PH热榜 | 2026-07-29

Hi Product Hunt, I’m Wojtek, one of the founders of Prelint.

Coding agents now produce more code than most teams can properly review. The obvious risk is bad code, but the more dangerous risk is good code that quietly builds the wrong product.

We saw this repeatedly in large production projects when we started using agents. A change would clear technical review, pass CI, yet still skip the transactional outbox pattern, introduce an unapproved dependency, change a permission rule or invent a business requirement. Nobody had made that decision, but it was now part of the product.

Prelint catches these decisions before they ship. It reads each change alongside your specifications, tickets and existing product context, then explains what the agent decided, what the consequences are and how difficult the choice will be to reverse. Your team can approve, correct or replace the decision, and Prelint carries that context into the next piece of work.

If you work with ADRs, our clients say it is the best tool they found for enforcing them with agents. On teams running Prelint alongside other AI reviewers, ~40% of the review comments that actually get fixed come from Prelint.

We started in GitHub and now support CLI and MCP, allowing agents to check product decisions while they work instead of waiting for a human to discover the problem at the end.

Prelint is not another technical code reviewer. These tell you whether the code works. Prelint tells you whether you should be building it.

If coding agents are contributing to your product, sign up at https://prelint.com, connect a GitHub repo and see what they've been deciding on your behalf. Then tell us what you like and what you'd love to see on top of it!

Use code: PH100 to get 100$ in additional free credits

My honest question: how much setup does this need before it's useful? A lot of "review your PRs against your docs" tools sound great until you realize the doc structure has to be near-perfect first.

Multiple AI reviewers catching different things makes sense, no single model sees everything a fast-moving codebase actually needs.

The stat about catching 40% of pre-merge issues on teams already running multiple AI reviewers is the part I'd want to see broken down further. What kinds of issues specifically? Style violations are cheap to catch. Actual drift from architectural intent is a completely different, much harder category and I'd bet that's where the real value is.

I'd be curious how noisy this gets in practice. Docs and ADRs drift out of date constantly, how do you keep the source of truth from becoming stale itself?

The 40% number is the kind of stat I want to stress-test rather than just admire. Is that measured across teams with mature, well-maintained ADRs or does it hold up even when documentation is patchy, which is the more common situation? I ask because a lot of tooling in this space performs great in the demo and struggles the moment real docs are inconsistent or out of date. Would genuinely like to see that breakdown.

Congrats on the launch! I’ve run into this exact issue with AI agents: the code works, but it still builds the wrong thing. What kind of specs work best with Prelint?

I like that Prelint checks against ADRs specifically. Most reviewers just look at the diff, not whether it fits the actual architecture.

Really interesting approach. How do you handle situations where multiple product specs or architecture decisions conflict with each other? Is there a way for Prelint to detect outdated or contradictory documentation before reviewing a PR? That could be incredibly valuable for larger teams.

Few weeks in with Prelint and it's been a genuinely good experience.

Best part for me is that it notices when a change goes against what the docs already say, which means you deal with it early instead of in a painful review later. Solid tool.

Nice to finally see you on PH, good luck!

I've watched AI-generated PRs balloon in volume while nobody's actually re-reading the original design docs. Tying review criteria back to ADRs feels like the missing piece most teams skip.

QA engineer here, been using Prelint for a while. A lot of our PRs are AI-assisted now code compiles, tests pass, looks right, but whether it still matches what the team actually decided used to land entirely on me.

Prelint does that pass before I get there. It catches spec drift early, and the notes it leaves on the PR are genuinely useful input for my testing, I go in knowing which areas need attention instead of reading a diff cold.

Good luck with the launch!

Prelint hitting "product drift" specifically is such a sharp framing. Most AI code tools obsess over syntax or bugs, but the thing that actually scares me is the AI quietly building something slightly off from what the product was supposed to be

The decisions I never see in an ADR are the ones that were made to a customer.

Support creates them constantly. Someone asks when the renewal reminder goes out, support answers before the charge, and that is now a promise a few hundred people are holding you to. It lives in a support thread and nowhere an agent can read.

A PR that moves that email passes every doc in the repo and still contradicts what those people were told. The expensive kind too, because the customers already know the old answer.

Does your ingestion pull anything from the support side, or is ground truth engineering artifacts only?

My team's biggest AI-coding headache isn't bugs, it's silent scope creep, code that works but drifts from what we actually decided months ago. Tools like this feel overdue.

I tried catching this manually with a "does this match my architecture doc" checklist in code review. It worked until velocity picked up and reviewers started rubber-stamping. Would love to know how it handles reviewer fatigue over time.

Technically correct code that builds the wrong product describes AI-assisted development pretty well. "Code became cheap" mantra doesn't mean we ship better products; it means the bar for what's worth building dropped. In practice this results in more of what users never asked for, faster.

