects the drawn zone boundaries properly, not just a sloppy clip mask. That gradient-to-line opacity taper is the kind of detail that makes it feel made for real cartography work, not just a tech demo.

using real SRTM elevation data instead of just letting you doodle contour-looking shapes is what makes this actually useful for real design work, not just a pretty toy. how fine does the 10m GeoTIFF option get in practice - is that resolution available everywhere on Earth, or only for certain regions where that survey data exists?

the gradient control on the elevation taper is genuinely thoughtful, gives you that subtle handmade cartography feel instead of looking like a flat computer readout. love how clean the SVG export is too.

Adding a quick elevation labeling option directly on the contour lines would save a ton of time, especially for detailed terrain work where you need to identify specific altitudes at a glance without opening a separate legend.

It would be great if you could add an option to import an existing SVG or image so we can trace over real terrain data instead of having to draw zones from scratch every time.

Love how clean the export looks and the Figma output is a huge plus for my workflow. One thing that would save me a ton of time is letting me save custom gradient presets tied to elevation ranges, so I can quickly reuse the same color scheme across multiple terrain projects without re-tuning each one.

Love how the elevation taper keeps the contours feeling organic instead of mechanical. The Figma-ready SVG output is a really thoughtful detail for designers.

Honestly the gradient and elevation taper feature is super useful, way better than other tools I've tried. Exported SVG dropped cleanly into Figma without any cleanup.

Drew a quick test zone and the contour output came back surprisingly clean, with smooth gradients and crisp lines ready to drop straight into Figma. The SVG export saved me a ton of cleanup time.

Would love a way to import a simple list of elevation points or a CSV so I don't have to draw zones manually when I already have survey data. That would speed up my workflow a ton.

The contour generation is impressively clean for such a quick draw, and the elevation taper really makes the depth read well. Exporting straight to a print-ready SVG saved me a ton of cleanup.

The elevation taper setting is such a thoughtful touch, it keeps the contours from feeling flat at the edges. Clean Figma export is the cherry on top.

Love that the elevation taper actually shows up in the exported SVG, makes the contour map feel tactile rather than just flat geometry.

Drawing a freehand zone and watching the contours snap out so cleanly was a nice surprise. The HD PNG export is already perfect for dropping straight into a deck.

Honestly kind of surprised how clean the contours came out, even when I scribbled a messy zone. The Figma-ready export saved me a bunch of fiddling.

The zone drawing UX is genuinely satisfying, feels closer to sketching than configuring software. Love that the contour output drops straight into Figma without a cleanup pass.

The SVG export is genuinely crisp and the elevation taper control feels like something a cartographer would appreciate, not just a design toy.

Drew a quick mountain ridge and the contour output looked surprisingly clean for a browser tool. The Figma-ready SVG was the real treat, dropped straight into my layout without any cleanup.

Drew a quick zone on a hillside photo and the contours came out way cleaner than I expected, especially with the elevation taper. Exporting straight into Figma saved me a solid hour of fiddling.

Drew a quick zone around a coastal area and the contour lines came out clean and print-ready without any fiddling. Love that the SVG drops straight into Figma without weird cleanup.

Honestly the contour lines come out way cleaner than I expected for something this quick. Drawing a zone and getting print-ready SVG basically instantly feels like cheating, you know

Smooth workflow from sketching a zone to getting clean contour lines, and the export quality actually holds up when I dropped the SVG into Figma. The elevation taper feature is a nice touch I didn't know I needed.

The zone-based drawing flow feels really intuitive, and the elevation taper on the contour output looks beautifully clean. Love that it spits out print-ready SVG without any fiddling.

Would love a way to import a simple point list with elevation values so I don't have to draw zones by hand every time, basically just paste lat/long/alt and get contours out

The contour smoothing looks genuinely sharp on the preview, and the gradient-to-taper blend feels really considered. Love that the SVG output drops straight into Figma without weird cleanup work.

The contour lines came out surprisingly crisp on the first try, even on a quick freehand zone. Love that the SVG dropped straight into Figma without me having to clean anything up.

the fact that the contour taper responds to your drawn gradient rather than just snapping to preset levels is a really thoughtful touch. the svg export stays clean too, no weird node cleanup needed when I dropped it into figma.

Love the open-source approach and the multi-agent timeline idea. One thing that would help adoption a lot is adding a simple share/export-to-web preset that bakes the project into a viewable link, so non-technical folks can preview cuts without installing anything. That alone would make sharing work-in-progress much easier.

