PH热榜 | 2026-07-28

Hi Product Hunt - Matt here, co-founder of Prefactor with Simon.

We've been heads-down on this one for a while, so finally getting to show you is a real thrill.

Let me start with the question this whole thing is built around: do you actually know what your agents are doing in production right now?

For most teams we talk to, the honest answer is "…not really." Your evals pass, everything's green, you ship - and then every real run vanishes into a black box. Quality quietly drifts. Risk creeps in. Costs climb. And eventually someone asks which of your agents are still doing their job, and the room goes quiet.

That silence is the entire reason Prefactor exists. Gartner reckons 40%+ of agentic AI projects get scrapped by 2027 - and honestly, this gap is a big part of why.

So we built the thing we kept wishing we had: Prefactor scores every run in production the moment it happens - quality, drift, risk - then wires those scores straight into action. A failing agent gets caught live, not charted three days later.

How it actually works

1. prefactor init - one command connects your workspace and discovers your agents across your runtimes. First traced run in under 5 minutes.

2. Drop in the SDK (TypeScript or Python) - native for LangChain, Claude, Vercel AI, OpenClaw and LiveKit. Every call becomes a span, streaming in live with cost and data risk attached.

3. Run the evals you define on every run - LLM-as-judge, technical checks, qualitative metrics. Custom spans pull context from GitHub, Linear, Jira or your database, so every eval is grounded in what actually happened, not a guess.

4. Act. Hold, approve or block the second a run crosses a line - automatically at runtime, or routed to a human. Every decision logged and enforced through the SDK or API.

The payoff: you get to ship agents like real software — versioned, staged, promoted through dev → staging → prod only when evals pass, with instant rollback when they don't.

Why it's different

Most tools observe and score, then hand you the problem. Prefactor closes the loop - observe, evaluate, act, all inside the same run. A risky agent gets caught, not just charted.

My favourite bit of feedback so far: a customer with 40 agents in production and, in their words, "no honest way to say which ones were still doing their job." We gave them that answer - and the brake pedal for when one wasn't.

Who it's for

Engineering teams shipping agents to real customers, on any stack. Agent frameworks work natively; everything else plugs in through OpenTelemetry or the core SDK.

A little something for the PH community

Sign up today and you get 1,000,000 free agent steps. Every span counts, so that's a serious amount of live production evaluation on us. Valid until Friday 11:59pm PT, once you've set up your first agent.

Our ask

If you're running agents in production, tell us how you keep tabs on them today - even if the honest answer is "we're mostly hoping." We'll be in the comments all day and we'd genuinely love to hear what's working, what's breaking, and what you'd want a tool like this to do next.

Get started free at prefactor.tech. First 25,000 spans a month free, no card needed.

@simon_russell1 @joshgillies @ethan_lee8 @rheu @joeys and I will be here all day.

Big thanks to @rohanrecommends @rohanchaubey4 hunting us.

Genuinely curious what "scoring every run" looks like at scale, is that sampled or are you actually evaluating 100% of production traffic? That distinction matters a lot for cost and for trust in the numbers.

Drift is the one nobody talks about until it's already cost them a bad week. Real detection here feels like the right instinct.

Congrats Ethan and team!

The line "most agents pass their evals and fail in production" is going to resonate with more teams than you probably realize. I've said some version of that sentence out loud in at least three postmortems this year.

I spent months building internal scripts to catch exactly this kind of drift. Would've saved me a lot of late nights to just plug into something like this instead.

Hi everyone, Joseph here. I’m a designer at Prefactor.

Something I’m particularly interested in that’s easy to overlook: the moment Prefactor flags something mid-run, a human has to look at a screen and decide whether to step in or let it ride. That decision is only as good as the interface it happens on.

Agents generate an enormous amount of data, and most tools just show you all of it. My job is making sure that when something's going wrong, you can tell in seconds, not after scrolling through a wall of spans. Real-time control needs real-time legibility.

If you try Prefactor, I’d love to know: did you know where to look, or did you have to dig?

Hey guys Ethan here, I'm part of the team at Prefactor.

My main focus is on the go-to-market side which means I pretty much spend my week on calls with teams and engineers running agents in production.

I usually hear the same stuff all the time.

For example: The agent worked in testing and POC, it went live, but now nobody can answer a basic question: is it doing what we told it to?

That’s where we’d come in, Prefactor gives those teams visibility into what their agents are actually doing in production, so accountability doesn't stop the moment it's in production.

If you've got an agent running live right now, what does your monitoring actually look like? Keen to know if anyone here has actually solved this properly and if so how.

Cheers!

Hi folks, I'm Simon - CTO and other co-founder of Prefactor. The real challenge for companies now is not building agents, but managing the infrastructure and process around them. A lot of the lessons learnt from deploying traditional software at scale still apply, but agents pose new and unfamiliar challenges. It's a combination of traditional devops and HR. The approaches to risk and quality that have worked in the past need to be updated.

Prefactor makes this approachable for any engineering team. Closing the loop on turning feedback into improvements to the agent; ensuring consistency over model and prompt updates; simple ways to understand and contain risk; ways to monitor and control the actions of agents across frameworks and deployment environments.

Friends, Josh here. AI Engineer @ Prefactor. I built the agent instance tracking that scores quality and flags data risk while your agent runs, plus the SDKs and docs that get teams from zero to live without guessing.

Here's the thing I keep coming back to. You can't have confidence in something you can't see. An agent in production is a bucking bronco; it'll throw you the moment you stop paying attention. I've talked to too many teams who deployed, watched it work for a week, then realised they had no idea what it was actually doing. No visibility, no guardrails, just hope.

That's the piece Prefactor owns: the quality and data risk scoring that runs in-process, flagging misbehaviour mid-gallop rather than in a trace you read after the damage is done.

If you've got agents live right now, how do you actually know they're behaving? Or are you just holding on and hoping?

How do you define "quality" here? That word does a lot of work and I imagine it means something different for a support agent than a coding agent.

How does Prefactor decide what parts of the code need attention without creating unnecessary changes? Congrats @ethan_lee8 & team!

Hey Product Hunt, I'm Mukil — an engineer at Prefactor. 🔥

People always ask me:

"Mukil, you work with AI all day, isn't it crazy what it can do now?"

Sure, it's pretty cool. But I spend most of my time on the mostly ambiguous half of that sentence. Yes, it did something but was it the right thing? Did it actually do it or did it just say it did? Did it just hand someone the account details for the completely wrong account?

When an agent screws up, it often doesn't like to stop. It keeps going, fully confident and you find out hours or days later. And at that point its often when a customer's already upset or a refund went out that absolutely should not have. As an engineer this made me a little annoyed. The problem was never that I couldn’t see what the agent did. Dashboards will happily show me — right after it’s already done something I didn't want it to do. It's already leaked that PII, emailed the wrong guy or somehow spun up 15 subagents to run a database query.

With Prefactor you can see runs happening live, stopping the agent in its tracks. Every run gets scored live and lets the human know to step in if they need to. Instead of waiting 30 minutes to see this failed run on your custom dashboard (it's very pretty, I'm sure), you can step in and kill it while its executing.

In short, almost anyone can build an AI agent these days. And as the barrier to building them gets lower, the standard for running them well needs to get much higher.

We obsess over that second part, so you can spend less time investigating what your agent did at 3:14 a.m. and more time letting it do actual work.

Would love for you to try it and tell us your experience with it!

Ethan asked how people keep tabs on live agents, so here is my honest answer. I run scheduled agent jobs daily for SEO reporting and roughly one in three submits used to report success while the page never actually changed, so every job now ends by rereading the rendered result and comparing it against what the agent claimed. Mukil's question of did it actually do it or did it just say it did is the exact failure I kept measuring. Can an eval in Prefactor check an external side effect like that, the page actually updated or the row actually written, instead of scoring the run transcript?

Fantastic product and team, very excited to see them launch here, congratulations on the successful launch of the product!

IMO the 'act inside the same run' is what actually separates this from the dashboards that just score runs after the fact. Charting a bad run three days later never stopped anything. QQ - once an eval can hold or block at runtime, that eval is sitting in the critical path - how much latency does the inline check add before it lets a step through? congrats on the launch!

kind of worried this just becomes the new green checkmark people stop questioning. same problem, one layer up?

I really like the idea of tracking how the agent is working, as it is performing actions. Being more proactive is better than finding out something went wrong several days ago. It's not to be overlooked that the interruption isn't just killing the agent run, but handing it to a human too. I am going to have to play around with this and my own agents.

The runtime controls are the strongest part here. Being able to hold or block an agent when an eval fails turns evaluation into an actual safety layer, not another dashboard.

