PH热榜 | 2026-07-20

Hey Product Hunt 👋

I’m Serena, founder of Fuzzy AI.

We kept seeing the same thing happen: someone sends a cold message to a complete stranger, it gets ignored, and they’re told to solve the problem by sending even more messages.

But what if volume isn’t the real problem?

What if the problem is that you’re still a stranger?

That’s why we built Fuzzy.

Fuzzy is a relationship-first AI sales platform that warms up your prospects before you ever reach out—so your first message doesn’t feel cold.

Instead of stitching together separate tools for research, LinkedIn engagement, content, outreach, follow-ups, and replies, you can manage the entire journey in one place.

Here’s what you can do with Fuzzy:

🔎 Find the right prospects

Tell the Fuzzy Agent who you want to reach. It can research companies, build targeted lists, enrich contacts, and find people already engaging with relevant creators or competitors.

💬 Build familiarity before reaching out

My Feed brings together posts from your prospects and key accounts. Fuzzy helps you write thoughtful comments and LinkedIn posts in your voice. You can also choose who Fuzzy engages with automatically.

🧠 Create outreach that sounds like you

Memories teach Fuzzy about your company, product, audience, offer, and writing style. Fuzzy combines that knowledge with prospect research to create personalised LinkedIn and email outreach.

⚡ Keep campaigns moving

Launch multi-step campaigns and let Fuzzy handle the follow-ups without manually managing every step.

📥 Turn replies into conversations

Fuzzy can draft or send replies using the full context of each prospect and conversation. You decide how much control to keep: review everything yourself, automate selected steps, or let Fuzzy continue until it’s time for you to step in.

👥 Work from shared context

Your team can share company knowledge, research, contacts, templates, conversations, and reporting in one workspace.

Fuzzy is built for founders, sales teams, agencies, consultants, recruiters, and anyone who relies on relationships to generate business.

Our goal is simple: less manual work and warmer conversations.

🎥 Watch the demo: https://youtu.be/PNs_VTpEbPo
👉 Try Fuzzy: www.getfuzzy.ai

Thank you to everyone who has supported Fuzzy. I’m excited to hear what you think 🙌

Serena

Hey everyone, I'm Bladen, CTO at Fuzzy 👋

My focus has been making all the different parts of Fuzzy work as one reliable workflow, from prospect research and message generation to campaign execution and reply handling.

One of the hardest engineering challenges has been balancing automation with user control.

We don't want Fuzzy to operate like a black box. Users should understand what the system is doing, review important actions, and decide when they want the AI to continue or when they want to step in.

If you test Fuzzy today, I'd appreciate feedback on:

Whether the review steps give you enough confidence

Whether the workflow behaves as you expect

Any bugs, slow moments, or unexpected behaviour you encounter

I'll be monitoring the launch and technical feedback throughout the day, so please don't hesitate to challenge what we've built.

Does Fuzzy adapt its commenting style depending on the founder's writing voice or industry?

Fuzzy got us 2 highly qualified leads in the first 3 days in a brand new country. Pretty amazing results, give it a try!

I've been trying Fuzzy, and what I like most is how much context it gives me before I reach out. I can see who's already engaging, understand why they might be relevant, and create a message that feels personal, all without switching between several tools. It makes the whole outreach process feel more natural and focused. Congrats on the launch!

There are many comments about the technology, so let me focus about the people behind Fuzzy: Fuzzy's team has been the most responsive and helpful to us. And Serena and her team were genuinely happy when we get leads :))

Thank you for your support!!

I've been using Fuzzy for the past couple months, really loving how centralized the app is and how I can execute multiple GTM needs in one app. My favorite feature so far is stealing my competitor's audience, I really recommend everyone to check it out!

Love the approach to warming up prospects naturally. How are you handling content matching/comment personalization so it stays authentic and avoids sounding AI-generated?

congrats on the much awaited launch! outreach without warming up your leads is like aimlessly shooting.. fuzzy is a great platform for outreach and increasing touch points.

So the fix for cold outreach was... not being cold. Incredible. Jokes aside this makes sense, curious what the reply rate diff looks like?

This reminds me of networking at conferences, you usually chat before asking someone for something. It would be awkward if someone just went and asked or pitched randomly without building the rapport.

I have used a lot of LinkedIn automation tools, they mostly optimize sending. You're optimizing recognition first, which is actually the right way. Congrats on the launch.

Serena, do you know what I like about your product positioning and why it immediately stood out? You're solving the problem before outreach even begins. This is exactly how every cold outreach tool should have been.

The idea of warming prospects before the first message makes much more sense than simply sending more cold emails which all the apps have normalised. Congrats Serena and team.

Personal branding and outbound have always felt disconnected. Good to see them combined into one workflow. Congrats!

When someone subscribes, they get all three sides of the product by default or there is an option to select what we want?

Bringing research, engagement, and outreach into one unified workspace is a massive UI win @serena_lam_!

In team setups where multiple reps or account executives are targeting different stakeholders within the same enterprise account, how does Fuzzy prevent duplicate 'warming' efforts? If Account Executive A is commenting on an Exec's post, does the workspace flag that context to Account Executive B so the team's outreach stays coordinated rather than spammy?

I always wished LinkedIn should have close connections list so we can engage with them in a separate feed. I will use Fuzzy mainly for this feature. :D

I like the idea of warming propects beore launch. Building familiarity first can improve response rates. How do you measure which interactions create the strongest trust? Showing those insights could help users refine their strategy.

Congrats on the launch! Curious how you're warming up contact without appearing insincere (I'm guessing there's the AI aspects on the automation)?

I like the idea because yes, you have low chances of getting a response from a cold outreach. But I feel like there's an issue which is authenticity because if I found out that someone had been warming me up with AI comments I'd have a different view on the person.

Interesting concept! Does Fuzzy adapt its messaging based on how prospects engage over time?

the warming-up approach actually works, saw real engagement on my comments before i ever pitched anyone. way less spammy than the linkedin tools i bounced off of last year.

The 'only comment when there's real overlap' call is the whole ballgame, and it's harder than it sounds. Running automated engagement myself, the asymmetry that bit us: a false positive, a comment where you don't actually have standing, costs way more than a miss, because a slightly-generic reply on someone's post reads as automated and burns the exact warmth you were trying to build. How are you tuning that overlap threshold, and do users get to see or veto the borderline ones before they go out?

Congrats on the launch, Serena. The 4% to 14% cohort you shared in the thread is the number that makes this interesting: warming as a measurable stage instead of a vague brand effect. My question is about audiences that live mostly off LinkedIn. We sell to healthcare professionals, and the busiest ones barely post, they read quietly if they show up at all. Does Fuzzy still find enough surface to warm an ICP like that, or does it need prospects who are actively engaging in-feed?

Finally tried this for my consulting work and the comment-warming angle actually feels different from typical scrapers. Surprised it didn't take much setup to get a sequence going.

One thing that would really round this out for me is a built-in deliverability health dashboard that flags risky sends before campaigns go live. Warming up inbox reputation is usually a separate chore, so baking it into Fuzzy alongside the sequences would save a lot of juggling.

This is cool. Which data sources do you source from?

Hey PH 👋 Baltej here, Founder of Nautis.

Over the last few years, I've been immersed in the startup ecosystem—building products, working alongside founders, collaborating with investors, and seeing firsthand what it really takes to build a company from zero.

No matter the stage, I kept seeing the same problem.

Founders don't spend most of their time building.

They spend it switching between CRMs, docs, spreadsheets, meetings, investor updates, invoices, emails... and now a growing list of AI tools that still don't understand each other.

Every tool has part of the picture.
Every AI knows one thing.
Nothing knows your business.

That frustration became the inspiration behind Nautis.

Instead of building another AI assistant, we asked a different question:

What if your entire startup lived in one workspace, with one AI that actually understands everything?

That's what we've spent months building.

🚀 A Fundraising CRM that knows your investor conversations.

📄 AI that turns your pitch deck into a business plan.

📂 Data rooms connected to your fundraising pipeline.

🤖 An AI Copilot that can draft emails, update your CRM, prepare meetings, and help across every module.

☀️ And my favorite feature: the Chief of Staff Briefing. Every morning, Nautis reviews your workspace overnight and tells you exactly what deserves your attention before your day even begins.

We're incredibly excited (and honestly a little nervous 😅) to finally share Nautis with the Product Hunt community.

I'd genuinely love to hear your thoughts.

If you could have an AI take one thing completely off your plate as a founder, what would it be?

We'll be here all day answering questions, collecting feedback, and shipping improvements.

And because every launch deserves a little fun... we even made a Nautis rap 😂🎤

🎵 https://www.youtube.com/watch?v=LzF-n2Kzxso

Hope it gives you a laugh.

Thank you so much for checking out Nautis and supporting our launch. It truly means the world to us. ❤️

Find me on:
Linkedin: https://www.linkedin.com/in/baltej-singh-46b78112
Twitter: https://x.com/baltejsingh88

----
One more thing! 🎉

To thank the Product Hunt community for being our earliest supporters, we're offering 3 months of Nautis Premium for the price of just 1 month.

Simply sign up, then head to Organization Settings → Offers and apply the code PHEARLYBIRD to unlock the offer. 🚀
----

We can't wait to hear what you build with it. 🚀

— Baltej & the Nautis Team

I'm impressed by how much thought went into solving a workflow problem instead of simply adding AI everywhere. The feature that converts a pitch deck into a business plan especially caught my attention because founders often repeat the same information in different formats. I'd love to know how much customization users have over those generated outputs before sharing them externally.

I like how Nautis connects fundraising planning finance CRM hiring and meetings into one workspace instead of spreading work across separate tools. How does the shared AI brain keep context accurate as the company grows?

Congrats. Context shared across every module feels much more powerful than disconnected copilots. BTW, are you a solo maker?

Nautis caught my attention from today's leaderboard. How does Nautis decide what deserves attention in the morning briefing versus what can safely wait?

"Shared data layer across all modules" is the interesting architectural bet here — most startup-OS tools bolt features together with separate databases and call it "integrated." Curious how you're handling permission/access boundaries across modules when everything shares one layer — does finance data get the same access rules as, say, hiring pipeline data, or is there per-module scoping on top of the shared layer?

Curious how this handles multi-founder teams — does each founder get their own view or is it shared state?

Can the AI recognize patterns across meetings and suggest priorities before problems become urgent?

Congrats on the launch!

Bringing fundraising, finance, meetings, documents etc. into one founder workspace sounds very practical.