Local-first editing with a real multitrack timeline is a really thoughtful call here, makes the whole thing feel like an actual editor instead of a toy. Appreciate the AGPL stance too.

finally an editor that doesn't fight me on the timeline, and the local-first angle means i can actually trust it with rough cuts. genuinely impressed it stays free under AGPL.

finally a chat driven editor that respects the actual timeline instead of pretending it does, love that it stays local first

Love that the timeline stays multitrack and editable on export, not flattened into a render. Hooking Codex and Claude into real video editing through MCP feels like the first time agents can touch a timeline without breaking it.

been wanting a real open source editor like this for a while, the local first angle is huge honestly. one thing that would make me actually switch though is proper proxy generation for 4k footage, my laptop struggles with raw timeline scrubbing and most editors punt on this

the timeline preview when agents are working on it would be a game changer, like being able to see the rough cuts they're proposing in real time without having to open each export. would make the back and forth so much faster when you're collaborating with the agents

finally something that doesn't lock my edits behind a paywall or a cloud upload, the local-first setup with claude actually tweaking the timeline felt surprisingly natural and the agpl license is a huge plus.

just curious whats your relations with chatcut lol

I like that the agent edits a real, reversible timeline instead of handing back a flattened export. Keeping captions, music, transitions, and keyframes visible makes this feel much easier to trust.

The 'edits proposed for review before applied' loop on a real timeline is what makes this feel usable rather than a black box. When an external agent works over MCP (say Claude Code), does it get the full timeline state as context each turn or just a diff of the current edit, and where do those proposed-but-unaccepted edits live — in the project file on disk I could version in git, or app-internal state until I accept? I'm trying to gauge whether the whole editing history stays reproducible outside the app.

The real-timeline part is what sells me — chat-only editors feel like a black box, but coming from a DAW world, being able to drop into a multitrack timeline and hand-nudge things after the agent's pass is exactly the workflow I trust. Quick one: how deep does the audio side go — multitrack audio with volume keyframes/ducking, or mostly video-first for now? Local-first + AGPL is a great call btw.

i cut short product demos for social every week and the timeline is always where ai tools fall apart. they generate clips but you can't nudge anything after. agents on a real multitrack timeline over mcp is the missing piece. can claude code drive it today or do i need your chat ui?

Love the local-first approach and the real multitrack timeline finally being usable with AI agents. One thing that would really help me as an editor: a way to version the timeline edits the agents make, so I can diff or roll back specific agent actions when something goes sideways.

The "real multitrack timeline" part is what makes this stand out to me — most AI video tools spit out a black-box render you can't touch, so having the agent edit an actual editable timeline (and local-first) is a much saner workflow. Love that it's AGPL too.

Quick question: since it works with Claude/Codex/MCP agents — do I bring my own API key/model, or is there a local-model option for people who want it fully offline?

The part I'd poke at is how the agent addresses clips through MCP. Every time we've handed an agent a stateful document as tools, state drift is the thing that bites: it reasons over a snapshot it read a couple calls back, so 'trim the third clip to 5s' hits the wrong clip once the timeline moved underneath it. Are the edits keyed to stable clip IDs it can re-reference, or to positions and timecodes that shift when tracks change?

Local-first multitrack editing with agent support sounds genuinely useful. One thing that would make a huge difference for me is a simple undo history that spans across agent-driven edits, so I can roll back a whole AI cleanup pass without losing my manual tweaks from before.

Love the local-first approach and editable exports being free. One thing that would help a lot is proxy media generation so 4K timelines stay snappy on lower end machines. Would also be great if the AI agents could remember project wide style choices between sessions so you don't have to re explain your color grade each time.

honestly the multitrack timeline with MCP agents sounds super cool, but one thing i'd really want is a way to save and share edit "recipes" or agent prompts so i can reuse the same workflow across projects. basically a little library for your favorite prompt chains that you can drop in whenever.

finally a chat based editor that doesn't lock my timeline away, ran a quick cut with claude and the export actually stayed editable which is more than i expected from something free.

A collab mode would be huge, like letting two people hop on the same timeline remotely with comments pinned to specific clips. Since the agents already touch the project files, feels like a natural next step for a local-first tool.

Would love to see a simple preset system for common edits like "remove filler words" or "add jump cuts" so I don't have to prompt the agents for the same workflow every time. Could be a small dropdown that bundles the right tool calls together.

The local-first approach with real multitrack timeline support is such an underrated detail, makes it feel like an actual editor instead of a chat toy.

Love the local-first angle and the multitrack timeline with real agent integration. One thing that would make a huge difference for me: a built-in proxy/media cache so 4K footage doesn't bog down the timeline when the agents are running analysis in the background. Would make the whole editing flow way smoother on mid-range hardware.