How do teams prevent a noisy or misconfigured eval from repeatedly stopping healthy production runs?

'no honest way to say which agents still do their job' nails it. when the llm-judge itself drifts, what keeps the eval honest?

Can I use Prefactor for my fleet of Hermes and OpenClaw agents ?

Working in high-trust environments, I don’t think the future is humans approving every AI action. It’s agents operating independently 95% of the time, with enough visibility and confidence that when they drift outside the guardrails, a human can intervene before it matters. Looks like you’re tackling that problem head on. Congrats @matt_doughty @simon_russell1 and team!

Send me my 1M spans to give it a solid rev up! ;)

Congrats on the launch.
When teams try Prefactor for the first time, how do teams catch the first bad run before users notice?

The "tokenless until you actually need an LLM" design is the part that stands out, most eval tools default to LLM-as-judge for everything and eat the cost whether it's warranted or not. Given you're scoring 100% of production traffic, curious how Prefactor handles agents that call other agents (sub-agent chains). Does a breach in a nested sub-agent bubble up and pause the parent run too, or does each agent in the chain get evaluated as its own isolated instance?

That's exactly the gap that matters. An agent can produce a perfect transcript and still fail the task in the real world. We've seen jobs report "success" while nothing actually changed. That's why the real evaluation isn't what the agent said—it's what actually happened. Can Prefactor validate external outcomes, like confirming a page was updated, a record was written, or an API state changed, instead of only grading the execution log?

scoring the run live instead of charting it three days later is the part that actually matters. catching a bad agent after it already shipped the damage is just reporting.

I like that this is built around production instead of just benchmarks. That's where things usually get interesting. Congrats on the launch!

I like that Perfector focuses on real production behavior instead of only passing evaluations. Many teams discover issues after release. How do you help teams trace the exact reason behind a quality drop across different agent workflow?

What kinds of failure modes or edge cases have you seen most often when an agent passes evaluation but fails in production, and how would real‑time scoring and drift alerts need to look for your team to trust them enough to act automatically?

Congrats on the launch! Does your tool make agents more smart over time by building right context around them, or is it still a responsibility of an agent?

Hey Product Hunt!

Sidhant here, co-founder of Cekura

Today we are launching self-improving loops for voice agents.

Fixing a voice agent has always been fragmented. Your testing tool tells you what failed, you diagnose it from transcripts, patch the prompt, re-run, and something else breaks. The tools find problems, but the fixing has always been a human walking between them.

Cekura collapses that loop. It runs thousands of simulated calls and groups every failure, explained in plain English. Optimise agent hands them to the Cekura Agent, or any coding agent you use, Claude Code, Codex, anything. It reproduces each failure, makes the change, and reruns until everything passes, then verifies nothing else broke. You review the diff.

Two rules: it must reproduce a bug before fixing it, and every fix is proven on simulated calls on a clone, never your live agent.

Free for everyone to try, starting today. We are in the comments all day. If you want to chat more, please feel free to book time here

Root cause diagnosis for voice agents is genuinely hard because failures compound across turns. If it traces that accurately, that alone is worth the pitch.

Reading this made me think about how much time my team spends manually reproducing a bad call just to figure out what went wrong. Automating that alone would be huge.

What stands out to me is the regression sweep after every fix. I've seen too many "self-healing" tools introduce new bugs while patching old ones. Curious how it measures whether a fix actually generalizes across edge cases.

How does Vocera handle accuracy and context when conversations get more complex or users have different speaking styles? Congrats @kabra_sidhant & team!

I've been burned by agents that pass QA and fail in production because the test scenarios were too clean. Genuinely curious how realistic it's simulated conversations sound.

Voice agents are unforgiving compared to chat, a bad pause or mistimed interruption sticks out immediately. I appreciate that it treats voice and chat as needing the same rigor rather than bolting voice support onto a chat-first tool.

The rule that a bug has to be reproduced before it gets fixed, and validated on a clone instead of the live agent, is the part that stands out to me. I build a voice companion that calls aging parents daily, and my worst fear is a silent regression in an emotionally sensitive call that nobody catches until it actually matters. How do you generate realistic edge cases for messy real speech (long pauses, hearing loss, someone talking over the agent) versus clean synthetic voices? That gap is where most of our failures hide.

the loop-closing part is the interesting claim. the failure mode i would worry about is the simulator quietly becoming the thing you optimise against, so the agent gets very good at passing your scenario set and no better in production.

we run agents against third-party systems we do not control, and the split that ended up mattering most was "the agent did the right thing" versus "the other side actually accepted it". those two look identical in a transcript and come apart constantly in practice.

curious how the regression sweep decides a fix generalised rather than fit the scenarios. is the held-out set drawn from real production calls, or generated the same way the training scenarios were?

It's like living in the future

The rule that a fix has to be reproduced and validated on a clone before touching the live agent is the right engineering call for anything running in production customer support. The part I want to understand is the CI/CD integration: when a new prompt version or agent config gets committed, does Cekura run the regression suite automatically as a pipeline step that can block a deploy, or is the test run still something the team triggers manually? That boundary between an automated gate and a manual check is usually where teams cut corners under deadline pressure.

Love seeing the focus on proving that a fix doesn't introduce regressions elsewhere. Reliable AI systems need repeatable validation, not just faster debugging.

Simulating messy real-world conversations with interruptions, pauses, and background noise feels much closer to production than traditional scripted evaluations.

Lets Go Team!!!!

Congrats on the launch. The hard part of a self-improvement loop is the blast radius. When an agent rewrites its own behavior from production calls, a fix for one flow can quietly bend a compliance flow sitting next to it, and in healthcare that is the thing every security review hunts for. Closing that loop safely, so the agent improves without drifting out of its guardrails, is the whole game, and it is a genuinely hard problem to have taken on.

the without overfitting part is doing a lot of heavy lifting in that description, and its the right thing to be worried about. an auto-fixer that cant prove the fix generalized is just prompt roulette.

I'm a huge fan of the work the Cekura team is doing. Closing the development loop with testing, evaluation, and paths to verifiable improvement is the single most valuable thing you can do in an agentic coding context. This is hard for voice agent development, for a bunch of reasons, including that there are many moving parts, gathering metrics robustly can be tricky and subtle, and success criteria are complex. Cekura has created great building blocks and high-leverage agent skills/workflows for solving these problems. Congratulations to the team on this launch, and all the effort and good thinking that went into it!

Love seeing more attention on production reliability for voice AI. The regression validation especially stood out. Curious whether Cekura can prioritize issues by business impact (for example, failed payments vs. minor conversation hiccups) or if everything is treated equally.

Congrats on the launch! Really like that you’re focusing on proving fixes instead of just identifying failures. One thing I’m curious about: how do you decide when a suggested fix is reliable enough to recommend versus flagging it for manual review?

Voice agents fail in ways text evals never catch — interruptions, latency, someone talking over the bot. The "loop" framing is the right one: testing voice once at build time is basically useless. Would love to know how many simulated calls it takes before the improvements show up.

Thanks — the GitHub Actions path is the answer to that question. The follow-up I'd have is whether the infra regression suite runs against a live agent clone or replays recorded sessions, because customer-support agents with integration state tend to behave differently on replay versus a live environment.

Does Cekura provide regression testing capabilities to catch quality drops when agents are updated or retrained?

How customizable are the evaluation metrics? Can teams define their own quality benchmarks based on their specific use case?

For teams building voice based agents specifically, does Cekura account for latency and audio quality issues, or is the focus primarily on conversational logic?

How does Cekura evaluate more nuanced qualities like tone, empathy, or conversational flow, beyond just accuracy or task completion?

Does Cekura support testing across multiple languages and accents, given how critical that is for conversational AI reliability?

When monitoring production calls, how does Cekura detect quality issues in real time, and what kind of alerting or reporting does it provide to teams?

Nice launch! What CI/CD tool integrations do you support?

For teams already using CI/CD pipelines, how much setup or configuration is typically required to integrate Cekura, and is it compatible with most existing tooling?

Hi PH!

On behalf of the Creator team, we are beyond excited to finally share our latest milestone with you: Lottie Creator 2.0!

Over the last year of building Lottie Creator, the same problem kept surfacing: on the modern web, there is a clear gap between simple tools and tools that enable you to unleash your full creative potential. That's where Lottie Creator 2.0 comes in to bridge that gap!

We’ve done our best to put YOU in the driver’s seat, all while automating the boring stuff away.