I’m curious, is Nautis mainly designed to replace tools founders already use, or does it also integrate with tools like Notion, HubSpot, and Google Drive while keeping the shared AI context?

honestly the chief of staff concept is cool but it'd be way more useful if it could draft investor update emails from your actual metrics instead of you having to feed it context every time

Me seeing fundraising finance planning and CRM connected through one system makes this feel designed for real founder workflows instead of isolated features. That shared context could save hours each week. How easy is it to migrate existing company data?

Been using Nautis for a few weeks now. One thing that would make it even better is a simple integration with my calendar to auto-log meeting notes and action items into the CRM module. Right now I have to copy-paste, which slows things down.

The connected workspace angle actually felt useful rather than gimmicky, and the chief of staff suggestions picked up on context from my notes without me having to spell everything out.

This seems especially valuable for first-time founders juggling fundraising while still building product. I used to work with founders on fundraising so I know how chaotic things can get.

Founders who run multiple startups would they have two different workspaces??

I spend too much time moving informtaion across disconnected app so one operating system with shared context caught my attention immediately. The idea that every module uses the same AI brain feels more useful than isolated assistants. How do you handle permissions when different team members need different access?

Building products is fun. Managing startup operations... not so much. Nice problem to tackle. How long has been this into making? Seems like you went too deep with the product.

I already run something close by pointing Claude Code at separate repos per function (fundraising notes, finance model, CRM export), and the reasoning quality is fine. The wall I hit isn't the model, it's shared state: getting one agent to actually see the deck, the cap table, and the CRM at the same time without me re-feeding context every session.

Just interested - where do you think your value sits? The connectors and persistent memory layer or approvals/ audit trail for actions within Nautis?

I really like that fundraising isn't treated as a separate workflow from everything else founders already manage. This is such a refreshing launch.

good luck!

the "every AI knows one thing, nothing knows your business" framing is spot on. curious how the shared AI brain handles context switching between modules — like does something you discuss in fundraising planning automatically inform your hiring roadmap, or do you still have to connect those dots manually?

Great launch, congrats to the team!
Love how Nautis ties everything fundraising, CRM, finance, meetings into one shared AI brain instead of another disconnected tool.
The Chief of Staff Briefing feature is a genuinely smart touch for busy founders.

I have built a quick livedemo for you guys, if you ever think about adding a livedemo on your landing page
Or maybe sharing it with prospects

https://app.livedemo.ai/livedemos/6a5ebe6020e56dfe811abe04

@baltejsingh88 Congratulations. And happy product launch.

The 'one AI brain across all modules' angle is what makes this interesting, most all-in-one tools just bolt together existing features without shared context. Curious how you handle the cold start problem when a founder first onboards and the AI has no context yet. Is there a structured intake or does it learn purely from usage?

Congrats on the launch Baltej. I have a feature idea... you can also make a friendly leaderboard for team members. Based on their task completion, attendance and other contributions, they can climb up the leaderboard and win team member of the month badge. :D

One of the great products I have used, for the past few days, I was using Claude Co-Work, but it does not have multiple connections that I am expecting. Nautis gave me a variety of tools within the same platform, and I am very thankful to have this application.

Great job @baltejsingh88

fundraising, finance, CRM, hiring and ops all sharing one AI brain sounds great on paper, but each of those is normally its own deep, opinionated tool for a reason. how do you compete with a dedicated ATS on hiring or a dedicated CRM on pipeline management when you're spreading the same AI across 6 different jobs instead of going deep on one?

congrats on the launch! honestly the shared context layer is the part that gets me—so many founder tools just glue a bunch of features together and call it “all-in-one.” curious to try the morning briefing, that’s the kind of thing i’d actually check every day - and can share with my team.

Congratulations on the launch!

The "nothing knows your business" line really lands. founders already have enough tools, the real problem is that fundraising, finance, meetings, docs, CRM, and daily decisions all live in separate systems with separate context.

The morning Chief of Staff briefing sounds especially useful. as someone building a company with a small team, I'd probably want it to surface what changed, what is blocked, and what actually needs my attention before I open ten tabs :)

Curious how Nautis decides what is important enough to include, and whether founders can tune that over time based on what they act on or ignore.

Hey hunters 👋

I’m Anbang, founder of Loova.

We built Loova Ads Studio for one simple reason: making one good ad is not enough anymore.
To find what works, teams need to test more hooks, more formats, more products, and more creative angles. But producing all of that manually is slow and expensive.

Loova Ads Studio helps you turn proven ad formats, product URLs, and images into ready-to-test creatives.
You can create video ads for under $2, clone winning formats, swap in new products or characters, and generate unlimited product images in the same workflow.

Our goal is to make Loova feel like a 24/7 content engine for marketers, ecommerce brands, agencies, and creators.

We’re excited to launch here today and would love your feedback, ideas, and questions.

Thank you for supporting us ❤️

For the BGM it generates alongside visuals, does it actually sync to pacing/cuts, or is it more of a generic background track?

Great launch team, BTW, quick Q, when you don't give it specific shot directions, how good is the AI's judgment at picking angles that actually match the mood of the story?

A unified asset library with tagging and version history would be huge, especially when juggling lots of product images and video variants across campaigns. Right now it feels like outputs can get lost between generations, so being able to filter by model, style, or campaign and compare versions side by side would save a ton of time for marketers running multiple briefs at once.

Wishing you Good luck with the launch! :)

Congrats on the launch. The ability to produce more creative variations without adding a lot of cost is useful, especially for smaller ecommerce teams that cannot keep feeding an agency brief every week. I’d be curious about how Loova handles brand consistency. Can it learn a company’s visual style, tone and creative rules, or does each ad still need a fair amount of manual cleanup? Also, when you clone a winning format, how do you stop the output from feeling too similar to everything else already running in the market?

Sick product! How much is it?

@constance_tong Congratulations. And happy product launch.

the under-$2 UGC/avatar angle is genuinely compelling for testing volume. one question I haven't seen asked: since these are AI-generated avatars and UGC-style ads rather than real customers, do you build in any disclosure/labeling for the brands using this, given Meta and TikTok are both tightening rules around synthetic media in ads? or is that left entirely to the advertiser to handle on their end?

The under $2 per ad is what caught my attention. About to launch a mobile app and figuring out the creative side is honestly the part I was dreading most. Does the UGC format tend to work for apps or is it mostly ecommerce so far? That would actually make this a no-brainer for me.

My only question is what makes someone switch from the AI ad tools they're already using ?

Once generation stops being the bottleneck, the next wall is the ad platform itself. When we ran paid tests for our own launches we found that past roughly 6-8 creatives per ad set the spend splits too thin for any single variant to reach significance, so more variants past that point slows learning instead of speeding it. Curious whether Loova does any pre-flight ranking of the variants it generates, a predicted-CTR or angle-strength score, so you can ship the top few instead of firing all of them into the same budget and waiting for the platform to sort it out.

Would love to see a built-in brand style memory where I upload logos, fonts, and color palettes once and every generated asset automatically stays on-brand. Right now I still have to re-specify these for each new campaign.

I have a head full of video ideas and no patience for stitching a bunch of tools together, so this really speaks to me. Going from a rough thought to something watchable in one place feels freeing.

the campaign workflow is honestly smoother than i expected, basically just plug in a brief and it spits out usable assets without a ton of tweaking

would love to see a built-in brand kit so i can save my fonts, colors, and logos once and have every generated asset automatically stay on brand without re-uploading everything each time

honestly looks slick, but one thing that would help a lot is a built-in brand kit where you can save your fonts, colors, and logos so every asset stays on brand without re-uploading stuff. would save a ton of time for anyone running multiple campaigns

Would love to see a built-in brand kit where I can save my fonts, colors, and logos once and have every video or image auto-apply them, saves a ton of back and forth when running multiple campaigns

Congratulations on the launch!

Can users upload their own successful ads as references to create new variations while keeping the same winning style?

congratulations! which type of users have seen the biggest impact so far ecommerce brands, agencies or solo creators?

the workflow sounds very marketer friendly. how long does it typically take to go from a product url to a finished video ad ready for testing?

congratulations! the idea of creating video ads for under $2 sounds impressive. how does the platform maintain quality while keeping production costs so low?

Congrats on the launch!

How does Loova decide which ad formats are most likely to perform well or does it leave that entirely up to the user?

curious how you measure 'converts, not just looks good' before a campaign has actually run - is that based on historical performance data from similar ads, or is it more of a stylistic best-practices thing until you get real click/conversion data back?

Congrats on the launch! Turning product URLs into multiple ready-to-test creatives looks like a really useful workflow for smaller marketing teams. Can the finished ads be published directly to platforms like Meta or are integrations with ad platforms planned?

Congrats on the launch @anbangx @constance_tong

I'm definitely going to try this. as someone building a product and preparing more launch content, being able to test more hooks and formats without turning every ad into a full production sounds really useful :)

The under-$2 angle is also strong, especially for smaller teams that need volume but do not have agency budgets. Do you offer any special pricing or credits for early-stage startups? :)

The "gives your coding agent the root cause and fix" part is what stands out to me — most QA tools stop at "here's a failing session," this one seems to actually close the loop back into the fix. Did you find agents act on that context reliably, or does it still need a human to sanity-check the suggested root cause before it gets applied?

had the opportunity to collaborate on this launch and am blown away by how fast the first-time experience is 🤯

go to qa.replay.io and get your first report in seconds.

what's the model you're using in the background, ?makers?

This feels like a natural next step from the replay/devtools side, turning session capture into something that actively goes and finds bugs instead of waiting for a user to hit them. Since it's exploring the app on its own, is it built to dig into things like auth boundaries and cross-user data isolation too, the bugs that never throw an error and just quietly leak or drop data, or is the current focus mostly on crashes and broken UI flows? That category is brutal to catch in vibe-coded apps and it would be a big deal if an agent could reliably surface it.

Congrats on the launch, this is a sharp read on where QA broke down — AI cut build time to hours and testing never caught up. We ran into something similar: half our early bugs only showed up after the flows ran a few times, never on the first pass. The GitHub PR flow is the part I keep coming back to — root cause + fix posted right on the PR. How do you handle false positives from apps that behave non-deterministically (race conditions, flaky async UI)? Does Replay QA re-run a failure before flagging it, or does that triage still land on the dev?

The recorded session + root cause combination is what makes this interesting. finding that something broke is useful, but giving a coding agent the exact interaction, state, and context that caused it could remove a huge amount of back-and-forth between QA and engineering.

The first-time experience also sounds almost suspiciously fast :)) Curious how Replay QA handles flaky or environment-specific behavior. can it tell the difference between a real product bug, a temporary network issue, and something that only breaks under one browser or session state?