You say "the agent can self-correct before a human ever looks" - what does that look in practice? Is it an automated loop, or does it require human involvement?

The code reviews are the real bottleneck now, good to see a product that tries to deal with it!

I've been using Prelint for a few weeks now, and it's been a great experience so far. Solid tool, cool to see you finally on PH. Good luck with the launch!

Hey, I've been using Prelint for the last couple months - I really like that it reviews the product reasoning behind a PR. And it's smart about scaling effort - most small, low-stakes PRs just get a quick approval, while it goes deep on ones that actually touch permissions or auth, even if the diff is tiny. It can be a bit too nitpicky at times, but that's a much better failure mode than rubber-stamping things through.

Quick question - do you charge for one-line PRs, or are those free/exempt from billing?

Congrats on the launch - rooting for you!

We were using similar thing in my project but built in-house, that's actually super valuable

AI-generated code will only become more common, so governance tools like Prelint are going to be must-have!

Prelint is interesting because it seems to be focusing on the product intent. A lot of times AI can take its own direction, catching it soon from drifting by a tool is a great idea indeed!!

Congrats on the launch guys! What kind of integrations do you guys have?

Congrats on the launch! "Product drift" is a great name for a problem I didn't have a word for. The code compiles, tests pass, and it still quietly isn't the thing you asked for. Catching that at lint time is the right layer. Does it work across any codebase, or are you starting with specific languages?

Congrats on shipping this! Product drift is such a sneaky problem, the code works fine but the product quietly became something else. Does Prelint only catch drift in new pull requests, or can it also scan an existing codebase for decisions that already slipped through?

The "decision ledger, not a code reviewer" framing is the part I'd have paid for a year ago.

One gap I'd want to understand. Every ingestion source you've described is an artefact of a team — Slack, meetings, tickets, ADRs. I run a one-person company and I have none of those. Nine months of architectural decisions exist only in AI chat transcripts and in commit messages I wrote to myself. There's no PR either; changes go straight to main.

So: is the CLI a first-class path for the no-PR case, or a fallback for the GitHub flow? And is there any way to seed the ledger from conversation history rather than from repo artefacts? The decisions I most want caught are the ones I made at 11pm in a chat window and had forgotten I'd made by the following week.

Congrats Wojtek and team on the launch!

We've had both human and AI PRs pass every check and upon prod review we end up reversing old decisions, excited to try this out.

Congrats on the launch. The line that got me is "good code that quietly builds the wrong product" — that's a failure I didn't see coming until it bit us. We build AI products too, and the agent passes every test and still changes a rule nobody agreed on. One question: how does Prelint know what the "right" product was meant to be — only from the ADRs and docs you feed it, or does it also learn from what your team approves over time?

Looks awesome!

Hey Product Hunt! 👋

We're the team behind SoundGate Guitar - a practice companion that actually hears you play and responds in real time.

Most guitar apps follow the same formula: watch a video, try to play along, get a pass/fail score. That's not how you learn guitar. You learn by playing, making mistakes, and getting feedback in the moment - the way a real teacher would.

So we built SoundGate to be that practice companion:

It hears every note. Play any note or chord and the interactive fretboard lights up instantly - real-time detection, not a 2-second delay. You can actually see what you're playing as you play it.

Interactive fretboard, tabs, scales, exercises - all interactive. It's not just detection. There's a full exercise library with scoring based on your actual playing. Chords, strumming, fingerpicking, scales - from beginner to advanced.

Ask it anything. Stuck on a barre chord? Wondering what scale works over a progression? The built-in AI coach gives you specific, personalized answers - not generic YouTube links.

Real native app. SwiftUI on iPhone, iPad, and Mac. Plug in a USB audio interface on your Mac and it just works. No Electron, no web views.

It's 100% free. Not freemium-free. Actually free. Every feature unlocked. We're in early access - we want feedback from real guitarists, not credit card numbers.

Whether you picked up a guitar yesterday or you've been playing for 20 years, we built this for you.

Would love to hear what you think - drop any questions below! 🎸

Music might be the most beautiful language we have...

Time to switch from code to strings for a while 🎸!

Congrats 👏 @aren_barseghyan on taking this live... solving real-time audio detection without delay is no small feat, Rooting for you guys today, exactly the motivation I needed to dust off my guitar this weekend.

Congrats on the launch! 👏 Just dropping by to say hi as a fellow guitarist.

I’ve tried a few similar products before (e.g. Yousician), so I’m curious what makes SoundGate stand out in such a competitive space.

In particular, is your differentiation around areas like:

• Polyphonic (multiple-note) detection?