Boss on your back and you need to quickly align an entire animated brand kit? Use the Motion Systems and standardize look and feel across multiple animations. Working on a one-of-a-kind masterpiece? Use the new Graph Editor to tweak every important frame to tell your story (and lets be real: every frame is important). Need to blaze through adding interactivity for your animations? Use a combo of State Machines and Copilot to perform the repetitive clicks, while your brain focuses on your next creative idea.

We hope you give Lottie Creator 2.0 a shot, and let us know about your experience. The team is actively sourcing feedback (the good, the bad, the ugly) for us to take the product to the next level, so please don’t hesitate to reach out to us.

We’re so excited to see your animations out there in the wild!

Hi Product Hunter!  

Lottie Creator creator here (no pun intended). This is what's behind the Lottie Creator 2.0.

• We started Creator with a challenging goal: bring advanced, production-grade animation tools to the browser while lowering the barrier to creating motion.

• Historically, producing sophisticated animations for digital products has required complex desktop software, specialized expertise, and continuous installation, maintenance and updates.

• With Creator 2.0, we’re building a different kind of motion tool—one that is approachable when you’re juts getting started, yet powerful enough for experienced animators to create detailed, high-quality work.

• Creator gives you the precision you’d expect from a professional animation environment, including keyframe animation, a graph editor, and advanced easing controls to build complex motion directly in your browser.

• We’ve also designed Creator to be AI-first. Motion Copilot helps you move from an idea to an editable animation, within seconds. AI-powered raster-to-vector conversion remove repetitive steps and give you a stronger starting point. AI doesn’t replace the creative process—it augments your creativity where your judgment and craft matter faster.

Some of the biggest additions in Creator 2.0 include:

• Advanced keyframing and graph-editor controls for precise motion

• Motion Copilot and AI-powered creation workflows

• State Machines for building interactive animations

• Motion Tokens for creating dynamic, data-driven motion

• Motion Systems for standardizing animation across products and brands

• Multiple animations and states packaged into a single dotLottie file

• An extensible plugin ecosystem for expanding what Creator can do

Creator is for motion designers, product designers, developers, and teams that want to treat animation as a core part of the product experience—not as a decorative asset added at the end. Our larger vision is to make sophisticated motion easier to create, systemize, and ship at scale, without taking control away from the creator. We’d love your feedback on the advanced animation workflow, our AI features, and how Motion Tokens and Motion Systems could fit into the way you build motion today.

Thanks for checking out Lottie Creator 2.0! 

Really proud to be part of the team behind Lottie Creator 2.0!
A lot of work went into this release, and it’s great to finally share it with everyone.

I’m especially excited about the improved Motion Copilot. You can just describe what you want, and it can help you create animations using powerful features like Motion Tokens and State Machines.
It honestly feels a bit like magic. ✨

If you try it, please share your feedback with us. good or bad. We’re very open to hearing what could be better and what you’d like to see next!

Massive congrats to the Creator team on this huge 2.0 milestone! 🎉 While I’m not on the core Creator team myself, but my team had an absolute blast collaborating with them to improving and integration dotLottie Motion Tokens and State Machines into this release.

Seeing these features come to life directly in the browser is honestly surreal. State Machines completely change the game by making interactive animations incredibly accessible, and Motion Tokens opens up so many dynamic styling possibilities for production. Really proud of the cross-team effort here. I genuinely can't wait to see all the interactive experiments the community builds with this! 🚀

Congrats to the Lottie Creator team! I'm impressed by how easy it's gotten to make animations. Something like this would have required hours of tutorials in the past, but Creator is really intuitive.

I also really appreciate that it works on the web. I've been playing around with Linux a lot lately (on top of my Mac/Windows computers), and browser-based software makes compatibility a breeze. I don't need to worry about downloading and installing anything.

And the Motion Copilot AI is probably the first time I've seen a motion design tool implement AI properly, so major kudos to the people who put that together!

Best part of any launch: seeing what the community builds next. Show us what you make with Creator 2.0 especially motion copilot 2.0!

Motion Copilot has changed the way I experiment with animations. Instead of starting from a blank timeline every time, I can quickly generate a direction, iterate on it, and spend more time polishing the final result. It's become a regular part of my workflow.

One thing I really like about Creator 2.0 is that everything happens directly in the browser. There is nothing to install, and I do not have to keep switching between different tools just to try something out. I can open it, start working, and see how an idea feels almost immediately.

That small difference makes the whole process feel much more approachable. I am more willing to experiment, make quick changes, and test different motion ideas because there is less setup getting in the way. It feels easier to stay focused on the work itself instead of thinking about the tools around it.

I have spent a lot of time with LottieFiles Creator over the years, so seeing it out in the world feels really special. It has been a rewarding journey helping shape a tool built for motion designers. I hope it helps you create faster, experiment more, and enjoy the process as much as i have.

Directly gonna share this with our team of designers :)

First off, congratulations to the team on Lottie Creator 2.0! It’s been a milestone we’ve all been counting down to.

For me personally, my favorite features are Motion Tokens and State Machines. Interactive and responsive animation is just so fascinating to me, and being able to build that with the power of dotLottie still blows my mind.

As a non-designer, getting started was definitely a learning experience. But once I got the hang of a few tricks, everything started to click and the logic became more clear. I can't wait to see what everyone creates! 😊

What I love about Motion Copilot is that it gets me past the hardest part, which is starting. I feed it a rough idea, it gives me something to react to, and from there I just keep refining until it feels right.

I honestly didn't expect Motion Copilot to become such a big part of my workflow, but here we are. 😄

As someone who works with motion design and animation regularly, I constantly have ideas I'd love to try—but many of them never make it past the "what if" stage because prototyping takes time. Motion Copilot has completely changed that.

Instead of spending hours building a rough proof of concept, I can quickly explore different animation ideas, iterate on them, and see what works. That freedom to experiment has made me much more creative because the cost of trying something new is so much lower.

It's become one of those tools I open without even thinking about it. Whether I'm exploring interactions, testing animation concepts, or simply looking for inspiration, it helps me move from idea to prototype in minutes instead of hours.

The biggest value for me isn't just the time it saves—it's that it encourages experimentation. I've ended up creating animations I probably would have abandoned before because they seemed too time-consuming to prototype.

If you work with motion design, animation, or interactive experiences, Motion Copilot is one of those tools that quietly becomes indispensable once it's part of your workflow.

I've been using it for the past few months and I'm so proud of what the team has built here. I love the Physics Editor: makes confetti animations so fast and easy. Also Motion Presets and Easing Curves really help companies build their definitive Motion System, which enables a whole new level of consistency and speed on teams.

But without a doubt, my favorite part is still Motion Copilot. Watching an idea turn into an animation in seconds with a simple prompt never gets old. Feels like absolute magic.

It's been great working on Creator 2.0. I'm not an animator myself, and honestly Motion Copilot is what finally let me make animations I couldn't before. A lot of this release came from feedback and comments we received, from small fixes to State Machines and the new Graph Editor. Still plenty on our list. Can't wait to see what the community builds. 🚀

The State Machines feature is what I wanted to understand more — specifically how complex the exported dotLottie gets when you layer multiple interactive states. I use Lottie for small animated microinteractions in web projects and the dev handoff is usually where the friction shows up. Does the dotLottie bundle the state machine logic so the player handles transitions automatically, or does the developer still need to write state-triggering code around it? That answer changes whether this replaces the whole handoff or just the design side.

Hey product hunters, so stoked to see what you’ll be creating with Creator! 

Been experimenting a lot with the Motion Copilot myself. As someone without motion design experience, being able to quickly animate and iterate ideas has been really fun. Can’t wait to see how it fares in the hands of everyone else, whether you’re a seasoned motion designer, a complete beginner or somewhere in between!

Congrats to everyone who shipped this 🎉

Motion Copilot is getting most of the attention today, and fairly so, but the update I'd point people to is the Graph Editor. Easing is where an animation either feels right or feels slightly off, and that's usually the part that sends you back to desktop software. Having real curve control in a browser tab is what makes Creator 2.0 feel like a tool you can finish work in, not just start it.

From design assets to full-fledged motion creation, watching the evolution of LottieFiles' ecosystem continues to inspire. Lottie Creator 2.0 makes adding interactive motion so accessible without compromising on quality or workflow speed.

Kudos to the team on another epic launch! Wishing you all the success on PH today! 🙌✨

I'm really happy with the result I get from Creator. Not only I can create complex animations with Creator, but my workflow also is much faster and more efficient. Simply copy and paste .svg, animate, save it to my workspace and ship it!, all in one platform. I'm also thrilled to try really cool plugins inside Creator as well. Great work, team!!!!