Congratulations for launching theproduct

This looks really useful. I like that you can record everything and inspect it later. It can save a lot of time when finding bugs and explaining issues to others. Great work! @Replay

Congratulations on the launch!

the recorded-session-to-root-cause handoff is the interesting part, most QA tools stop at "here's a broken screenshot" and this closes the loop back to something a coding agent can actually act on. one question I didn't see covered above: when the agent auto-explores an app that has real user data behind auth (not just a throwaway account), what happens to whatever it touches or records during that exploration - is there any scrubbing/redaction of sensitive fields before a recording gets stored, or is it on the team to only ever point it at a sandboxed environment?

This is awesome! I'm looking forward to digging into this. Might write a forum when I do. Congrats on the launch. Nice to see @mplacona in the mix. 🤩

Congrats on the launch folks! Can’t wait to take it for a spin

Used this on CodeTV! Amazing tool

Super cool! Just signed up, put in Git.Law, and it worked out of the box.

I've been looking for something like this.. surprisingly.. not found anything so simple and easy to use. Congrats on your launch.

The "give your coding agent the root cause and fix" handoff is the part I'm most curious about — I run multiple coding agents on a monorepo and the painful step is always getting the bug reproduced inside the agent's context. Is the handoff a replayable trace the agent can actually step through, or a text summary of the root cause? Congrats on the launch — the drop-a-URL one-time check is a smart low-friction entry point.

One thing that would make Replay way more useful for me: a "diff" view that lets you replay two recordings side by side and highlight the differences. Would be huge for QA teams trying to spot what changed between a working build and a broken one.

Root-cause handoff to coding agents feels genuinely useful. How do you decide which flows Replay QA explores first?

Recording right in their browser and being able to jump through the timeline with devtools feels genuinely useful, especially when sharing a weird bug with my team.

@mplacona Congratulations. And happy product launch.

the part that stands out is handing the root cause and fix straight to your coding agent. most QA tools stop at "here's a bug" and leave you to reproduce and diagnose it, which is honestly 80% of the actual work. closing that loop into the agent that can already fix it is the right instinct, the bug report was never the valuable part.

Congrats on launching a neat product! Do you guys crawl read-only paths first and earn its way into write actions, or is the expectation that people hand it a scoped throwaway account and let it break things?

Do we need to write test cases in our GitHub repo, or does the tool help us write the test cases for testing?
How does it work when we have feature updates

The 'gives your coding agent the root cause and fix' part is what makes this interesting. Most QA tools stop at finding the bug and leave you to figure out why. Been burned too many times by vague error reports that take longer to investigate than the actual fix. Any plans for mobile app support? Building a Flutter app and this would be genuinely useful for catching edge cases before users hit them.

This looks cool! How does Replay QA prioritize which bugs are worth fixing first?

Is there a way to give it a test email to enter a logged-in area of our browser-based software? Without giving Github access, just to act as if it's a real user setting up an account.

Congrats on the launch. The coding agent handoff is what caught my eye. Curious what the root cause artifact actually looks like in practice: is it a structured description of the broken flow, a reproduction trace, or something closer to a prompt the agent can act on directly in Cursor or Claude? Also wondering how it handles async UI states where the same event sequence sometimes resolves and sometimes fails, since that seems to be exactly what the thread is wrestling with.

I was given this tool during a time crunch and cannot overstate how useful it was in helping us find bugs in our application. It just...worked. Absolutely insane and feels like the natural next step in the evolution of dev-tools.

Hey Product Hunt,

Marcos here from @Replay.

Shipping apps has changed. AI coding tools make it much faster to build something that looks like it works. But the hard question hasn't changed: does this thing actually work in the browser?

We built Replay QA to answer that. Give it a URL and it explores your app like a real user, then produces a structured QA report showing what broke, where it broke, and exactly what your agent or team needs to fix it. Every session is recorded with our time-travel debugger, so you can replay the exact moment a bug occurred.

And if you're building something for the public - free tier included. If your app is posted to a public link (Show HN, Reddit, X, Product Hunt), you can test it for free. Create a project, set your public share link, and it becomes free immediately. Any existing credits are refunded.

Give it a try: qa.replay.io

We'll be here all day answering questions about what Replay QA checks, how the reports work, how this fits with coding agents, where it fits alongside existing testing tools, and how teams can use it before launch or in CI.

Would love to hear what you think - especially if you have an app we should test.

honestly this looks pretty cool, would love to see a way to redact sensitive info like api keys or user emails before sharing a replay. kind of a blocker for shipping these to support or external devs right now.

Hey PH 👋 Sagi here, founder & CEO of Skippr AI (backed by Bessemer Venture Partners and some other cool investors).

In the last year or so we've been obsessed with what happens when an AI agent can actually see your product, talk to your users, and act, live. Not a chatbot in the corner. A real (human-level) presence inside your product.

Here's what we learned: live agents fail when their human operators can't take a step back and treat them like employees. Ones you'd trust alone with your users, customers and internal users too. If the agent is too late to respond, out of context in long sessions, missing subtlety or tone, or needs the "boss" to work for it (updating the knowledge base, babysitting its learning loop), it brings more pain than gain.

So that's what we spent our time fixing.

We've been working with our design partners on our key use cases, such as onboarding and product adoption, and we're busy implementing more of them and building more agent skills. But today is about bringing what we learned to the wider community. We want anyone to be able to spin up, customize, and embed a live agent instantly. From Shopify stores and Lovable/Replit builders to any AI-first startup, just spin up an agent. Two lines of code to embed, or a simple meeting link.

You'll find live AI employees ("Buddies") inside for onboarding, training, demos, and support. But whatever you can imagine doing with a live agent, you can build it for your users or your team's needs.

Start free with 200 credits, startup package from $149. For serious volume or enterprise, come talk to me.

We're around all day. Try it at skippr.ai and tell us what you think.

Sagi & the Skippr team

Hey PH fam 👋

Excited to hunt Skippr AI for the global tech and startup community today!

Here’s the shift I think Skippr represents:

Software went self-serve two decades ago, and we quietly accepted the tradeoff. Nobody greets you when you land in a product. Nobody walks you through setup. Nobody notices when you’re stuck on step 3 and about to churn.

We optimized for scale and lost the floor staff.

Skippr puts that layer back. Think Apple Store Genius Bar energy inside your product, instead of a laminated FAQ binder. An agent that’s present in real time and handles onboarding, activation, demos, and support like a person would.

What impressed me when I tested it:

→ It doesn’t just tell users what to click. It clicks. Built-in browser automation means the agent completes actions on screen, which is the difference between a tour guide and a driver 

→ It holds agenda and memory across an entire session instead of resetting like most conversational tools 

→ Speaks 10 languages out of the box, so your onboarding works in São Paulo as well as it does in San Francisco 

→ Setup is genuinely instant. Trained on your docs, styled to match your brand, live in minutes

Sagi and the Skippr team are here all day. Free tier gets you 200 credits to try it, no payment info needed.

Question for the community: as agents get good enough to staff our products, which parts of the user journey should stay human, and which would you hand over first? Curious where people draw the line 👇

I see you support 10 languages with real-time switching. Does the user have to ask to change the language, or can they just start talking in the new language? Also, what if we need specific Asian languages that aren't part of the current 10?

The two-line embed is genuinely impressive, got it running in under five minutes. The on-screen automation feels surprisingly smooth for a self-serve setup.

This looks useful. Does the agent adjust its pacing if a user skips multiple steps quickly?

The browser automation is the part I'm most curious about. One thing that would make it feel way more trustworthy is a small audit trail window showing what the agent just clicked or typed, so users can follow along and intervene if it drifts. That kind of transparency would help a lot during onboarding flows where trust is everything.

I find the live employee concept interesting because it goes beyond chat support. What is the biggest limitation today? Being open about current boundaries could strengthen trust with potential customers.

What is the criteria for a problem that would be handed over to a human? Given that Skippr cannot fix it.

the "it clicks, it doesn't just point" part is the interesting bit and also the scary bit. once an agent is actually driving the UI instead of just narrating, what happens on a destructive action, like a delete or a cancel-subscription button, does it always pause for a real confirmation, or is that left up to how the workspace owner configures it?

How does the agent handle it when a user asks something totally outside what it was trained on? Does it hand off to a human or just say it doesn't know?

One thing that would make this a no-brainer for our team is a Slack or Teams integration where the agent can ping a channel when it gets stuck on something it can't resolve, so a human can jump in without the user having to leave the flow.

Actually impressed by how quickly the agent picked up on my screen during onboarding, felt almost like having a coworker guiding me. Wish the free credits went a tiny bit further but solid first impression.

The right boundary for this kind of product feels like assistance without taking agency away from the user. Let the agent complete obvious setup work, but make its planned action visible before it mutates anything important. That is where trust compounds.

kinda wild that it actually clicked around my app on its own without me hand holding it, felt like watching a coworker onboard a new hire. honestly the multi language thing surprised me the most, gonna throw a spanish speaking user at it and see

Embedded it in a quick test and the agent actually remembered context from earlier in the session, which caught me off guard. Browser automation felt snappy too, no lag clicking around the page.

Spent a few minutes with the demo agent and it actually clicked through screens like a real person would, not just scripted clicks. The multilingual response felt surprisingly natural too.

Spun one up in under five minutes and it actually handled a messy onboarding flow without me babysitting it. Surprised how natural the voice interaction felt across languages.

Congrats on the launch, Sagi! The line about live agents failing when operators can't treat them like actual employees really stuck with me, feels like the kind of lesson you only learn by actually shipping this stuff with real users for a year. Curious how the agent handles context across really long sessions, does memory ever get reset or does it carry the full history forever? Excited to try this on our onboarding flow.

How are you thinking about this vs. the tooltip-tour tools a lot of us already have? 

Serving every user autonomously is the scary part to watch as a founder, not the cool part. When it hits something it doesn't know, what happens, does it say so and pass it up, or guess? I run a hard no-improvising rule in my own product and even with that rule written down I still catch edge cases weekly.

Interesting. How much can I customise Skippr’s appearance? Can I incorporate my own design system?

Excited to see Skipper launch today! Congrats on the launch @sagi_shorrer1

I had the chance to try Skipper a few weeks ago! It was amazing how intuitive the whole experience felt. It doesn’t behave like a scripted chatbot or a traditional product tour. It actually understands what’s happening inside the product, sees the interface, talks to the user, and can take action in real time.

It feels much closer to having your best product specialist embedded directly into the experience! Ready to onboard, guide, train, and support every user.

The live see-talk-act loop is the hard version of this, and the part I'd want to hear about is what happens when an action doesn't land. Browser-driving agents break the moment the UI shifts under them: a modal pops, an A/B test moves the button, so the agent clicks and the page isn't in the state it assumed. Does Skippr verify the effect of each action before moving on, and how do you keep that vision-plus-action loop inside conversational latency so the voice side doesn't feel laggy?