• Expression/technique detection (e.g. bends, vibrato, slides, palm muting)?

Those have always seemed like some of the hardest technical challenges for AI guitar coaches, so I’d love to hear what you’ve improved.

Find it pretty useful. At some point in my life, I wanna learn to play any instrument. And if AI can teach me, this could be my option to try it. Do you also have any background in the music industry that you decided to create this?

Does the app support 4-string bass or ukulele, or is the engine currently tuned strictly for 6-string guitar? Congrats on the launch.. after seeking a product like SoundGate for a long time that's truly based on music, this is really great @aren_barseghyan 🙌

congrats on the launch! this looks like the perfect excuse for me to finally start learning guitar :)

are you planning to add support for more musical instruments?

Congrats on the launch!! The fact that it hears you play and responds in realtime is the cherry on top, I love how it recognises the note/chord you've just played and instantly shows on screen. I see it can also be your 'step-by-step- companion, are we talking beginner beginner? Let's say I don't know how to play guitar whatsoever - can I start today?

Congrats on the launch! You can tell a lot of thought went into making the experience approachable. It feels like the kind of product that could get more people to actually enjoy the learning process instead of giving up after a week.

Real-time feedback is the part that hooks me here. Most guitar apps just grade you after the fact, so catching a bum note the instant I play it feels way more like an actual teacher sitting next to me.

Congrats on the launch!

how cool! I remember how expensive music lessons were, so it's special to make that accessible for any audience

Real-time feedback is the one part of learning guitar you can't get without a teacher in the room. Everything else is already on YouTube. Curious whether it's catching timing and cleanliness or mostly pitch, since that distinction is basically the whole product.

Congrats on launch!

The real-time fretboard feedback is the part I would actually use - finally something that listens instead of making me watch a video and guess if I am doing it right. The thing I would want to know before making it a daily habit: does it save any record of what I have practiced or how I have improved over time, or is each session fresh? Knowing I can look back at what I was struggling with two weeks ago would make the practice habit much easier to stick to.

the zero lag note detection feels genuinely useful, especially when working through scales. the AI tutor suggestions were surprisingly specific to what i was messing up.

Congratulations with the launch. Do you plan to add more instruments or you wanna go niche?

Real-time note and chord detection could make practice much more useful than getting a score afterward. How well does it handle background noise or playing through an amp?

Im thinking about buying a guitar after this app. lol hope i can learn it fast because i don;t have time to go to the music classes

after this, I'm getting my first guitar.

congrats on the launch <3

SoundGate looks like a game-changer for guitar practice. Real-time feedback, personalized exercises, and an AI tutor in one platform make learning feel much more focused and motivating.

Excited for the journey ahead and to see how the product continues to evolve! 🚀

This really looks great and is a way for anyone to really go in and start getting a guide to learn to play a guitar. The AI feedback is cool and can help with where a user is doing well or not and overall, how they can improve. I just wish that it could be on Android. Can't leave the Android ppl out now, LOL.

As generative AI becomes capable of creating music, do you see it replacing musicians, or can tools like SoundGate help more people become musicians?

the use case i'd actually want is playing along with a real song, not just isolated exercises. if a backing track or the original recording is playing through my speakers at the same time as my guitar, does the detection still isolate my playing cleanly, or does it need near silence in the room to work? that's the difference between a practice tool and something i'd actually use to learn songs i care about.

Hi, on the initial start page it asks what type of guitar you play. I have an electric and acoustic but i was only able to select one option. A multi option for some of us who play multiple instruments might be good. Great work so far.

Love the look of this! Learning scales from static fretboard diagrams has always been the part I struggle with. Having something interactive or that actually lights up as you play sounds like it would make it click so much faster. Excited to give this a go.

Sounds interesting. Are you planning to launch a uke tutor as well?

Love this idea! I've wanted to learn guitar for years but never knew where to start. Does the AI create a personalized learning path for complete beginners?

Hey Product Hunt 👋

Saverio here, founder of Denovo.

I spent five years at Meta building ML models, and I've been doing gen AI research since 2018. I served four years as the AI advisor to the Italian Prime Minister's Office.

I'm the only employee at Denovo — I built the entire product from scratch, solo.

Here's the part that actually matters: my previous startups all failed because of go-to-market. Not the tech — I could always build. Finding customers is what killed them. Every time. 💀

AI made building easy. Getting paid is still brutally hard. You vibe-coded your app... now what? 🤷

Denovo is me solving my own failure mode, for everyone else — the whole distance between an idea and a paying customer:

✅ Feedback first — before anything gets built, Denovo pressure-tests your idea (market, competition, pricing) and gives a clear green/yellow/red verdict. Some ideas deserve a "no" before you spend a cent.