Congrats to the whole team on shipping 2.0! 🎉

I've been somewhat involved in Creator's development on and off since the earlier stages, most recently contributing to the plugin system. I've also been using it internally as a testing ground for some of the work I'm doing, and the improvements in this release are really noticeable. Smoother, faster, more responsive, and just a lot closer to what you'd expect from a polished production tool.

Proud of how far it's come. Excited for people to try it out.

Really love the new Creator 2.0, everything feels fluid. Can’t stop playing witht the copilot to quickly generate the animation with just a prompt.

Really proud to see what the team has been able to build and achieve so far! Been trying it out internally and am amazed to see the potential that the community will be able to achieve with this product.

Such a powerful motion design tool, especially now with Motion System! But my personal favourite hidden feature is being able to change themes 💙

I run an app service packed with micro-interactions, and animation is a big part of making it feel fun and alive. I’ve always wanted to use Lottie more, but as a non-designer, the complexity of existing tools made it hard to get started. LottieFiles Creator and Motion Copilot 2.0 feel like a whole new world ~_~ they make the creative process so much more accessible and inspiring ✨

Not gonna lie, motion design used to feel intimidating to me, but Creator 2.0 makes it feel approachable.

Motion Copilot is my favorite, feels like having a motion designer assistant right inside the tool. Congrats team! 🎉

After trying quite a few browser-based animation tools, Creator 2.0 stands out as one of the few that feels powerful without being overwhelming. Motion Copilot is a nice touch for getting ideas moving quickly, the Graph Editor gives you the control when you want to fine-tune things, and I was genuinely impressed by how smooth and responsive the editor feels, even with more complex animations. Looking forward to seeing where this goes. 🚀

Having been on Creator since the early builds, seeing 2.0 go live is pretty surreal. I spent most of this cycle on the performance side, and that's meant obsessing over a ton of under-the-hood details. That's making sure everything from State Machines and the new Graph Editor to Motion Copilot feels insanely smooth, stable, and ready for real production work. It honestly feels like a completely different product now.

The part I'm actually waiting on is what everyone does with it, as that's always the best bit of a launch. Thanks to everyone who reported the rough edges and kept telling us what was missing. A lot of 2.0 is your feedback.

My favorite part is how easy it is to iterate. Prompt, tweak, refine, repeat. Watching an idea transform into an animation in seconds still feels kind of wild!

Honored to be a part of Creator engineering team.
Creator has been an ongoing project that's been incredibly fun to work on, from the earlier versions up until today. It's so much more stable, intuitive and capable than ever before. I hope you enjoy using it as much as we've enjoyed building it.

Hey Product Hunt 👋

Hardbook is a booking portal + contract signing in one link for day-rate freelancers.

I lost a job to a client who agreed on dates, never signed, and by the time I followed up the window had closed. The gap between "sounds good" and "officially booked" is where most freelance work dies — not on quality, not on price, just on friction.

The flow: send your Hardbook link, client picks dates from your Google Calendar, signs the contract, and — if you’ve switched deposits on — pays one, all in the same session. No app, no account on their end. Nothing moves forward without a signature, and the deposit goes to your own Stripe account, not mine.

motion designers, video editors, photographers, anyone billing by the day — to try it on a real client and tell me what's broken.

What's your biggest point of friction between "agreed on dates" and actually being booked?

Btw
- PHUNT65 + monthly → ✓ "65% off for 3 months"

- PHUNT20 + annual → ✓ valid (20% off the first year)

As a freelancer, the hardest part is often not finding work it's getting the commitment locked in. Nice solution! 🚀

Congrats on shipping! Every freelancer has a story about starting work on a verbal yes — putting the contract right into the booking flow is a clean fix. Wishing you a great launch day.

this is a real problem, not a nice-to-have. i've watched exactly this happen to freelancer friends - verbal yes, calendar hold, then silence right up until the date, and there's no clean way to force the moment where it becomes real. the deposit-tied-to-signature bit is the smart part imo, way more than the signature itself. one thing i'm curious about: once a booking is signed and the client needs to shift dates (not cancel, just move), does that need a brand new link/contract or is there a lighter reschedule path that keeps the same deposit and terms intact?

This is actually useful. Freelancers waste so much time going back and forth before a project even starts.

The no app, no account part for clients is what makes this actually usable, nobody creates yet another login just to book a shoot. I can see photographers and video folks adopting this fast. Are deposits or partial payments at booking on the roadmap, since chasing the deposit is usually the other half of the problem?

Every freelancer I know has been burned by starting work on a "verbal yes." Folding the contract into the booking link means the paperwork happens when the client is most motivated. Smart sequencing. Are the templates jurisdiction-aware, or one standard agreement?

Congrats on the launch, Adam. Skipping a client account and doing everything through one link is exactly the friction Hardbook is trying to kill. Since that link ends up carrying all the access control, is it a long random token that can't be guessed or iterated, or something closer to a sequential booking ID? That's usually the first thing I'd poke at on a no-login flow like this.

Love the no-account approach, that alone removes so much friction. One thing I'd add: a simple payment trigger tied to the signature, like letting the client leave a deposit or set up the first milestone right in that same flow. It would close the loop from "agreed" to "funded" without needing a separate invoice step.

love the framing that the job dies on friction, not price or quality. one thing I'd want as a freelancer: a signed contract still doesn't stop a client from just not showing up or cancelling last minute once the date arrives. is there any deposit or cancellation-fee mechanism built into the flow, or is that still something you'd have to chase separately even after they've signed

Hey ProductHunt!

I am Kevin, maker at Leaping AI.

Today we are launching our AI voice and texting platform, specifically for home remodeling companies. We started originally as a horizontal voice AI calling platform and have gradually niched down to the home remodeling industry.

Why? We signed up our first home remodeling companies because they approached us inbound via our website and we discovered that they have a list of specific pain points that makes our voice AI solution almost a no-brainer.

Pain points that we address

Home remodeling companies miss calls on the weekend or after-hours because no one is in the office, struggle to reach out to web leads fast enough over the phone and therefore lose leads to competitors, cannot serve Spanish speaking customers because usually no on on the team speaks Spanish and cannot (at scale) reactivate old leads that didn't buy in the past but might now be interested again.

Our solution

Our solution is AI agents that can take inbound phone calls and make outbound dials, specifically to schedule appointments and confirm appointments. They can hold 100+ conversations in parallel, are always active and stick to the script (a big problem in the industry). We can deploy both English and Spanish speaking agents. Inbound leads are being reached out to in under 10 seconds.

If leads do not pick up, our AI texting agents will send them a text message and follow the same conversation flow.

Our differentiator

Something that differentiates us is that the AI calling and texting agents can now be weaved together in multi-threaded campaigns. It's similar to sales software, like HubSpot, where you can create multi-day sequencing with email sequences mixed together with LinkedIn outreach and cold calling. In our platform, our customers can create similar campaigns where new leads, that have opted in to be reached out to (strict requirement), will be contacted via calling and texting until they pick up and schedule an appointment. Leads can at all times opt out of these campaigns by texting STOP via SMS. We maintain our own DNC database and strictly respect that.

It can be configured how long the campaigns last, what the exit conditions are and also what the cadence should be (e.g., how many calls and texts on day 7 vs. day 14 of the campaign). Home remodeling companies already have a similar process in place with humans and now our AI agents can complement that team and provide more firepower.

We have been live with this already with several large brands in the space, e.g. Bath Experts, Aspen Contracting, etc.

Our ask

Please give us feedback on our solution and what features we should build next. We are constantly iterating and are always open for new inspiration.

The "similar to HubSpot sequencing but for calls/texts" framing makes total sense, and the channel-agnostic DNC record (spoken stop = texted STOP) is the kind of detail most teams skip until it bites them. We deal with a version of this on the WhatsApp/email side, once someone's mid-conversation with our AI and a human needs to step in, we keep the same thread going so there's no "starting over" moment.

Curious how the handoff feels on your end when a homeowner insists on a real person mid-call, does the human rep see the full call history instantly, or is there a summary step first?

Kevin, the multi-day campaign is the part I would stress test, and it is not the dialing. You covered opt in and STOP over SMS, which is more than most voice products say out loud. What I would check is the spoken revocation.

If someone tells the agent on day 7 to stop calling, that counts even though it never touched your SMS keyword path, and it has to end the remaining calls and the texts. I work in a regulated space where consent is tracked per channel, and the failure always has the same shape: the person revokes on the channel in front of them and the sequence keeps running on the other one.