Work in any language ? French and arabic?

Congrats on the launch, Sagi. In-product help that acts instead of pointing at docs is the right direction. We spent months getting new users to their first success moment faster (in our case, a clinician producing their first note in under ten minutes), and most of that friction lives inside the product, not in the marketing. One angle I have not seen in the thread yet: the agent sees the user's screen. For products handling sensitive data (health records in our world), can you mask fields or exclude zones from its vision, and where does that vision stream get processed?

Hi everyone!

Inkling is the first model from Thinking Machines.

It is a 975B MoE with 41B active parameters, a 1M-token context window, and native reasoning across text, images, and audio. The full weights are available under Apache 2.0.

Thinking Machines is clear about what Inkling is for. They say directly that it is not the strongest model available today. It is meant to be a broad base that can be adapted to a specific product or workflow.

You can control how much thinking it uses, fine-tune it on @Tinker, and deploy the resulting checkpoints through several inference providers.

They even had Inkling write and run its own fine-tuning job, turning itself into a model that avoids the letter “e”.

The bet is that a model shaped around your own work can be more useful than a slightly higher score on a temporary leaderboard.

Love that Thinking Machines put the honest framing up front, "this isn't the strongest model today, it's a base you shape around your own work." Most open-weights drops oversell the benchmark line, so leading with adaptability instead is a cleaner pitch, and Apache 2.0 on a 1M-context multimodal MoE is a real gift.

The thing I keep bumping on is that "open weights" and "actually touchable" aren't the same at 975B. Realistically, who fine-tunes a model this size outside of Tinker? Wondering whether the openness is meaningful in practice, or whether downloadable weights are mostly a trust signal and Tinker is the real on-ramp most people will have to take to do anything with it.

the framing of being explicitly not the strongest model but the best base to adapt is refreshing, most labs oversell the raw benchmark score. with 41B active params out of 975B total, what's the realistic hardware floor for someone fine-tuning this through Tinker, is this something a well-funded startup can do on rented compute or does it really need lab-scale infrastructure?

LoRA is great, but supporting full fine-tuning or QLoRA as a built-in option would make this way more useful for folks working on smaller models where LoRA just doesn't cut it.

finally an option that lets me skip the gpu setup dance and just get to the actual tuning. ran a small lora job last night and it just worked, which honestly surprised me

one thing that would make tinker a lot more useful for me is built in support for evaluating models right after fine tuning, like running a small benchmark suite automatically so you can see if your lora actually helped without wiring up a separate eval pipeline

Really interesting approach. What's the minimum dataset size you'd recommend for effective domain fine-tuning?

As someone who's never trained an AI before, it seems pretty approachable. You fine-tune open models on your own data, and they handle the heavy infrastructure stuff. Inkling also feels like it's meant to be adapted to your workflow. Good idea

Built a quick LoRA job on Tinker yesterday and the setup was honestly painless. One thing that would be a huge help though: a built-in diff viewer or summary that shows what changed in the merged adapter weights so I can sanity-check before pushing to prod without having to script it myself.

Would love to see built-in support for evaluating checkpoints mid-training, so we can compare LoRA adapters on a validation set without writing custom eval loops. A simple callback or webhook when a checkpoint saves would go a long way for experiment tracking.

Most of our pain when we LoRA an open model comes from the post-training rather than the weights. You tune for a narrow extraction job and the model keeps sliding back into the chatty assistant phrasing it picked up in RLHF, and you burn epochs beating that out. Is there a raw pre-trained checkpoint alongside the released one, or is the Apache 2.0 drop the post-trained model only? For something billed as a base to build on, that distinction decides a lot.

The 'goes stale the second I update it in one place' problem is the real test here: a shared context file only helps if every connected agent reads the current version, not a cached copy from last week. When I edit my Creed, does each agent pull it live at session start, or does it sync on a schedule and risk serving a stale snapshot mid-task? And is the source of truth hosted on your side, or a file I own locally that tools point at?

This scratches a real itch. I've hand-copied the same claude md into every new agent and watched it go stale the second I changed one copy, so one portable file every tool reads is the obvious right shape. Making it something you own and edit rather than a black-box profile is the part I like most.

One thing I'd want before wiring it into everything is read scope. My coding agent, an email assistant, and some random marketing tool don't all need the same slice of "who I am." Wondering if creed lets me control which connected agent sees which part of the file, or whether every tool I connect gets the whole identity doc. The least-privilege version of this is a lot easier to trust than all or nothing

Connor, the thing I re-explain most isn't who I am — it's my domain. I do research-heavy work, and every new agent has to be re-taught my field's vocabulary, which sources I actually trust, and the shorthand I use, before it's useful. So the tool I'd most want plugged in first is my knowledge base (Obsidian / Zotero / NotebookLM), because that's where the real "how I think" lives, not my chat history. One design question: do you separate stable identity context (role, preferences) from fast-moving project context? The first I set once; the second changes weekly, and I'd worry a single flat file quietly blurs the two together.

"Onboarding a new coworker every session" nails the exact tax I feel — I keep hand-copying the same CLAUDE.md into every new agent I try, and it goes stale the second I update it in one place.

A single portable creed.md that every connected tool reads is the right fix. Curious whether tools pull it in over MCP or need native support — but either way it's something I'd actually keep in my stack.

Nice one @connorhpbrn 👌

What I re-explain constantly isn't identity, it's rules, I keep a project guidelines file (folder structure, naming, "don't put business logic in components") for exactly this reason. But the failure mode I hit isn't staleness, it's silent drift: the agent has the file in context and still doesn't follow it three files deep into a task, and nothing flags that it happened. Does Creed have any way to verify an agent actually adhered to the context vs. just having it available? That gap matters more to me than sync timing.

the staleness/scope/precedence questions above are all the right ones. mine is more boring: if I stop paying or Creed shuts down someday, is the file mine to walk away with in a plain, tool-agnostic format, or does it live in a format/schema that's really only useful inside your own MCP layer? for something meant to be the one identity file every agent reads, portability out feels like it should matter as much as portability in.

Spent so much time trying to fix this in claude code, love this

Love how it puts the user in control of what agents actually know about them, rather than just scraping context behind the scenes. That transparent, opt-in approach feels like the right call for something this personal.

Finally something that fixes the blank stare my AI tools give me every time i ask for help with ongoing projects.

Really like this idea — "context that follows you across agents" is something I keep running into as a gap. Is it meant to sync automatically as your preferences change, or do you update the context file manually?

I've been solving this with a hand-maintained AGENTS.md per repo and the drift problem is real — it goes stale the moment my workflow changes. How does Creed handle updates: do I edit the file directly, or does it learn from usage over time? Also curious how you keep one shared context file from bloating once every connected agent starts contributing to it.

Congrats on the launch, Connor. This is a clean fix for the re-explain-yourself-every-time problem - once an agent's connected, does it only see the slice of your profile it needs, or does connecting an agent hand it your whole context file?

love the idea of giving my agents persistent context, honestly this could save me so much repeating myself across tools. one thing i'd love is a way to mark certain context as time sensitive, like project deadlines or temporary preferences, so the agent doesnt treat them as permanent facts about me six months later.

Portable context is a real need. The line I would want kept sharp is durable facts vs. session guesses: preferences, constraints, and project state should survive across agents, while assumptions need timestamps and an easy way to discard them.

This is clever. How does Creed handle context drift when you update the file mid-session?

The shared context layer is such a smart move. Finally my tools stop making me re-explain my preferences every conversation.

Finally feels like my AI tools are on the same page about my projects. Set it up once and noticed Claude and my other assistants stopped asking the same questions about my role and preferences.

Finally something that actually remembers my preferences across different agents, kind of a relief not to repeat myself constantly. Setup was quick and the context sharing feels pretty seamless so far.

Finally connected my daily tools and the agents actually pick up on context without me repeating myself, which is a relief. Setup was quick.

Honestly love how clean the onboarding is. The way it lets you shape context gradually instead of dumping a giant form on you upfront feels really thoughtful, like the team actually thought about how someone would want to build out their profile without it feeling like a chore.

the idea of one shared context file across all your agents makes sense, but it also means every connected agent gets the same level of access to everything about you - work style, preferences, whatever else ends up in there. is there any way to scope what a given agent can see, or is it all-or-nothing once you connect it?

Went through the site before commenting. The MCP pull is the right delivery choice, since a file fetched before the answer rather than pasted into a prompt is what makes this work across tools instead of inside one.

Andrew asked above whether the agent reads all of the file as it grows. I think the prior question is whether it ever shrinks. Both diffs in your own screenshots add lines and remove none, +18 and +9 with a zero on the other side, and that is the shape every memory product drifts into, because agents are good at noticing what to add and have no instinct for what stopped being true. Accept a year of those and it is a long preamble pulled before every answer, mostly stale, all of it paid for on every call.

So the question: does anything ever propose a deletion? A diff that says this line has not been true for four months, here is why, is the harder thing to build and the thing that keeps the file worth reading.

The other one is precedence. Your own example says ship to production Monday through Thursday. If my Creed says that and a repo's own context file says otherwise, I want to know which one the agent is meant to believe before I connect it to anything real.

the 'onboarding a coworker every session' line nails the tax. as the file grows, does the agent read all of it or learn which parts change its answer?

Congrats on the launch, Connor!

Not having to repeat my project structure and coding preferences to every AI tool I use would save so much time. Having a single context file makes total sense.

Rooting for you guys today!

This is dope crazy. I would normally spend over animated video more than one hour. You compressed the time. <3

Cool product, I need this for my channel. Quick question: for a series of videos, is there a way to keep visual style and pacing consistent across them?

Hi Product Hunt!!!!!! 🌊
besides being part of the Layerproof team, I also a content creator and do manage social media on daily basis. Tbh, I'm super excited about this launch because I truly believe it's going to be a total game-changer for my own content workflow!

as a content creator, making product videos used to be my biggest bottleneck. Turning a boring document or landing page into a clean motion video usually meant hours of scriptwriting, storyboard editing, voiceover recording and keyframe tweaking 😿

That's why building this feature was so personal to me.
I’d love for you all to give it a spin!
what’s your biggest struggle when creating product videos? Let me know below!

Hey Product Hunt 👋 Tracy here. I am a Product designer at LayerProof.