Love the local-first approach and editable exports being free. One thing that would really help adoption is a built-in shareable project link feature where collaborators can open the same timeline in a read-only browser view without installing anything, perfect for getting quick feedback from clients or teammates before they commit to the setup.

Love the local-first approach and the multitrack timeline is exactly what I was missing. One thing that would really help: add a visual diff or changelog view when the AI agents make edits, so I can quickly review what Codex or Claude changed before approving. Right now trusting invisible agents on a timeline feels a bit opaque.

finally something that doesnt lock your timeline behind a subscription, ran a quick cut with claude and it actually respected my existing tracks instead of nuking them

Hey Product Hunt! 👋 MD here, founder of DualStream, with my co-founder John (Tangent on Twitch where he's been Partner for over 13 years). Quick story on how this got built, because the origin actually matters:

I have spent 20+ years building digital products. For a big chunk of that time I've watch streamers as a sort of background environment while I code or design, I love chat and being a lurker. For most of that time, I watched streamers I know describe their setups in horror: plugin duct tape, multi-streaming nightmares, software so complex they'd lose entire weekends to it. The same stories, year after year.

A while back I reached out to Tangent - I'd been a chatter of his since the days of No Man's Sky drops - and asked if he'd be willing to test out some prototypes I was playing with about a better way to multi-stream. He said yes, and then walked me through everything streamers actually deal with. The list was long. We started talking to other streamers. The list got longer. Their wish lists were specific, repeated, and almost never met by the tools on the market.

Those conversations became DualStream.

A note before the pitch: OBS is one of the greatest open-source projects of the last 20 years and the foundation that taught most of the streaming world how this all works, we have nothing but love and respect for OBS and the dedicated team behind it. DualStream isn't here to replace it. We're here to be the next thing - for creators who want a different path. Every creative field got its intuitive layer (Squarespace, Canva, CapCut). Streaming was the last one waiting.

DualStream is one studio that streams to your Desktop and Mobile audiences at the same time - same sources, arranged for each scene, live everywhere in one click. The DualStream Relay - one upload from your PC, live on every platform from our servers. And if your connection drops mid-broadcast, Relay keeps every platform live on your standby screen and picks back up the moment you reconnect. Native Alerts, Per-Source audio control, VTuber support, instant replay clips (horizontal and vertical), cinema grade effects, pop-out multi-platform chat and so much more is all built in.

The app is free to download - build everything, test offline, go live to one channel. Premium is $9.99/month for access to the DualStream Relay, multi-format recording, water-mark free clips and streaming to more then one channel at a time.

🎁 For Product Hunt: use code PH-3M-HI6NIX at checkout and your first 3 months of Premium are free (or just go to dualstream.gg/ph, its applied automatically that way).

We're a three-person team shipping every two weeks, and we genuinely build what users ask for. We'll all be here all day in the comments. Ask us anything, about the product, the build, the streaming category, the design process, the streamer feedback that shaped it, whatever. Genuinely grateful for your time!

// MD & Tangent

Oh! Fun little P.S. ~ We also just released our first Open Source project for DualStream: An MCP server! This server lets AI assistants like Claude operate your DualStream studio: switch scenes, save instant-replay clips, build entire scene layouts from scratch, restyle your alerts and widgets, and react to what happens on stream. Everything runs locally on your machine, with your explicit approval, and every change renders on both your desktop and mobile canvases. Check it out here.

Greetings Product Hunt! I am John, co-founder of DualStream! I am a full time live streamer on Twitch with 13 years of Partnered, full time experience on the platform with 230,000 followers under the online handle 'Tangent'. I started streaming as a passion hobby once I discovered such things existed and I was instantly hooked. Over a decade later I am still here and have experienced every up and down a creator can imagine!

In that time the content creation space has evolved and changed (And grown exponentially!) in ways many creators might not expect! One place that has never changed though? The reliance on open source OBS for operating what 99% of the streaming world utilizes.

My contributions to DualStream are in the features I feel that content creators of all sizes not just want, but absolutely need in todays ruthlessly competitive streaming space. MD can handle the technical side of things, but if you are a content creator; a dreamer just starting out with a hobby or an experienced content creator hearing a streamer's perspective is, read on my friends!

OBS is responsible for giving me and many of my peers the careers that we have had and enabled us to turn hobbies into careers. OBS was never intended for live streaming, but with a little creativity and understanding of its (confusing) settings, we have been able to fit it to what we need.