Does a spoken stop write to the same DNC record as a texted STOP, and does it close the whole campaign or only the call leg?

the reactivation angle is the one that would worry me most honestly. an old lead who didn't buy could mean "not ready yet, try again later" or it could mean someone who told a rep months ago they weren't interested and that just never got logged properly. both look identical in a CRM as "no purchase, no explicit opt-out on file." if the campaign resurfaces someone who already said no once, even politely, that's a much worse experience than a missed call. how do you handle that gap between what's actually in the DNC list versus what a homeowner remembers telling someone on the phone six months ago?

@kevin_wu25 Congrats on the launch! You’ve deployed voice agents into messy, real-world workflows where conversations can span multiple calls and texts over several weeks. What surprised you most about where AI works well and where customers still clearly want a human?

congrats on the launch!

the multi-day part is what stands out to me too, but from a different angle than the consent/DNC thread below - a lead's situation can just change between calls. someone's roofing quote request from day 1 might be moot by day 5 because they already hired a competitor or got it patched themselves. does the agent pick up on that kind of signal mid-call and update the campaign, or is there a real risk of the agent sounding tone-deaf by following up on a problem the homeowner already solved elsewhere, which seems like it'd burn trust faster than a missed call ever would

Missed calls and slow speed-to-lead are massive pain for local service businesses, banks especially lol. Putting appointment scheduling and confirmations on autopilot with multi-day campaigns is a game-changer for trades. Congrats on the launch, goated!

multi-day is the part that sounds small and is not. a campaign running for weeks means the agent's picture of a contact has to survive between sessions, and every stale field is a chance to say something that was true last tuesday.

the thing i would want measured is the gap between "we placed the call" and "a human heard it". we work on an adjacent problem, confirming a form submission actually registered on someone else's system, and the honest version needed independent confirmation rather than our own send log. voicemail, a carrier drop, and someone who hung up in two seconds all look like a completed attempt from the sender's side.

how are you counting those?

Good luck with the launch guys!

Leaping has come so far from its first launch! What challenges did you have to solve to maintain context and memory across these channels?

That's interesting. When a homeowner insists on a real person, how does the handoff actually work?

The horizontal to home remodeling niche-down is a smart call, the physical-world trades are so underserved by voice AI. I work on the consumer side (daily check-in calls to aging parents), and the thing that quietly kills outbound for us is carrier spam labeling: even a wanted, friendly call gets flagged "Scam Likely" and never picked up. With 100+ parallel outbound calls, how are you handling number reputation and STIR/SHAKEN attestation so calls actually connect? Would love to hear what held up at scale.

Sourcing real in-stock parts is the underrated feature here, half the pain of hobby electronics is finding out the part in the tutorial went obsolete years ago. The breadboard to perfboard progression is a nice touch too. How does verification work under the hood, does it actually check electrical constraints like current limits or is it matching known patterns from the parts library?

The orchid chamber origin story is great, that is exactly the kind of project people abandon once the wiring gets confusing. This would get me to finally attempt a hardware side project instead of stopping at the idea stage. With 290 parts in the library so far, how do you decide what gets added next, sensors, motors, or whatever users request most?

Do you see EasyCircuit evolving into a full KiCad workflow? It would be amazing if beginners could go from a plain-English idea all the way to a PCB while learning why design decisions were made.

Hardware is the last place where iteration still costs weeks and real money, so lowering the barrier here matters more than another web-app builder. The question I'd have as a user: how far does it get me before I need an EE to check the work? Congrats on shipping.

the fried-sensor case is one thing, but I'd want to know about the failure modes that are actually dangerous rather than just annoying - anything touching mains voltage or LiPo battery charging. does the tool flag current/voltage limits that could cause a real safety issue (overheating, fire risk) as a hard stop, or is that still something a beginner could accidentally wire past since it technically "works" on the breadboard?

The prototype is the easy half, honestly. I do sourcing and QC in Yiwu, and what usually stalls people is the handoff: the factory asks for a BOM with acceptable substitutes, and the answer is "whatever the prototype used." Then someone swaps a capacitor to hit a price target and nobody catches it until the first batch lands.

Curious whether you export anything a factory could actually quote from, or is that outside what you're solving?

A schematic view export would be amazing, especially as a downloadable PDF or SVG. Right now you get the parts kit but if I want to tweak the design myself or share it with a maker friend, I have nothing to work from visually. That would make this way more useful for anyone learning along the way.

Looks like I'm about to lose myself in this for quite a few fun evenings! :)

the orchid growth chamber backstory sells this way better than a generic pitch would. one thing that gives me pause with AI-designed hardware specifically vs AI-designed software: a bad software deploy just rolls back, but a bad resistor value or wiring choice doesn't show up until you've already soldered a real board and maybe cooked a sensor. does the breadboard stage actually catch those mistakes before anything's committed, or is it mostly a physical-layout check rather than a real electrical sanity check?

Hey Product Hunt! 👋

We built Ycode to give designers and developers more control over how they create websites, and this launch takes that idea further.

You can now connect Claude, OpenAI, Gemini or Grok and use your preferred AI directly inside Ycode. The agent can help you refine designs, manage CMS content, and create or update components without leaving the editor.

Instead of generating a website once and starting over, the AI works with your existing project and helps you keep improving it.

We’d love to hear what you think, how you would use it, and what you would like us to build next. Thanks for checking out Ycode! 🙌

Love that I can use my own Claude/OpenAI key instead of buying AI credits with a markup. You pay your provider directly and Ycode charges nothing for AI usage.

Huge congrats🙌 on the launch @Ycode team.. giving creators choice over their AI model directly inside a visual builder is a game-changer qq Is there a version history rollback for AI edits, so if a prompt yields a messy layout we can instantly restore the canvas to the previous state?

the bring-your-own-model part is the interesting risk here. we run a browser-driving agent and swapped the loop model across four candidates on one fixed task, and the failure that cost us most was not a model that errored. it was one that reported every step as successful while committing nothing, so the run read clean and the output was empty.

if your agents write to the CMS, that shape is worth guarding against, because a tool result saying it updated something is the model's own claim rather than the CMS's. do you read the record back after a write, or take the tool return at its word?

Since you support Claude, OpenAI, Gemini and Grok, are users bringing their own keys per provider, or would you consider a single unified endpoint to abstract the model routing and billing? Curious how you're handling failover when one provider's API is degraded.

Love the approach here. Bringing your own API key and paying providers directly is a much more transparent model than the usual markup. Pair that with being open source, and it gives users real flexibility without vendor lock-in. Wishing the team a fantastic launch!

Giving users the freedom to work with Claude, OpenAI, Gemini, or Grok directly inside the builder is a strong level of flexibility. Can users easily switch models within the same project while keeping the previous context?

Most AI websites tools optimize for convenience @lunenas . This seems to balance convenience with flexibility

Congratulations! There are quite a few website widget tools out there. What do you think is the one feature that makes Jotform Website Widgets stand out from the rest?

jotform is been my all time used product and this new feature makes it even better

Jot has evolved a lot! Been watching Jot for so many years now. Congrats on the launch!

Honestly, the drag and drop builder is way more intuitive than I expected. Threw together a feedback survey in like two minutes without touching any code.

honestly the drag-and-drop builder feels really polished, like every element just snaps into place without that finicky alignment frustration you get with other form tools.

Used Jotform to spin up a feedback form for a small project and was honestly surprised how fast it came together. The drag and drop felt snappy on my laptop, and the conditional logic just worked without me needing to mess with anything.

Jotform has grown far beyond a form builder, and this feels like a logical extension of the ecosystem rather than a random add-on. The individual widgets may not be entirely new, but having reviews, popups, FAQs, booking, chat, and forms under one account could remove a lot of small-tool fragmentation for website owners. I’m especially curious about performance — how lightweight are the embeds when several widgets are used on the same page? Congrats on the launch!

One thing I'd love to see is a built-in conditional logic preview mode that lets me test branching paths visually before publishing. Would save a lot of back-and-forth with test submissions.

Hey hey fam — time for me to jump in the maker seat!

Since January I've done nearly 200 coaching calls focused on launch positioning and product feedback for Product Hunt.

I've watched more great products die on bad copy than bad code.

And now thanks to agentic coding, non-developers like myself can finally bring their ideas to life.

The problem is — all those non-devs need to connect with their audience, and AI slop isn't going to cut it.

To stand out in a leaderboard that regularly exceeds 1000 products launched every 24 hour period, you need powerful, muscular prose that grabs the audience by the lapels and demands they pay attention.

Firstpass gives you feedback before you launch⁉️

Not "make it pop" — a real adversarial content crit.

Here's what few people realize: the same tagline lives or dies differently depending on where it appears.