A little about the journey behind Mylar:

• I kept watching great products get flat, forgettable videos, not because the makers didn’t care, but because good motion design is hard and slow

• Every version meant re-cutting scenes, re-timing voiceover, re-exporting… and starting over the moment anything changed

• So we set out to make motion design feel less like a production and more like a conversation

What we obsessed over on the design side:

• Scenes that actually feel designed: real pacing, hierarchy, and rhythm, not just text on stock footage

• Regenerate any single scene without blowing up the rest of the video

• Refine every keyframe, so you keep full creative control

• Export cleanly for each channel, so it looks intentional everywhere it lands

I’d genuinely love for you to try it and tell me where the product shines and where it falls short. Feedback from this community makes the product better every single launch. Thanks so much for taking a look ♥️

Hey Product Hunt 👋

Ha here - one of the makers behind LayerProof. Excited to share LayerProof Mylar with you today.

If you’ve ever tried to turn a product into a video, you know the pain: you’re stitching together a script, hunting for scenes, recording voiceover, and re-editing every time something changes. It takes days, and by the time it’s done the product has already moved on.

Mylar collapses all of that into one flow 🎬. Point our AI agent at a product page, a doc, or just a prompt, and it generates a full motion video including script, scenes, and voiceover.

The goal was simple: a launch-ready product video in minutes, not days, without needing a video editor or a motion designer on call. We’d love your honest feedback. Tell us where it shines and where it falls short. We read every single comment and we ship fast 🚀.

Curious how it handles messy input. Does it still pull the right key points from a long, unstructured PDF?

When a teammate edits a scene I generated, does it keep my original context so the AI doesn’t lose the reasoning?

Congrats. Looks great. It's just what I need. Will give it a try. Where does it get images from? Your product page, you add them or does it generate them?

I want to try with videos that have human interaction showing real human faces but this is great with its stability to generate motion graphics video and voiceover

Does export lock in resolution and FPS, or can I export the same project at multiple settings?

The single-scene regeneration is probably the biggest win here. product videos change constantly, and having to rebuild the whole thing because one section, feature, or line of copy changed is painfully inefficient.

I also like that Mylar does not stop at "AI generated video", but still gives control over keyframes, pacing, hierarchy, and channel-specific exports. as someone doing launch prep, that balance between speed and editability feels much more useful than a one-click result you cannot really refine. Curious how well Mylar can preserve one consistent visual style across a full launch campaign with multiple videos and formats.

Hey Product Hunt 👋

A bit of the story behind Mylar:

As a designer, I kept seeing amazing products introduced through videos that didn’t quite capture what made them special. Not because teams lacked vision, but because creating high-quality motion design takes a huge amount of time and effort.

Every small update meant going back into timelines, adjusting animations, syncing voiceovers, exporting again, and repeating the whole process whenever the product evolved.

We started Mylar with a simple goal: make creating product videos feel as flexible and iterative as designing the product itself.

A few things we spent a lot of time getting right:

🎬 Videos that feel thoughtfully designed, with clear hierarchy, pacing, and storytelling
🔄 The ability to regenerate a single scene without affecting the rest of the video
✏️ Full control over keyframes and animations when you want to fine-tune details
📱 Exports optimized for different channels, so your content feels native wherever it’s shared

I’d love for you to give it a try and share your honest feedback. What feels great? What feels missing? Every conversation helps us improve the product and build something people genuinely enjoy using.

Thanks for checking out Mylar ❤️

When a teammate edits a scene I generated, does it keep my original context so the AI doesn’t lose the reasoning?

A built-in brand kit would be huge - let me upload our logo, fonts, and color palette so every generated video stays consistent with our brand without having to tweak each scene manually.

What's the tech / foundational model being used behind it?

would be great if it could pull in real product screenshots or UI clips automatically instead of just generating scenes from text, since that would make the videos way more accurate to what's actually on the page

The keyframe-level refinement is a really thoughtful touch, way beyond what most "AI video" tools offer. Cuts out that painful back-and-forth loop of regenerating an entire video just to fix one shot.

honestly the workflow looks really smooth but it'd be super helpful if the AI agent could pull in existing brand assets like logos and color palettes upfront so the generated scenes already match the brand without having to tweak every frame manually

Took my messy Notion doc on a side project and got back a clean 90-second walkthrough with actual scenes in under ten minutes. The scene regeneration was the bit that won me over, since tweaking one frame without redoing the whole cut felt way smoother than other tools I've poked at.

The script-to-scene flow actually works as advertised, the keyframe refinement step is where it really shines. Saved me an afternoon of editing for a launch teaser.

building a narrative from a doc instead of dumping scene-by-scene text is the real work here. how does it decide what to cut vs say everything?

Motion for launch assets is a narrower lane than most tools try to cover, which I like. Does it pull brand context automatically, colors, logo, tone, or does every launch start from a blank setup? Same problem I fight on the text side, most of my week goes into making sure nothing forgets what a brand actually sounds like.

Really respect the "no login, no cloud, doesn't train on your writing" stance — that's clearly a direct response to what writers are actually worried about right now, not just a feature checkbox. The index-card-wall analogy makes total sense as the thing to digitize. Does it support any kind of collaborative/multiplayer editing yet, or is it intentionally solo-first for now?

The infinite canvas idea sounds great for visualizing structure. One thing that would really help my workflow is a beat board view that auto-syncs with the canvas, so I can drag scenes around spatially but also see them in a linear outline that stays current. Right now I'd worry about losing the script's logical flow when everything lives on a canvas. A toggle between spatial and outline modes, with the outline updating in real time, would make this feel complete for actual drafting days.

finally a spatial tool that gets how my brain works during outlining — the infinite canvas for threading scenes and character arcs feels like it was made by someone who has actually wrestled with outlining before. exporting to fountain and final draft is the cherry on top.

Who doesnt love an index-card-on-a-wall tool for the browser. You can move scenes and ideas around visually, see the whole story at once, and then export to Markdown, Fountain, or Final Draft when you’re ready to write.

I applaud that it’s free, with no login, and nothing saved on their servers. It feels simple on purpose, like a visual development board. That’s actually what I would want from something like this :)

the index card wall replacement is such an obvious need once you say it out loud. does the export to Fountain/Final Draft preserve the spatial layout at all, or is it just a linear scene order once it leaves the canvas?

The Fountain and Final Draft export is the smart part here — most visual planning tools quietly trap your work inside their own canvas. When you export, does the spatial structure you've laid out (scene order, act breaks) carry through into the linear script, or does it flatten to a plain doc and leave you re-sequencing by hand?

the visual index card approach is how a lot of writers already work on a physical corkboard, so digitizing that directly instead of forcing everyone into a linear document is a smart call. is it built for solo writers only right now, or can multiple people work the same board together like a shared writers room?

The infinite-canvas-to-Fountain/Final Draft export is what would actually pull me in — most story-planning tools trap the structure in their own format and re-typing it into a script kills the whole point. When you export to Fountain, does the scene and character layout on the canvas carry through as scene headings and structure, or does it flatten into a plain ordered list? And if I keep moving things around on the canvas after exporting, is there any re-sync or is export strictly one-directional?

There is something about a wall of index cards that no outline ever quite replaces, so seeing that feeling recreated digitally really appeals to me. Knowing the whole story stays on my own device makes it even easier to trust.

 Congratulations on the launch! I read through the site but couldn't tell - when you rearrange cards on the canvas, does that order carry into the Fountain/Final Draft export, or does export just pull the text and leave sequencing to you?

the infinite canvas approach is honestly such a smart call for outlining scripts visually, beats cramming everything into folders or cards. also love that it exports to fountain without making you jump through hoops

Pinned my characters and locations to a canvas and the export to Fountain actually kept my scene structure intact. Wish I'd had this during my last rewrite.

Hi Product Hunt! 👋 I'm Akanksha, one of the founders of Backdrop.

AI has made execution dramatically faster. But as execution gets easier, the bottleneck shifts to deciding what to build, why it matters, how to prioritize it, and keeping everyone aligned. At the same time, the knowledge behind those decisions is becoming even more fragmented. It's scattered across Slack, docs, tickets, meetings, and now private AI conversations.

Customer feedback gets trapped in someone's ChatGPT. A feature gets rebuilt because nobody remembers why it was killed. Someone joins the team and suggests an idea that was already tried six months ago. Two teammates ask AI the same question without realizing the other already did.

The problem isn't that teams lack information. It's that they can't build on what they already know. Too often, what people learn in AI stays with them instead of the company. We think AI should make companies smarter, not just individuals. That's why we built Backdrop.

Backdrop provides AI coworkers for projects and operations that understand your company, work alongside your team, and carry context across every project and decision.

We're launching with Alex, our AI coworker for projects and operations. Alex connects to Slack, GitHub, Linear, Notion, Asana, Google Workspace, and more to turn customer feedback into product plans, discussions into decisions, meetings into action items, and plans into execution.

If your team ships software, Alex can also bring in Sam, our AI engineering coworker, to implement features, review code, investigate bugs, and turn plans into shipped products.

We're incredibly excited to finally share Backdrop with the Product Hunt community. We'd genuinely love your feedback, questions, and ideas. We'll be here all day, so ask us anything. Thanks for checking us out! 🚀

the "someone joins the team and suggests an idea already tried six months ago" scenario is such a specific and true pain point, that's the real cost of tribal knowledge living in people's heads instead of anywhere searchable. when Slack, Linear, and a doc genuinely disagree on the current state of something, does Backdrop surface the conflict to a human, or does it pick one source as authoritative and move on?

This is cool, all these shared context cos are building stuff for the user but for not the org and its hard to flow context (and importantly, source of truth) across the org. Also daisy chaining tool use is pretty unique in this context.

Is orchestrator multi model?

love how you pulled together so many integrations into something that actually feels cohesive instead of bolted on. the shared company context piece is honestly the part most tools skip, so good to see it done right.

the part that's interesting to me is the cross-tool synthesis rather than just another single-tool assistant - pulling customer feedback and turning it into an actual plan across Slack/Linear/Notion is the annoying manual glue work most teams still do by hand. what's been the hardest integration to get genuinely useful vs. just surface-level read access?

What's the differentiation between Backdrop and other AI team apps?

This really resonates! AI has made building faster, but deciding what to build and keeping everyone aligned on the context behind those decisions is still incredibly hard. The approach of bringing that scattered knowledge together and turning it into actual execution feels very timely.
Congrats on the launch!

Really cool and useful product backed by a great team!

Really like the vision behind this Keeping context inside the company instead of losing it across chats feels incredibly valuable Congrats team

this is so exiting. the shared context and information, along with actions to do tasks. will definitely cut down the time our team spends on working with AI agents rather than these agents for working for them. Are Alex and Sam purpose built? Or can they be used interchangeably?

the integration list is honestly impressive, basically covers every tool my team already lives in. love that the AI coworkers actually build shared context instead of just answering questions in a vacuum

AI coworkers become useful when the handoffs are legible. For projects and operations, I would want every automated step to leave behind the why, the source, and the next human decision needed, so the team gains leverage without losing accountability.