But that is one of the big challenges of OBS! It is robust in the hands of somebody who is technically experienced with the software, but to a new user it is a daunting trial and error experiment. Multi streaming has been the recent hot topic people have been trying to crack in the last few years. If you want to branch your stream from Twitch to Youtube, Tiktok, Kick, etc AND have both a vertical feed for phones and a horizontal for PC users, you have to jump through many hoops to make this possible.

OBS is capable of doing these things, but requires the use and understanding of plugins and mods, all of which are designed by good (usually, there have been security issues) intentioned users. The specific installation instructions and settings required to use many of these plugins results in frustrations and radical changes to the OBS UI. On top of that, it bloats the OBS software and makes it perform in unexpected ways depending on your PC setup. You shouldn't have to have custom docks installed and complicated modifications to your streaming software to branch your stream out in 2026.

When Twitch removed their no compete clause on multi streaming, the wild west once more opened up to creators. The push to crack the easy multi streaming code was the hot topic. For OBS, this meant bigger bloat, more plugins, worse PC performance, many crashes, and a lot of experimentation. For a new streamer you can afford to make mistakes and adapt to the multitude of glitches and crashes if you mess it up, but if you are a full time caster where your mortgage payments are reliant on you keeping your content schedule steady and professional, this can be harrowing!

Remember where I mentioned OBS has 99% of the market for users? Well, all of the innovations for multi streaming (And everything else) began flowing through OBS. The industry decided that the best way to cash in on this, was to bloat OBS even more, and design entire companies and software around utilizing the already strained OBS systems simply because it was so prevalent in the space. Virtually nobody has innovated streaming software from the ground up, too concerned about leaving the security of the confusing, bloated, and increasingly expanding software just because the assumption was that everything must be tied to OBS. The companies that did try to make their own software charge ridiculously high prices for essentially what OBS does with a somewhat more simple approach.

This is where we come in! MD has been a long time viewer of mine in my twitch channel, and witnessed me having some extreme problems with OBS. Crashes, disconnects, encoding problems, multi stream failures. At one point Twitch kept closing my connection to OBS mid cast due to a problem with OBS that would require them to patch it out, leaving me helpless for months.

He approached me and offered to build me a private connection to Twitch, my own personal streaming software because he was very kind and wanted to help me. He built me a very simple, but extremely effective piece of software that proved to be incredibly stable. He asked me if there was any features that I would like to have on it, and this turned into multi hour brainstorming sessions about the industry and tools that I need for my content.

This would eventually become DualStream. I was so floored by MD's ability to implement my requests (And I have a lot of them!) so quickly and so simply that the discussion started to steer in the direction of "If this is helping me so much, it has to be able to help others too."

In no time we had a complex streaming software that was so SIMPLE to use that required ZERO plugins or modifications and sidesteps virtually all of the pain and agony associated with trying to multi stream in multiple formats. I have been using DualStream exclusively for thousands of hours and proudly use it every day for my live streams. I entrust my financial wellbeing as a full time creator in DualStream to put my money where my mouth is. Without wanting to sound cocky, I have had more financial success in the last year of utilizing DualStream on my casts than previous years with OBS, even with the bumps in the road of developing the software and live fire testing it on my live streams.

But it isn't just the raw ability to reliably connect. It is all the fun features for interacting with audiences, manipulating your feed with special effects and on screen novelties that normally require browser sources and many plugins to use. We even have a fully customizable on screen Alerts that offer more customization that Streamlabs and Streamelements without the need of coding knowledge (And it is less taxing than using browser sources.. Though we have browser sources available as well of course!). A lot of streamers don't realize that browser sources are TAXING on your system and the reason every streaming software utilizes them is because the streaming world is chained to OBS and everyone is too scares to take the chance on innovating and those that do see even small creators as cash cows and will charge very high prices knowing that there are no better options.

So that's it! DualStream is a from the ground up approach to live streaming, designed by creators aimed specifically for creators, not merely by people trying to cash in on creators. I have not abandoned streaming to focus on DualStream, I am actively streaming 60+ hours per week and feel that if I am to properly guide DualStream in its future creator features that I will need to be on the front lines of content creation still.