Product Hunt gives you 60 characters. The Glaze Store caps at 40. The browse grid where people actually scan cuts near 33. A tagline that's perfectly legal in one place is unpublishable in another — and you find out after you've shipped.

⚡ How it works: you paste your launch copy once. Firstpass renders it on every surface a stranger will actually meet it — PH page, the leaderboard row, the store grid — and pins handwritten crit notes to whatever breaks in each. It scores against a real rubric: tagline, description, gallery, maker comment, message consistency. Calibrated against 305 live listings.

🔎 Why it's different: other tools rewrite your copy. Firstpass shows you where a stranger stops reading — per surface, before you publish. It's a design crit wall, not a chatbot.

🕵🏻 The proof: I ran Firstpass on this launch. It caught that my tagline gets cut mid-word on the Glaze grid — the exact failure the app exists to find — and flagged an earlier draft of this comment for ending on two questions. It critiques itself.

It's early — pre-beta and free while under development. If you find the concept compelling, you can get it from the Glaze store and if you like it, I'd love your support in the Glaze Awards: 🫶🏻

The useful part for me is that it shows where a stranger stops reading instead of rewriting the copy for you. One question: when a crit note flags a break, does it explain the principle behind it so it carries to the next launch, or is it scoped to that one listing?

Launching today myself, sitting at 3 points, so your timing is either perfect or cruel.

The feedback I actually needed came earlier than launch day. Our tagline went through four versions and the one we shipped came from cutting it to fit the 60 character limit, not from anyone telling me the earlier ones were weak. That forced edit did feedback's job by accident.

Does it work on a pre-launch draft, or only once the page is live?

Reading copy where it actually renders instead of where you wrote it is the whole trick, and it generalizes past launches. Last week I ran a copy and alignment pass on my agency site and every fix that mattered came from checking each string on the rendered page, the mega menu and the footer both read fine in the file and wrong in the browser. Out of the 305 live listings you calibrated the rubric against, which surface kills the most taglines? My money is on the 33 character browse grid, since nobody writes for a surface they cannot see in the editor.

Chris, the per-surface render is the right primitive, and I want to push on one axis the crit wall can't see by construction.

I do this dance on the App Store, where the same string holds two jobs at once. The title is a headline a human skims in a results row, and it's also an indexed retrieval field. So when a character limit forces a cut, the visible fix and the invisible cost land in different places: the tagline reads tighter and cleaner, and you've quietly dropped the word you were ranking on. Nothing on the rendered surface tells you that happened. You find out weeks later in the impressions, if you go looking.

Product Hunt is friendlier here because the leaderboard row is mostly a reading surface. It isn't purely one though. There's search, there are topic pages, and there's whatever the models are ingesting now when somebody asks them for "tools like X".

So: does the rubric have any notion of a term that has to survive, or is it scored entirely on where a stranger stops reading? A crit that says "this cut reads better and costs you the only word anyone searches for" is a note I've never gotten from any tool, and I've had to run that check by hand every time.

A company brain for AI agents (the question is why)

So I posted one simple question on Reddit about a frustration I have while using AI.

got 8.5k views in 5hrs with 45+ comments

Almost nobody said the problem didn't exist, instead everyone shared their own workaround.

Some use Chatgpt Projects. Some maintain PROJECT.md or CLAUDE.md files.

Some build knowledge graphs. Some create automation pipelines with n8n or Zapier.

Some use RAG connectors and custom memory systems.

The pattern was surprisingly consistent.

The real problem is Claude doesn't know what's happening minute to minute people said use an API or connect it manually. But then you're still feeding it everything including the useless stuff so we added an approval layer you decide what goes to AI combined with MCP, no need for Make, Zapier, or n8n it just works.

And what's the output

AI agents 10x better, knows what's happening minutes by minutes + no daily updates

Very Happy to announce FlowTask on Today's Product Launch. After working for months on out SaaS we have finally launched FlowTask 2.0 which has much better Operations Management, now comes with a larger AI context for companies with large number of employees. Hope ya'll enjoy it!! Happy Coding.

I like that you're tackling the context problem instead of trying to cram more data into the model. Congrats on the launch!

Congrats on the launch. The approval layer is the part I like most — most tools that pull from Slack and WhatsApp just take everything, so personal and work chats get mixed. Letting you decide what actually enters the brain is a smart call.

One thing I keep thinking about: when the memory updates every few minutes from so many channels, how do you handle facts that go stale or contradict each other? Like a decision made in Slack last week that gets reversed today — does the brain catch that on its own, or does someone have to approve the change?

The "approval layer so personal stays personal" line is the part I'd want to understand more. Once Slack, WhatsApp and Gmail are all feeding the same brain and any agent can read from it via MCP, is the approval/redaction happening per-source before it ever enters the brain, or per-agent at query time? Asking because those give very different guarantees — one keeps sensitive stuff out entirely, the other trusts every agent to respect scope once it's already in.

This is something I say to every C-level I talk to: if a person's knowledge doesn't become the company's knowledge, the hire was meaningless. Most teams lose that expertise the moment someone's out sick or leaves. Turning scattered Slack, email, and WhatsApp threads into one living memory solves something I run into constantly. Well done.

The approval layer on top of the ingestion pipeline is what makes this safe to use with real company data — without it, connecting WhatsApp and Slack is effectively giving every AI session access to every conversation that got pulled in. The thing I want to understand is the MCP server lifecycle: when an agent session connects to the FlowTask MCP, does the knowledge snapshot stay consistent for the duration of that session, or can it update mid-session as the brain ingests new messages? That matters for multi-step workflows where context drift mid-task would break the run.

the staleness questions above are the obvious concern, but there's a related one nobody's asked yet: if Claude and ChatGPT are both reading from the same brain at the same moment via MCP, and the brain is updating minute by minute, could two agents working in parallel end up citing two different 'current' answers to the same question just because one queried a few minutes before an update landed and the other after? not a bug exactly, just a consistency question that matters more once you've got multiple agents acting on the same context instead of one person reading it.

"No more stale CLAUDE.md files" is the right problem to name — we run agents across eight repos and those files rot faster than anyone gets around to updating them.

The thing I'd worry about with an always-updated memory is a quieter kind of staleness: a fact that was true the day it was written and silently stopped being true when the code changed underneath it. Nothing in the original Slack thread ever says it expired. Does the approval layer carry any notion of a fact aging out, or is it mainly a gate on what gets in?

Congrats on the launch! 👏

Curious—what's the one workflow your users keep coming back to?

both the Google Workspace and Slack connectors are broken which means there's no way to actually proceed into the app and try it out. bit disappointing and not a great experience for new potential customers

Peter, the fax detail is the part nobody outside these workflows believes. Healthcare runs the same rails, and the queue shaped exactly like yours is insurance benefit and eligibility verification, which sits directly on whether a practice gets paid.

What changes there is the content of the call. The moment your agent says a patient name to a payer, the recording, the transcript and the auditable result are all regulated records, and whoever holds them needs a signed agreement before a compliance team approves the pilot. That requirement stops this kind of automation more often than the phone tree does.

Have you pointed Ava at payer calls yet, and could a customer have the recording and transcript dropped so only the result survives?

Hey Product Hunt 👋 I'm Peter, co-founder of Superunit.

Every time someone gets hired, applies for a mortgage, or signs a lease, somebody has to confirm they actually worked where they said they did. In practice that's millions of phone calls a year to HR and payroll — leaving voicemails, sending emails, sending faxes (yes, in 2026), waiting days for a callback, getting transferred, leaving another voicemail.

Background screeners, lenders, and property managers hire whole teams to do this. It's tedious, expensive, and turnover is brutal because the work is mind-numbing.

Tbh we didn't set out to fix this. We went through YC's S24 batch working on something else entirely (in the accounting space), and at Demo Day it landed with a thud, which sucked. By the end of the year we made the call to rip it all out and start over.

The restart came from a chance conversation: a product leader at Checkr told us employment verifications were "soul-crushing." We got curious, started building in January 2025, and haven't looked back.

So we built Ava (not the most original agent name, I know…), an AI agent that runs employment verifications end to end. She does the contact research, calls HR and payroll, navigates phone trees, sends the emails and faxes, handles compliance, and publishes a clean, auditable result back to your system.

In the past year we've made 1.5M calls and emails to complete 200k verifications across 17 countries for 50+ customers, including some of the largest background screeners and lenders in the US. 70% completed, most back in under a day.