Would love to see a quick weekly digest mode where the AI coworker summarizes what it touched across Slack, Linear, and Notion so the team can skim changes without digging thread by thread. That kind of cross-tool recap would save a lot of context switching for everyone.

honestly the shared company context idea is really compelling, one thing i'd love to see is a way to mark certain decisions as final or source of truth so the ai doesn't second guess them when pulling context for new projects, basically a confidence flag for historical calls

Pulled feedback from our Slack threads and turned it into Linear tickets without me babysitting the process. Genuinely useful for cutting down the back-and-forth.

Looking forward to this. It is going to help employees with task management and delegation to agents that get stuff done efficiently.

Cool!

Information silo is so detrimental to most teams. Pumped to see how this performs!

I like the idea of AI coworkers having a persistent understanding of company context instead of starting every task from scratch. Continuity is something most current AI workflows still struggle with.

This is neat. How does it decide which tasks to handle autonomously versus loop you back in?

Everyone here is reacting from the "team with scattered Slack/Notion knowledge" angle, which makes sense since that's who you built for, but I'm curious where the line is for a one-person team. If it's just me, GitHub, and a task tracker, is there still enough fragmented context for Alex to be worth wiring up, or is the value fundamentally about reconciling multiple people's half-knowledge? Not knocking it, genuinely trying to figure out if this category applies below a certain team size.

The shared company context across Slack, GitHub, Linear and Notion is the piece I would test first — for a community or product team the risk is not capability, it is a coworker surfacing something from a private channel into the wrong doc. On setup, can I scope what each coworker reads — connect Slack but exclude specific private channels, or restrict a coworker to one Linear team — or is it all-or-nothing per integration? And when it synthesizes customer feedback into a spec, does it cite the source threads so I can trace a claim back before acting on it?

that's a great breakdown, especially the point about reading data being maybe 20% of the work. the Google Workspace / Figma example makes sense too, since a PM and an engineer looking at the same file are basically doing 2 different jobs even though the underlying API call is identical. are you building the role-specific behavior as hardcoded rules per app, or is there a more general framework so a new integration doesn't mean starting from scratch each time?

Congrats on the launch. One question tough, does this integrate with any note taking apps or can you save the explanations you got in some way? I think saving this for later review would be really useful for studying.

Congrats, Harshit! Fellow extension builder here — quick one on “your data stays yours”: since scans save to a workspace and diagrams go through a vision model, is highlighted content sent per-request only (not used for training), and can I delete a scan? That’s what decides whether I’d keep an in-page AI tool always-on.

honestly this solves something that's bugged me for ages. one thing that would make it even better is keyboard shortcuts to highlight and summon the popup without needing the mouse, basically so you stay in flow while reading long PDFs or docs. right now i imagine reaching for the trackpad breaks that focus you're trying to preserve.

Finally, something that doesn't make me bounce between tabs like a pinball. Highlighted a dense section of a research paper and got a clear answer right there without losing my place.

Highlighted a dense section of a medical textbook and the in-context answer actually clarified the terminology without me having to switch tabs. Took a little fiddling to enable it on a PDF but once it clicked, the workflow felt way smoother than my usual copy-paste routine.

been wanting something like this for a while honestly, the copy paste dance into chatgpt is so annoying. one thing though, it would be really nice if you could save snippets or highlighted passages into a little notebook inside the extension so i could come back to them later without losing track. like a quick personal knowledge base built up as i browse.

Would love to see a quick keyboard shortcut to toggle Kogvio on and off without losing my highlight. Right now I have to reach for the mouse every time, which kinda defeats the staying-in-flow part.

honestly the highlight-and-ask flow sounds really clean. one thing i'd want though, basically a keyboard shortcut to open the side panel so i'm not hunting for a small button every time. something like cmd+shift+k would make it feel instant and keep my hands on the keyboard.

the highlight-and-ask-in-context flow is exactly the kind of friction removal that actually gets used daily instead of abandoned after a week. for something like a long research paper, does it feed the whole document as context or just the surrounding section you highlighted? that tradeoff between accuracy and cost seems like the hard part to get right.

Harshit, what makes it different from the Comet or Chrome browser capabilities?

the copy-paste-back workflow is such a small tax but it adds up over a day of reading docs. how does it handle technical diagrams specifically, is it reading the image content too or mostly working off the surrounding text

Love that it works on PDFs, not just webpages. So many highlight tools ignore that, but textbooks and research papers are exactly where I get stuck and need context the most.

As someone who lives in research papers and medical textbooks all day, that copy → open ChatGPT → paste → lose-my-place loop is painfully familiar, so this clicks immediately. Two things I'm curious about: when I highlight a passage in a dense PDF, does Kogvio pull in the surrounding context (or the whole document), or only the exact selection? And for the technical-diagram case like the Kalman filter example in your demo, is it reading the figure itself or the caption and text around it? That's usually where in-page tools fall short for me.

The highlight-in-context flow on a PDF or docs page is the part I would actually keep installed — that copy, open ChatGPT, paste, come back loop is exactly the focus-killer you describe. Before trusting it on technical or medical material I would want to know the data boundary: when I highlight a selection, does it send just that selection to the model or grab the surrounding page/PDF context too, and is any of that page content retained server-side per query or is it stateless? For sensitive docs that is the difference between leaving it always-on and only enabling it on tabs I trust.

Hey everybody,

My name is Asaf, and I have been an entrepreneur since I was 17 years old (or since I was born, depending on how you look at it 😄).

After several startups and many failures and lessons along the way, I had my first exit at 27 as the CTO of a Fintech company I co-founded with two partners.

After the exit, I decided to change directions and go towards what I love - creating productivity tools and automations, and naturally, I found myself in the Mac app world.

Today, my company AppitStudio owns 4 macOS apps, serving more than 3,000 users worldwide, and once again, I ran into the same problem I had with my past projects… I can't figure out which distribution channels, platforms, or collaborations ACTUALLY work for my business…

I partnered with several content creators, including some big names with millions of followers and smaller ones with 10k to 500k followers…
The problem was that, except for the timing of the video's appearance, the statistics were completely unreliable.

The existing solutions that I found were too broad and aimed at enterprises.
From a general URL shortener without influencer tracking, to a big influencer agency management tool.
I didn't find a balanced tool that will provide me with both branded links and a good influencer management platform.

So I decided to solve that issue by creating a tool that will do exactly that.

Full link tracking, clicks, and statistics on the server side, not on the client side, where ad blockers filter data, with a full influencer management tool that allowed us to find the real cooperations that worked.

Meet LnkFlow - the solution that helps me, and is going to help you scale your business by actually understanding what brings in the money.

Using LnkFlow, I immediately understood what my money bought…

For example, big follower bases actually brought in a bunch of bot traffic, and of the 3 small content creators I collaborated with, one was responsible for 85% (!) of the clicks on my websites. I immediately stopped working with the other two and doubled down on the one that worked, and the results were amazing.

Of course, I had to make it fully agentic-ready. I am a big believer in automating manual processes.
This is especially important for solopreneurs and small teams, where time and focus are everything.

We already have a beta version of the platform that tracks the full cycle from click to sale. We are using it internally for our company and a few design partners, and I can say this is gold.

If you relate to the problem I had, check out LnkFlow. I decided to run a promotion with a big discount for early adopters, and that price will stay for every renewal, for a lifetime.

Congrats on the launch, Asaf! The bit about partnering with creators who had millions of followers and still not being able to trust the stats is such a real problem, feels like something you'd only build right after living through it yourself. Curious how the server-side tracking holds up against ad blockers on mobile specifically, since that's usually where the numbers get messiest. Excited to try this for ranking our own collabs.

Congrats on the launch! I love the idea of attribution in this space. What makes it "agentic" in practice — does it act on the data for you or surface recommendations?

Looks like a really clean setup, especially the server-side tracking part since ad blockers basically wreck most shorteners. One thing that would help me out: a way to set up expiry rules on links, like auto-redirect after a campaign ends or push old tagged links to a fallback URL. Would save a lot of cleanup work when collabs wrap up.

the tension I'd want answered before switching is that "counts what ad blockers hide" and "filters out bots by default" pull in opposite directions - one undercounts, the other risks overcounting if the bot filter is too conservative. how do you validate the bot-filtering, is it just user-agent/IP heuristics or something closer to behavioral scoring?

Congratulations on the launch! Counting on your own server before the redirect is a smart way past ad blockers, but it's also where bots and the link-preview crawlers from places like LinkedIn and Facebook tend to pad the count. The launch says bot filtering is on by default; how do you separate real clicks from those? That's usually the hard part with server-side counting.

the bot filtering baked in by default is a smart call. so many short link tools still treat every click as a person, which makes the reporting basically useless for anyone paying creators.

Congrats on the launch, Asaf. The AI-assistant flow for creating links without touching the dashboard is a nice shortcut - when it authenticates, does it get scoped access to just that user's own links and campaigns, or a broader key that covers the whole workspace?

honestly this looks solid, the bot filtering by default is huge. one thing though, would love to see a way to set up custom conversion events beyond just clicks, like signups or purchases tied back to each link. that way you could actually see which creators are driving real revenue not just traffic.

Love the idea of filtering bots automatically, that alone saves so much headache. One thing I'd really want though is a simple way to A/B test two destination URLs under the same short link so I can quickly see which landing page converts better without spinning up a whole separate campaign.

Finally a link tool that doesn't get fooled by ad blockers, and the bot filter caught what looked like a suspicious spike from one creator. The AI pulling reports in plain language saved me from opening the dashboard at 11pm. Solid update.

Finally a short link tool that doesn't lie to me about clicks. Tested it on a few collab links and the bot filtering actually caught stuff my usual tracker missed, which was a nice surprise.

love that the bot filtering is on by default - feels like you actually thought about the real headaches people run into instead of just shipping a pretty link shortener. the agentic piece is pretty cool too

congrats on the launch! 🚀 most attribution tools tell you where clicks came from, not where revenue came from. tying campaigns and creators back to actual money is the part everyone gets wrong - good to see it front and center!

click tracking that shows what actually grows the business implies some kind of causal attribution, not just raw click counts. is that based on downstream conversion data you connect separately, or is it inferring growth signal from click patterns alone? those are very different claims and I want to know which one this is before trusting the dashboard.

Having built self-hosted MCP agents, governance is always the bottleneck: defining unattended vs. human-in-the-loop actions and auditing the aftermath. You can ignore this solo, but it kills enterprise adoption.

The "allow reads, approve writes" approach with unified audit logs targets this perfectly, though I'm curious if it survives a real security review.

For anyone running agents in production: where do you draw the line for human intervention? That’s the boundary I find hardest to define.