If you have any questions about live streaming, the successes and failures (and risks!), or anything in the creation space as well as DualStream, please feel free to ask! Thank you for your time! - John / Tangent

twitch channel for verification: https://www.twitch.tv/tangent

Simultaneous desktop + mobile streaming is a great use case for hybrid content creators. Does it sync audio/video timing automatically, or is manual alignment still needed?

honestly this looks really solid, the simultaneous desktop and mobile push is a great angle. one thing i'd love to see is built in collaborative moderation for chats across all linked platforms, basically a shared mod queue so my team can handle trolls without jumping between five tabs. would make running a bigger stream way less chaotic

How does the disconnection protection actually work? Does it buffer on your servers or keep something local? That's genuinely useful if it can hold a stream alive for even 30 seconds, but I'm curious what happens if someone loses internet completely.

finally something that handles desktop and mobile at the same time without me juggling two setups. honestly the relay feature saved me when my wifi dropped mid stream

Something different for a change, everyone else seem to be trying to clone each other :0)

Good luck with the launch you guys

the fact that everything from chat overlays to VTuber support lives natively in one window instead of stacking extra apps is honestly kind of a game changer, you can tell the team actually streams themselves

Pushed to twitch and tiktok at the same time with no extra setup, that alone sold me. The chat pop-out keeping everything in one view while streaming is genuinely nice.

The multi-platform chat pop-out is great, but it would be super helpful if there were a way to set per-platform chat moderation rules so I don't have to ban the same spammer three times across Twitch, YouTube, and Kick.

One thing that would make DualStream a no-brainer for me is built-in multitrack recording with sync markers. Being able to pull a local high-quality archive after every stream and drop the markers straight into Premiere or DaVinci would save a ton of post-production time.

Finally a tool that doesn't make me choose between desktop and mobile viewers. The relay server thing is basically magic - my stream stayed up when my internet died mid-game, which honestly never happens with other setups.

The dual desktop and mobile stream actually looked crisp on both without me tinkering with bitrates, which never happens for me. Also appreciate the relay kicking in when my wifi stuttered mid broadcast.

Tried it out for a quick stream and honestly the dual desktop and mobile push just worked, no fiddling. The relay keeping the stream alive when my wifi hiccupped was kind of a lifesaver.

The way the per-source audio control is baked right into the same panel as the stream health indicator is a really thoughtful touch. It feels like the team actually streamed before building this.

Pushed a test stream to desktop and mobile at the same time and it just worked, no fiddling with OBS scenes or restream plugins. The disconnection protection saved me when my wifi hiccuped mid-broadcast, which honestly sold me on it alone.

The relay setup genuinely saved me when my internet flickered mid-stream, no awkward goodbye screens. Love that I can finally chat with both desktop and mobile viewers without juggling windows.

The cross-platform streaming idea is solid, especially with that disconnection protection built into Relay. One thing that would really level it up is a built-in collaborative editor where co-hosts can queue clips, soundboard cues, or scene changes in real time without needing a third-party tool like Streamdeck. It would make multi-person streams feel way smoother.

12 hours in — quick update: This thread has been better product feedback than most user interviews. The three most-requested things today — a turnkey GitHub Action for CI gating, first-class Slack/Teams drift alerts, and a side-by-side spec↔consumer diff — all went onto the roadmap today. Keep it coming. And my question from this morning still stands: what's the worst "the docs lied to me" moment you've had? 👇

the request/response drift coverage in this thread is thorough. one direction I didn't see covered: outbound webhooks. those break consumers just as often as REST responses do, but there's no request to inspect, just whatever payload you decide to send on an event. is webhook payload schema part of the same spec/drift pipeline, or is that a separate problem you're leaving to the consumer to validate on their end?

Finally tried this on a messy internal API setup and the drift detection caught two stale endpoints we totally missed. The MCP piece feels like the real unlock for our coding agents.

honestly the drift detection part is what pulled me in, super useful. one thing though, it would be amazing if you could add some kind of slack or teams alert when drift actually gets caught in prod, so the right person gets pinged right away instead of someone having to go check the dashboard. would save a lot of "wait when did this break" moments.

finally something that tackles the spec drift problem head on. hooked up our openapi and it immediately flagged two endpoints in our docs that were out of sync, which honestly saved me a headache before our next release.

love that the MCP angle means my agents actually pick up the same spec updates instead of working off last week's version — that detail alone saves a ton of "wait why did this break" moments

Finally tried routebase on a messy node project and it caught three routes that drifted between our swagger doc and tests, exactly the kind of thing that always sneaks past review. The MCP angle is genuinely useful too since our agents can pull the same source.

The drift-detection angle is genuinely useful, most teams I know are drowning in stale API docs that nobody trusts anymore. Love that it hooks into MCP so agents stay in sync too.

Drift detection sounds great, but it would help a lot if Routebase could show a side-by-side diff view between the published spec and each consumer like the docs or mocks so I can spot what changed at a glance instead of just getting a list of drift alerts.