Building this also gave us a look at the verification layer almost nobody gets to see, so we're publishing some of it today in our State of Employment Verification 2026 report. One finding that surprised even us: when a verification routes through a named third-party platform, ~73% of the time it's a single company. More in the report → superunit.com/launch

Happy to answer anything, how Ava handles a live call, the compliance side, the international data, whatever you're curious about. Ask away, excited to be here. 🙏

this idea is great! one thing I'm wondering about is how the calls work. does it clarify its an AI?

with voice-phishing sounding this convincing now, I'd actually worry less about HR being cautious of you and more about the reverse: what stops a bad actor from cloning Ava's script and running the same kind of call to social-engineer real employee data out of an unsuspecting HR rep, using your legitimacy as cover? is there a way for the person on the other end to verify a call is genuinely from you, like a callback number or portal link, rather than just trusting whoever calls and says they're an automated verification agent?

Great idea and product for solving an everyday problem. As a business owner, I would get verification phone calls (and even faxes), all the time. Businesses want to respond, to help employees with their loans, rentals, etc, but the existing system was very cumbersome and annoying. Anything that streamlines and simplifies it is valuable, great use of technology.

Congrats on the launch, sounds like you've made some good progress. Question - what happens to a verification you can't complete? Does that happen often?

Congrats on your launch! Do receivers of the call like HR or employers get cautious cause its an AI Agent on the phone and don't want to confirm sensitive information?

Sheeta, the founding story has a detail the rest of the thread hasn't picked up. Your friend went quiet for ten months because she didn't have it in her to have a conversation. That same person also doesn't have it in her to complete a daily check-in.

Which is why the reciprocal gate is the thing I'd pressure-test. You offered it to Irene as the retention answer and it clearly works for the engaged middle. But it's a lock, and the people it locks out are the tail the app was built for. Three weeks into a bad stretch someone opens Cercle, gets asked to log first, and closes it. Her friends see nothing, and she sees nothing either, on the day it matters most.

I build in a different category, in-the-moment support for parents, where the user is depleted by definition, and the rule I keep coming back to is that any gate in front of the thing that helps gets paid by whoever is worst off. Mutuality is a fair principle when everyone is fine.

Two things I'd want to know. Does lighting the bat signal bypass the gate in both directions, so a low mood logged after a long silence still reaches friends who also haven't logged lately? And do you treat prolonged silence as a signal in itself? In your own story the ten months was the message, but a mood log renders silence as no data.

the reciprocal-gate thread above is the real design tension here, but there's a second one on the other side: what happens when someone logs low mood often, not once after ten months of silence, but as their baseline. does the friend group eventually stop reacting the same way to a widget that's been on "struggling" for three weeks straight? the whole point is catching the person who's usually fine and suddenly isn't, but I'd worry the signal gets noisier the more someone actually needs it long-term, not less.

People almost never say they're struggling, they just go quiet. Having your closest circle's state show up on your home screen before the phone call, without anyone having to perform okay-ness, matches exactly what I've seen after years of running group retreats. Loneliness spreads in silence long before anyone names it. Rooting for this one.

The reciprocal check-in mechanic — you have to log before you can see — is smart because it makes the value exchange explicit and keeps it feeling mutual rather than one-sided. What I'm curious about is the Cercle size itself: is there a cap on how many people can be in one circle, or does it scale to the size of your friend group? That changes whether this works as a tight inner-circle tool or something bigger with looser ties.

Great idea! Wish it was on Android too.

Love that the mood lives on a home screen widget, so support becomes proactive instead of waiting for someone to find the words to ask for it.

Would you consider making available on Android as well? :)

I really like the idea of focusing on close friends instead of trying to build another social network.

I'm curious: after people complete the first few check-ins, do they tend to keep using it consistently, or is building the habit the biggest challenge?

This is a considered take on a hard problem. A question about the self side rather than the friend side: does a daily mood check-in change how well people come to understand their own patterns over time, or is the log mainly a signal for others?

Sheeta, the screen that stopped me was the third check-in step, because Want to Isolate is the hardest state to build for and you already named it.

The preference lives in the log, but the delivery surface may not honor it. A low mood still lands on a home screen widget, and the person seeing it is the same friend from your story who sent countless texts. I build AI for clinicians and the rule there is identical: a stated preference about being contacted only counts if the system downstream enforces it, otherwise it is a note nobody acts on.

Does the widget hold or soften an update when someone picks Want to Isolate, or does the mood go out either way?

The decision to keep the metering core open source while charging for the rest is genuinely smart, basically lets people trust the math before they trust the product. Love that there's no revenue share either, feels refreshing.

The decision to ship the metering core as free open source while keeping the full stack paid is a really thoughtful move, it lets developers build trust with the hardest part before committing to the boilerplate.

The webhook story is the right thing to lead with. I shipped a Stripe integration this month where everything was green in test mode and the first real card 400'd on a currency param test mode never asked for - and separately found receipts failing silently because the send helper never throws and a try/catch was eating the reason. Money bugs don't announce themselves; you find them by walking the path with a real card and reading the logs like a skeptic.

Which is why I'd second Grace's checklist ask, and add one architectural vote: treat metering as a ledger, not a counter. Reserve before the metered call, settle exactly once keyed on the attempt, refund on failure - Gal's retry-dedupe question mostly dissolves when the settle is idempotent by construction. The counter tells you what happened; the ledger makes sure it only happened once.

Publishing mcp-metering free and inspectable is a genuinely good trust move. €79 for the edge cases someone already paid to discover is fair.

The zero-downtime API key rotation is the detail that would save the most pain in practice — keys usually get replaced when something breaks, not during planned maintenance, and any rotation downtime cascades into support tickets. The question I would ask before building on this: does the usage-based metering support different pricing tiers per API key, or is billing flat across all keys on the account? That matters when you want to give different API keys to different customer tiers without running separate instances.

The OAuth 2.1 + PKCE half is the part I'd have paid for. We ship an MCP server next to a local CLI binary, and the awkward case was authenticating a client that has no redirect target of its own — we ended up binding a loopback listener on 127.0.0.1 and handing back a http://localhost:<port>/ca... redirect URI. That took longer to get right than the billing did.

Does the boilerplate cover that non-browser client path (loopback or device code), or is it aimed at hosted MCP servers where a normal redirect URI already exists? Also curious what led you to one-time €79 with no revenue share when the metering core is already open source — that's a deliberate-looking choice.

The Stripe webhook example is a strong trust signal. For MCP servers, billing is one of those areas where AI-generated code can look complete while the edge cases quietly decide whether the product is usable.

One onboarding thing I’d want as a small builder: a checklist or test mode that proves the money path end to end — auth, metering, retry behavior, failed payments, and key rotation — before I connect a real server.

the webhook silent-failure story in your maker comment is the one that would actually keep me up at night, not the OAuth stuff everyone's asking about. on the usage-based billing side specifically: if Stripe retries a webhook (which it does on any non-2xx, including your own transient DB hiccups), does mcp-metering dedupe on the event id before incrementing usage, or is double-counting on retry something the integrator has to guard against themselves? that's the kind of bug that doesn't throw an error, it just quietly overcharges someone until they notice their invoice looks wrong

The OAuth half of this is the part I would pay for before the billing half. I connect MCP servers to my Claude setup as a user most weeks, and the pattern from that side is blunt, a server whose auth works on the first try gets used the same day, and one that fails sits unauthenticated for weeks. I have one in that exact state right now. Does your OAuth 2.1 flow cover the interactive consent dance the desktop AI clients run, or is it aimed at headless API consumers with keys?

the "calls nobody picks up at 11pm" line is the real value prop here, but that's also exactly when the riskier calls happen, a guest locked out with a medical issue, a smoke smell, someone who sounds actually panicked versus just annoyed about checkout time. does the voice agent have a hard trigger for escalating straight to a human on-call instead of trying to resolve it itself, or is that judgment call left to the model reading tone in the moment? for something touching door locks and payment systems already, I'd want the emergency path to be the most boring, hardcoded part of the whole system.

Hey Product Hunt 👋

I'm Justin, the maker of Liminal, and I'm stoked to finally share what I've been building!

Heads up: this is the first cut. I'm launching now to find out if the friction I've been feeling is real for other people too. Please let me know if you run into any bugs or problems, I'll get them fix asap!

Quick backstory

Last summer, the way I work changed. Agents got good enough to become the place I did the work. Claude Code became my main interface, the intelligence layer everything flowed through. My job shifted from writing and producing to directing an agent to do it.

Notion and Google Docs weren't built for this. They were bolting agents into their apps, but that's not where my agent lived, and their MCPs were slow and expensive to talk to. Claude Code worked much better and faster with local markdown files, a format the agent already speaks. So I stopped writing in Notion and Google Docs and went all in on local markdown.