I've had access to Lunen for a few weeks and have an agent that runs daily that is saving me hours of work per week, and will only continue to get better as I refine and have an agent that can build these agents for me.

The governance view is genuinely useful, I could actually see what ran and approve the risky steps before anything shipped. Wish more AI tools handled the boring compliance stuff this smoothly.

The thing that stands out to me is keeping a clear record of what the AI actually did. Wanting the convenience without giving up the ability to answer for it later feels like the sensible way to approach all of this.

Finally a tool that keeps the consumer feel but actually shows you what happened under the hood. The approval flow for risky steps felt thoughtful, not bureaucratic.

The control surface matters as much as the agent. For team-run agents, I would want permissions, dry-run mode, approval points, and logs to be first-class, because the failure mode is not that the agent is slow; it is that it acts confidently in the wrong system.

Clean interface that doesn't feel like another enterprise dashboard, and I was surprised the approval workflow didn't slow things down.

Would love to see a side-by-side replay of the agent's exact tool calls with diffs against what it intended to do, makes post-hoc review way easier when something feels off.

The "2FA checkpoint" framing for risky calls is the interesting part to me. Most agent tools either gate everything behind a human click or trust the model fully — you're doing something in between. How is the read vs. write / auto-allow vs. approve line actually drawn? Is that a threshold an admin configures per tool/data source, or does Lunen infer it from the action type itself? And for scheduled agents running unattended — if a write needs approval and nobody's there to click it, does it just sit and wait, or fall back to a pre-set policy?

the framing really nails it, putting governance on equal footing with experience instead of bolting it on after. that takes discipline most teams skip.

The approve-the-risky-step checkpoint is exactly what's missing from most agent tools — they fire and you reconstruct what happened afterward. Before wiring this into a real workflow I'd want to know where the action record actually lives: a hosted store only, or something I can export or self-host to hand the security team? And is the approval gate rule-based per action type, or does a human get pinged on every write?

Congrats on the launch, Mike. The control layer is the part most agent builders skip, and it is the first thing a regulated buyer asks about. I run a HIPAA-compliant AI suite for clinicians, and "who can run what, with which data" is half of every security review we go through. How granular do the permissions get, can an admin scope which data sources an agent touches per team member? And can the audit log be exported to hand to a compliance team during a review?

One thing that would help my team a lot is a side-by-side replay viewer that shows what the agent did step by step with diffs, so I can review an overnight run in a few minutes instead of digging through logs. That kind of timeline with redo and undo per step would make the whole governance story actually usable day to day.

Finally something that doesn't make me choose between a nice interface and actual visibility. Tried a small workflow and the audit trail was already there without me digging for it.

Finally tried it this morning and the approval log before risky steps actually feels useful, not just busywork. The fact that I can see exactly what the agent touched makes me way more comfortable letting it loose on real files.

Finally got access this week and the activity log is the standout for me, I can actually see every file and step it took without digging through chat scrollback. The approval prompt for risky actions feels like the right amount of friction too.

Great launch! If CaptureKit routes the file straight into a folder, does that markup window still come up? Does it become save-then-reopen-in-Preview? Arrows and boxes are half of why a screenshot is worth sending I guess!

My desktop is basically a screenshot graveyard right now, so this is an instant "yes" for me. One thing I'm curious about: does CaptureKit intercept the native macOS shortcuts (Cmd+Shift+3/4/5), or do you have to trigger the capture through CaptureKit's own menu/hotkey for the auto-save-to-folder behavior to kick in? Asking because macOS doesn't make it easy to hook the system screenshot flow — if it's the former that's a genuinely hard problem you've solved.

The auto-route-to-a-chosen-folder instead of dumping on the desktop is the exact friction I hit daily. Curious about the routing model: is it one global destination, or can I set per-project rules so shots taken while a given app or Space is focused land in different folders? And does it hook the capture itself or just watch the default screenshot directory via FSEvents and move files after?

This is a nice Mac-native utility surface because the value is in staying out of the way. The details I would care about most are predictable rules, quick undo, and keeping originals easy to find, so organization never feels like another inbox to manage.

clean idea, my desktop is basically a screenshot graveyard right now. how does it decide which folder a screenshot goes to, is it just one default folder or can you set rules per app/project?

I take screenshots daily. This app seems very helpful!

Finally, a screenshot tool that just gets out of the way. Love that it saves straight to a project folder instead of piling up on my desktop every time I hit cmd+shift+4.

Would love to see an option to auto rename screenshots with a pattern like project-date-number, makes it way easier to find the right one later instead of scrolling through "Screenshot 2024-05-23 at 3.47.11 PM" files.

Finally something that gets the folder routing right out of the box without needing another Hazel rule. Love how clean the menu bar icon stays, no extra chrome cluttering up the workspace.

This is such an ongoing frustration with macOS and quite crazy this is not natively a better experience on Macbook. Biggest frustration is naming and labeling. Almost all screenshots I take are in some way 'bookmarks' of something I want to read at a later stage. I often forget about them. Auto-prioritization and context related labeling would be a big win.

My desktop is a graveyard of "Screenshot 2026-07-..." files, so this one hits. Can I route screenshots by rule — e.g. by active app or window title — or is it one destination folder? Per-project folders that switch based on the frontmost app would be the killer feature for me.

would love to see an option to auto rename screenshots based on the active app or window title, like "Slack - design-feedback.png" instead of just "Screenshot 2024-01-15 at 4.32.18 PM". would make sorting through project files way easier honestly

Congrats on shipping. One thing worth knowing before you build the hotkey, and I say this as someone who would happily use the version of this that goes further.

macOS already does the base case natively. Cmd+Shift+5, Options, Save to, Other Location, and every screenshot from then on lands in the folder you pick. There is a defaults write for the same setting. So for the exact job on the landing page, saving to a chosen folder instead of the desktop, the system has shipped that since Mojave, and per your reply to Tarandeep, CaptureKit currently only applies when someone uses your menu rather than the shortcut already in their muscle memory. That is a hard sell against something built in and free.

The good news is that the interesting half is the half macOS cannot do, and it is already sitting in this thread. Gal asked about rules per app or project. Uğur asked for pattern based renaming instead of the default timestamp filenames. Neither is possible natively at any price. A screenshot that lands in the right project folder because that project was focused when you took it, named something you can actually search six months later, is a real product. A different default folder is a settings pane.

If it were mine I would park the hotkey and build the routing rule first. The hotkey only gets you to parity with the system. Routing is the part nobody has done.

The local processing is the part worth paying attention to. A lot of transcription tools quietly send your audio to a server, which matters alot!

The drum transcription accuracy is impressive for a local tool. One thing that would really round it out is adding a confidence indicator on each detected hit so you know where to focus your manual corrections first, maybe color coding notes the algorithm is less sure about.

The transcription side looks solid, and running it locally is a big plus. One thing I'd love to see is a way to nudge the detected tempo before it lays out the grid, since slight drift in the source clip often throws off the whole chart and forces a lot of manual cleanup.

as someone who's tried a few "audio to sheet music" tools before, drums specifically are the part everyone gets wrong because hi-hat vs ride vs crash bleed into each other in a busy mix. does it do okay on a fast double-kick section or does that turn into a mess you end up hand-correcting the whole bar anyway?

Hannes, transcribing by ear is the part every drummer I know quietly dreads, so seeing this exist makes me happy. The fact that it all stays on my own machine is a real comfort too.

Local is a good fit for this job. Drum stems and rough mixes are exactly the kind of material people should not have to upload just to get a starting chart. I like that the score stays editable too - AI should remove the blank page, not trap the musician in its first guess.

Dropped in a stem and the transcription came out close enough that I only had to nudge a few hits by hand, which is way better than I expected. The PDF export looks clean and ready to print for a real chart.

Dropped in a drum stem and was honestly surprised how clean the notation came out, even with some messy ghost notes. Being able to hand-fix the chart and export straight to MIDI without anything leaving my machine is a really nice touch.

Dropped a drum stem in and was surprised how cleanly it lined up to the staff, fixing a few ghosts by hand took seconds. The local-only approach is a nice touch for anyone wary of uploading sessions.

Hand-correcting beats on a five-line staff feels way more natural than clicking grid cells, and the GM export is a nice touch for sending charts to a real kit. Glad it all stays local too.

Dropped in a rough drum stem and it picked up the ghost notes on the snare way better than I expected, the editable score is genuinely handy for checking fills. Exporting as MIDI instead of just PDF is a nice touch for reworking parts later.

the drum score preview alongside the playback is such a thoughtful touch, makes checking hits actually feel doable instead of guessing. and running it all locally is exactly the right call for this kind of tool.

This is cool, did you make it out of your own need or what inspired you?

This is such a great, unglamorous problem to pick. Everyone does melody transcription and quietly skips drums because they're the messy case, so going straight at it is a nice move. All the best with it!

This sounds too good to be true! As a drummer myself, I know how time-consuming it can be to transcribe parts by ear, so turning drum audio into an editable score feels incredibly useful! Thanks so much for making this.

The correct-by-hand-then-export flow is the actual selling point for me — most audio-to-score tools trap the result in their own editor. When I hand-fix a hit — say move a stray note from the hi-hat lane to the ride, or delete a ghost snare — does that edit carry into the General MIDI export as the right instrument on the correct GM drum note, or does export re-run detection and drop my corrections? And does the printable PDF keep the same five-line kit-piece lane mapping I see on screen?

The GM MIDI export is what sells me here — an editable staff you can actually pull back into a DAW beats a static PDF every time. As a producer I live in swung and triplet-feel grooves (trap hats, shuffle) — does it transcribe those to real triplet subdivisions or snap everything to a straight grid? That's usually where audio-to-score tools fall apart for me.

I have some friends that are drummers. What's been the problem in getting the notation right?

The hero we needed, loved the fact you made this without creating another account or installing other apps. Have you noticed people coming back and using the tools ?

Tried it with the JWT decoder — loaded a sample token, decoded instantly, header + claims right there, no login. Clean. The one thing I’d love spelled out: you say “supported files stay in this browser,” so which of the 35 are fully client-side vs the ones that do a server round-trip (PDF compressor, image stuff)? Even a small “runs locally” badge per tool would make the no-upload promise obvious at a glance. Congrats on the launch!

the no-account-needed approach is honestly so refreshing, you know how rare that is for a hub of this many tools in one place

Solid set of no-signup utilities, the image resizer was snappy and handled the batch I threw at it without a hitch. Wish the JSON formatter had a tree view but for quick tasks it does the job well.

No-login is underrated — a lot of "quick task" tools add too much friction with signup walls. What made you decide to keep it fully login-free rather than adding optional accounts for saving history?

No login is the right call for utility tools — nobody wants to create an account just to resize an image.