I was doing all of this in VS Code because that's where I coded. But viewing and editing markdown there is clunky. You use the preview to render it nicely, then jump to the raw side to actually edit. So I built a WYSIWYG editor on top of the local files.

Then I tried to share something. Nobody wants to be sent a raw markdown file, and Google Drive doesn't render it nicely. Passing files back and forth killed any hope of collaboration. So I built real-time live collaboration on top of local files: a lightweight CLI watches your workspace folder and syncs bidirectionally with the cloud. Share a link and your teammate opens the same live workspace in a browser. Their edits land back on your disk instantly.

Once that worked, the thing I hadn't designed for showed up. My agent and my teammates' agents were all reading and writing into the same workspace. Liminal had quietly become a shared second brain. Company knowledge that keeps updating and compounding as the team works, sitting on every agent's local disk as plain files they can directly access.

What Liminal is

A collaborative workspace for humans, agents, and teams. Your AI-native second brain. Bring your own agents. Own your files.

How it works:

🟢 Work with your agent, not just through it. Your agent writes files to disk in markdown or HTML. Liminal renders them as clean UI the moment they save. Review and edit naturally in the browser, without wrestling with raw markdown.

🟢 Share instantly, without the friction. A lightweight CLI syncs your workspace to the cloud in real time. Send a link and your teammate is in the same live workspace. No sending markdown over Slack, no hosting your own HTML, no waiting on Google Drive.

🟢 A shared second brain for every agent on the team. Files live locally, so every agent has direct access to everything the team has written, decided, and learned. No external calls, no lag, no losing context between sessions.

🟢 No middleman, no MCP tax. Direct file access means no MCP round-trips, no proprietary formats, no wasted tokens. Context is cheap, fast, and always fresh.

The ask

Would love your feedback, especially if you're living in Claude Code, Codex, or another agent all day: does this map to your experience, or am I solving a problem only I have?

Justin

the "everyone with the link sees the same live workspace" part is the piece I'd want to understand before using this for real team work - once an agent is writing files into a shared space, it's going to write things that were meant for one person's eyes at some point (a draft perf note, a client's rough numbers, whatever). is there any per-file or per-folder permission layer, or is the model still "anyone with the link sees everything," just like the underlying markdown files would if you handed someone the folder

Everyone's building second brains for individuals. Almost nobody's built one for a founder plus their agents plus their people sharing the same context. That's the harder and more interesting problem. Good luck with the launch.

You're not solving a problem only you have. I've run exactly this architecture for months — a folder of markdown that multiple agents on multiple machines read and write, synced through git because nothing better existed. It quietly became the most load-bearing thing in my stack, so the friction is real and so is the payoff.

Some conventions from living it, offered as free product opinions since you said Liminal doesn't have any yet:

Per-writer append-only files — each machine and agent gets its own daily log — make conflicts structurally rare instead of resolved-after-the-fact. Shared files become the exception, and the exceptional ones are the ones worth merge UI.

One small always-loaded index with a hard rule that detail lives elsewhere. At 500 files the index is the search: the agent greps the index, not the corpus. Omri's question is the one that decides whether this scales.

An authoritative current-state file that wins when it disagrees with older notes. Abdullah's staleness case bites hardest — my standing rule is that notes record what was true when written, and anything actionable gets re-verified before an agent acts on it. A tool that aged or flagged entries when the thing they point to disappears would be the first genuinely new feature in this category.

And merging over last-write-wins is the right call — this week two of my agents wrote the same knowledge as different commits and taught me that multi-writer brains want merge semantics everywhere, not just the editor. Good launch. This one's aimed at a real hole.

this is one of those launches where the backstory tracks exactly with my own workflow - i ended up doing the same VS-Code-preview-then-raw-edit dance before giving up and just living in Claude Code with a folder of notes. the part i'm most curious about is scale rather than sync: once a team's second brain grows to a few hundred files across months of work, how do you actually find the right one? plain markdown in folders works great at 20 files and turns into its own kind of clutter at 500 - is there real search/structure on top, or is it still mostly relying on the agent to know where to look?

Love the positioning.

What type of teams are seeing the most value so far?

@justinonthelam The concurrent write answer above covers editing at the same time, but I'm curious about the offline case: if you're working locally with the sync CLI not running, say on a flight, does everything just work as plain markdown with nothing lost?

'work with your agent, not just through it' is the real reframe. when two agents write the same file at once, does it merge or last-write-win?

You asked whether this maps to people living in Claude Code all day, and from my desk the answer is yes, the friction is real. My whole setup runs on a folder of markdown memory files with an index loaded each session, and it quietly became the most durable part of my workflow because notes survive every context reset. The failure mode I keep hitting is staleness, a note written three weeks ago names a file or a flag that no longer exists and the agent recommends it confidently anyway. Does Liminal do anything to age or flag entries when the thing a note points to changes, or is pruning still on the human?

finally something that doesn't make me wrangle markdown back and forth. plugged in an agent output and it just showed up as a clean page my coworker could edit, which honestly felt kind of magical

the reversibility question above is the big one, but there's a smaller version that would hit way more often: if it's voice-first and always accessible, is it listening for a wake word continuously, or do you have to trigger it? because the failure mode I'd actually run into daily isn't a misheard delete, it's being on a screen share or call and someone else's voice on the meeting triggering an action mid-conversation. does it know the difference between you talking to it and you talking to a person while it happens to be open?

Finally tried Phantom and the notch placement feels genuinely clever, I can just dictate a quick question without leaving whatever I was working on. Screen context awareness works better than I expected for grabbing text from a doc.

I like the idea of making AI available everywhere instead of living in yet another chat window.

I'm curious: after using Phantom internally, what are the tasks where people naturally switch to voice instead of typing? I'd love to know what became the biggest "aha!" moment for your users.

the part I'm stuck on isn't privacy, it's reversibility. "click, type, organize, carry out tasks just as you would" means it can also delete the wrong file, send a message before you meant to, or submit a form, and voice commands are inherently less precise than a mouse click you can see land. what's the undo story for an action that's already irreversible by the time you notice it misheard you? until that's answered clearly I'd be nervous giving something this much control over my actual Mac, not just a sandboxed chat window

How about privacy? How and where do you process it? Do you then store it anywhere?

Seeing my own setup the way a stranger on the outside would is oddly reassuring, and I like that it just tells me plainly what looks open without any fuss.

Nice that curl here just streams text back, no piping into bash like typical installers. The 15-second abort window before scanning starts is a smart consent mechanism too.

Really like how frictionless this is. curl qsa.sh scanning the IP you're calling from means it's guaranteed your own host with no account or consent faff, which is a genuinely clever bit of design, and not retaining results is the right call for trust.

One question from my corner: for those of us mostly on serverless/edge now (Cloudflare Workers and the like) where there's no box with an open IP to point at, is there a version of this that makes sense, or is it firmly aimed at VPS/server setups? Either way, nice work.

honestly the one-curl approach is kind of genius — no signup wall, just instant results streamed to your terminal. love that the free scan stays honest about what the internet actually sees instead of burying the real findings behind a paywall.

This i s great! Would love to try it! ...but .....Is it opensource? I couldnt find a github link. Runing curl is in it self a security risk. Basically you allow anything to run on your computer. No gatekeeper nothing. So having it be closed source is a very red flag 😅. Its fine to do from big vendors like claud eor openai. But small vendors, opensource + curl is a must. IMO. Thanks 🙏

How is it different from Lovable or Replit (or any other similar tools that are able to code your own solution)?

Congrats, what early workflow usually gives users the quickest win?

the search-then-buy flow is clever, but the thing i'd want answered before using this commercially is rights, not workflow. "publicly available" and "cleared for training a model you're going to sell" are two very different bars, and that gap is exactly what's driving a bunch of the current lawsuits against scrapers. is there any licensing/consent layer behind the 25M clips, or is it on the buyer to figure out whether a given clip is actually safe to train on?

Searched for a tricky cooking motion I'd been hunting for and the re-rank after a few votes actually nailed it. The no-sales-call, pay-by-hour setup is refreshing.

It looks good, but a problem for me and for most people now i guess is that we don't store our music files, just use spotify or apple music

This app has been the highlight of my summer.

My music library feel alive now.

10/10 excellent project and incredible dev behind it.

This is very cool! Like a self-sovereign Spotify AI DJ :D

Love the self-hosted angle and the AI DJ taking plain-language requests is a great touch. One thing I'd love: a "discover from my own library" mode where the DJ only pulls from tracks you haven't played in a while, so it actually helps you rediscover music instead of leaning on the same favorites.