The no-account-required approach is refreshing, and I love how cleanly the JSON formatter handles nested structures without any clutter. Great execution on something that feels simple but is genuinely useful.

Handy little toolbox for quick browser tasks, no login fuss. The JSON formatter is what surprised me, clean output and surprisingly fast for a free tool.

would love a batch upload option for the image resizer, having to resize 30 product photos one by one is painful. drop in a folder, pick dimensions once, and let it churn through everything would save so much time.

Love that everything runs right in the browser with zero signups. The all-in-one layout for image, JSON, and PDF tools feels really thoughtful.

honestly the JSON formatter was way faster than the one i usually use, no login popups or anything getting in the way. solid little toolbox honestly

Resized a batch of product photos straight in my browser without signing up or seeing a paywall, which honestly caught me off guard. Also tried the JSON formatter on a messy API response and it cleaned things up nicely. Pretty handy bookmark.

A bulk upload option for resizing multiple images at once would save a ton of time, especially when working with large batches of product photos or social media assets. Right now it seems like one at a time only.

order-to-cash is a good pick for agents specifically because the steps are so repetitive and rule-bound most days, and the exceptions are usually well-defined too. since this touches real invoices and payments, does finance get a full audit log of every action the agent took, or just the summary result? that matters a lot when someone has to explain a discrepancy to an auditor later.

One thing I'd love to see is a built-in simulation mode where you can run the agents against historical invoice data to predict how many exceptions they'd catch before going live. Would make the ROI pitch much easier when rolling this out to finance leadership.

Love that Rex tackles the whole order-to-cash loop instead of just one slice of it. One thing that would help us evaluate it faster: a sandbox mode with pre-loaded sample invoices and dispute scenarios so we could watch the agents work through edge cases without touching real customer data.

the way rex tackles exceptions before cash actually gets stuck shows really thoughtful sequencing, like the team actually understood how messy ar workflows get in practice

Took Rex for a spin on a messy test account and the exception handling actually caught a duplicate invoice flag our team missed. Pretty solid for enterprise scale.

Order-to-cash is exactly where agent work needs boring reliability. Idempotency, audit trails, and clear exception handling matter more than a flashy demo, because a duplicate invoice or missed handoff turns automation into operational debt very quickly.

One thing that would help us a lot is a built-in dashboard showing which exceptions Rex resolved vs. flagged for human review, so we can track accuracy over time and identify any patterns in the cases it escalates.

The exception-handling angle caught my eye, since most AR tools just chase dunning emails. Curious how the agents handle messy real-world disputes vs clean test cases.

Looking crisp, congrats on the launch!

you asked for it, so: what's the actual failure mode when an agent gets a judgment call wrong on a real invoice, does the customer just get an incorrect collections notice before a human catches it, or is there a hold step before anything customer-facing goes out

Finally gave Rex a spin with some messy test invoices and it actually flagged the FX mismatch before the cash hit the bottleneck. Slick setup if you deal with cross-border invoicing headaches.

Congrats on the launch @merlin_kafka! The Order-to-Cash process is notoriously tricky because writing incorrect data back into an ERP like NetSuite can wreck month-end close or mess up revenue recognition schedules.

When Rex agents handle exception resolution (like applying partial cash or mapping unmatched remittances), does the system maintain an immutable, step-by-step audit trail of the agent's reasoning right inside the ERP/CRM, or does it operate via a separate proxy logging layer for finance auditors?"

Order-to-cash is a great wedge for agents — high-volume, judgment-heavy, and the "copy from portal → NetSuite every single day" drudgery you described is exactly where automation historically fell over. Taking you up on the "ask what they get wrong" invite @merlin_kafka:

How do the agents decide the boundary between acting and routing to a human — a confidence threshold, a fixed policy per action type, or something learned?

On portal uploads specifically — I do a lot of browser automation against sites that change without warning, and that's usually where things get brittle. How do you keep those flows stable across hundreds of customer portals?

And since a wrong AR action touches real money, how do you eval that an agent did the right thing before it goes out?

Honestly the 3D rotation on big correlation networks is super smooth, no lag at all on my m1. Felt kind of like having origin but actually fast and local, which is basically what i always wanted

as someone who's bounced between a notebook for the analysis and a separate app for publication-quality figures, the vector export at the end of the pipeline is the part that stands out to me. what format does that export to - straight SVG/PDF, or does it keep layers editable in something like Illustrator? and does "raw data never leaves your Mac" mean fits run fully on-device too, or just the raw import step?

Unifying analysis and plotting in one place is a real pain point — too many tools force you to jump between a notebook and a separate charting tool. Does it support live data sources, or is it mainly for static datasets right now?

I lose whole afternoons to bouncing between spreadsheets and notebooks and image editors, so the idea of it all living in one calm place lands well. Keeping the raw data on my own machine makes it feel even more trustworthy. Thanks Lale!

Would love a built-in scripting pane so I can automate repetitive fits and re-run an analysis when the underlying CSV gets updated. Right now I have to re-click through the same steps every time the dataset refreshes.

the Metal-backed 3D plots are genuinely smooth, way better than fighting with matplotlib on a big dataset. kind of nice that everything stays local too

Would love a Jupyter notebook style export so I can pair each plot with the exact code or transformation steps that produced it. Also makes it way easier to share reproducible figures with collaborators who don't have AutoPlot.

The reproducible-Python export is the right call, because with AI-written analysis the plot that renders cleanly is the one you stop checking. Every bad LLM-generated analysis we hit was silent, never a crash: NaN rows quietly dropped before a mean, or a default bin width smearing a real bimodal peak into one hump. The chart looks fine so nobody re-reads the code. Do you surface the assumptions the model made, rows dropped, null handling, aggregation choice, somewhere near the figure itself, so a wrong-but-pretty plot is catchable without opening the export every time?

the cc-into-inbox approach is genuinely clever, feels way more natural than opening yet another app to delegate. curious how well it picks up on tone and priority cues over time.

The "own inbox" framing is smart — most AI assistants just react to what you send them. Does Deck proactively flag things in that inbox, or is it purely on-demand when you check in?

The 'its own inbox' design is what separates this from a chat assistant — CCing it on a thread and having it take over is an actual workflow, not a demo. Two setup questions before I would hand it recurring work: when it signs up for services or gets added to threads, is it acting under its own inbox identity so those accounts belong to Deck rather than me, or under my credentials? And the preferences and priorities it learns over time — is that context portable if I move to a new inbox or workspace, or is it pinned to the original address?

CC-ing an AI assistant on a thread and having it just take over is a genuinely different interaction model than opening a chat window, that part is clever. what happens when it takes an action based on a misread of an ambiguous email though, does it check with you before actually replying or scheduling something on your behalf, or does it just act and let you catch mistakes after the fact?

Been waiting for something like this, the idea of just CCing an agent into a thread and walking away is genuinely useful. One thing that would make me trust it faster is a clear rollback trail, so I can see exactly what it sent on my behalf and undo or edit a message within a short window after it goes out.

Cc'd it on a messy vendor thread and it actually pulled the right context back without me re-explaining anything. The morning briefing landing on its own was the part that sold me.

Congrats on the launch, Stasia. Giving Deck its own inbox instead of building on top of your real one is a cleaner boundary than most agent-inbox tools start with. When Deck signs up for a service on your behalf, does it hold that service's login itself, or hand off to a one-time OAuth flow for that step? That's the design choice I'm most curious about. Skipping blanket inbox authorization by default is a good call — most tools ask for way more access than they need on day one.

"most capable"... as in? Are there any benchmarks published? :D

Handed Deck a messy client thread to summarize and it pulled out the action items plus flagged two follow-ups I'd have missed. Liked that it just sat in my inbox waiting instead of demanding a chat window.

The "give it its own inbox instead of blanket access to mine" call is the right instinct. An agent with its own address that only emails me is a far cleaner trust model than yet another tool asking to read my whole mailbox, and leaning on email as the delegation surface is smart since that behavior already exists.

Where my mind goes is the flip side: acting on anything it's CC'd on means inbound email becomes the attack surface. Wondering how Deck tells a real instruction from a spoofed or injected one. If someone emails its address pretending to be me or my boss, does it act with the same authority, or is there a trust boundary around who's actually allowed to make it do things?

This is a fascinating angle for video creators — most analytics tools tell you what happened (views, drop-off points) but not why. Curious what's actually being measured here: is this modeling predicted neural response from visual/audio features, or do you have real EEG/biometric data feeding the model? That distinction matters a lot for how much I'd trust the output as a creator deciding what to cut.

saw your reply to Aidan below - appreciate the direct answer that it's a forward model against published neuroimaging weights, not fitted to real retention data. genuine follow-up: if it isn't validated against actual audience retention, what's the evidence that the engagement score correlates with what real viewers do, versus just being internally consistent with itself? asking because "stimulus-driven, not viewer-specific" is honest but it also means the score could be confidently wrong in a way no refund policy would catch.

The audio-aware angle is genuinely interesting, since so many tools ignore the sound side. One thing that would help me decide if it's worth using long-term though is some kind of comparison view, so I can drop in two versions of the same clip and see how the engagement scores and emotion timelines actually differ. That would make the suggestions feel way more actionable instead of just a score in a vacuum.

I like that the product is explicit about what it measures and, just as importantly, what it doesn't claim to predict. Tools that acknowledge uncertainty tend to be much more useful than ones that always return a confident score.

One thing I'd love is the ability to compare two versions of the same clip side by side, so I can see how a re-edit changes the engagement score and emotion timeline. Would make A/B testing hooks and thumbnails way easier.

Honestly the audio piece is what sold me here, finally something that gets that podcasts and music-driven clips need attention too. One thing I'd love: a way to compare two versions of the same clip side by side so I can see which edit is actually pulling more engagement second by second. Would make the refund policy almost unnecessary.

The refund on low confidence reads is a smart call. Most analytics tools will give you a number no matter what, even when the input is too messy to trust.

Honestly the audio part is what sold me, most engagement tools totally ignore that. One thing though, would be cool if you could compare two versions side by side before exporting, like drop in a re-cut and see how the emotion timeline and score shift. That would make the refund policy almost irrelevant because people would just keep tweaking until the numbers move in the right direction.

Is it that Meta model behind this product or what are you using?

The thing I'd want to understand is what the prediction is validated against. "How a brain would respond" can mean a model fit to EEG data, or one fit to actual retention curves from real viewers, and those two behave very differently once a clip looks unlike whatever it was trained on.

Very cool - congrats on the launch.

uploaded a 30 second podcast clip and the per-second emotion timeline actually picked up on the pause before the punchline, which i didn't expect from a tool like this. the refund promise is a nice touch too.