🎨 Your brand, then your website — built for you, on-brand, no page builder. Want changes? Just chat.

💳 Stripe from day one — the moment you pick your domain, payments are wired in automatically with Stripe Connect. Revenue isn't an afterthought. It's the point.

🎯 Customers, found for you — Meta ads set up automatically, plus outreach matched to your ideal customer profile from a database of 100M verified leads. You wake up to work you didn't do.

🎧 Then you just ask — "How's business?" Denovo tells you what it shipped, where revenue stands, and what's next.

15,000+ businesses have been built here — real domains, real Stripe accounts, real customers. 🚀

For the PH community: 20% off yearly plans (including Build and Launch) with code PRODUCTHUNT20 🎁

What will you be building on Denovo?

I'll be here all day — ask me anything, including the hard questions. 🙌

— Saverio

Congrats on the launch Saverio. The honest bit about previous startups failing on go-to-market makes this more interesting than the usual “build anything with AI” pitch. Denovo is taking on a lot though. Idea validation, branding, websites, payments, ads, lead generation, reporting. I’d want to know where the system is genuinely making decisions and where it is mainly automating setup. The green, yellow, red verdict also caught my attention. Has Denovo ever told a founder not to build something, and what evidence does it use to make that call? That could be one of the most valuable parts if it is willing to be blunt.

Congrats on shipping! The gap between "I built a working thing over the weekend" and "someone paid me for it" is where most of these projects die. Nice to see someone aiming at that specific step. What does the first week with Denovo actually look like for a solo builder?

Impressive scope for a single builder, transition from ML researcher to building a full-stack solo product is no joke, massive respect, qq how are you handling domain warmups so the automated outreach emails don't instantly hit spam folders? Congrats on the launch @saverio_pulizzi3 🙌

Congrats Saverio! The framing about the distance between an idea and a paying customer rings true, so many vibe-coded apps die at distribution rather than at code. Does the validation step ever tell people plainly that an idea is not worth building, or does it always suggest a pivot?

As a solo founder with a good portion of this already setup (stripe, website, etc) how much can Denovo can be used a la carte vs everything together? Meaning can I utilize Denovo simply for growth?

Does Denovo work well for B2B SaaS, or have you seen ecommerce brands succeed with it too? 🤔

How is this different from the 100000000 other things that do that same thing.

Congrats on shipping. What are the other GTM options you plan to add?

Some feature ideas:
- LinkedIn outreach
- SaaS directory submission
- Social media marketing

Denova = Vibe coding app + AI SDR + AI Performance Marketer. Did I get it right?

Love the positioning, particularly because building has become easy, but distribution is still where most founders struggle. Congrats on the launch! 

This looks like @Nas.com for startups. We do it for creators ;)

Congratulations. This feels like the logical next step after vibe coding. Build, launch, acquire customers. Repeat. Good stuff. 😊

How is the outreach personalized? Are the emails generated uniquely for every prospect or based on templates??

Having Stripe added to the product from Day 1 is amazing. Too many makers / builders leave it for later.

How accurate the market validation is before building? Have you seen founders completely change direction because of Denovo's feedback?

The gap between a working prototype and someone actually paying is where most of these projects die. I'm curious whether it touches pricing and packaging or just the checkout plumbing. When you say "paying customers," does it help figure out what to charge, or only collect it once you've decided?

@Denovo - this looks great, but what if I already have an app and website - how does this work, do you have plans just for marketing and growth.

The gap between building the thing and getting someone to pay for it is where most vibe-coded projects die, so that's the right target. The autonomous framing is ambitious though. What does the first paying customer actually look like, and how much is the founder still in the loop?

I work with a lot of startups in the education space and something like this can totally get them to market faster, nice job!

solo-building this whole surface is wild, congrats. the part I'd want to understand before handing over Stripe and ad spend: when the growth engine sets up a Meta ads campaign on its own, does it launch and start charging the card right away, or does it wait for the founder to see the campaign once before it goes live? "you wake up to work you didn't do" is a great pitch for the email side, but ad spend is real money leaving an account unattended, that's a different risk tier than a cold email that just gets ignored if it's off target.

Congratulations on the second launch. I noticed the acquisition side runs on cold email. The hard part there is choosing who to email, since a small, precise list usually beats volume. How do the agents decide who the right people are?

Congrats on shipping. The Stripe piece is the part I keep coming back to — wiring Connect the moment someone picks a domain means onboarding a merchant account per business, and platform review is usually what slows that down. Did you have to get Denovo approved as a Connect platform before any of those 15,000 could take a payment, or does each founder end up with their own standalone account?

I'm pretty new to Product Hunt but I have been building with Denovo for 2 months now and can say I absolutely love it and all the upgrades I've seen added in that time. I can also say that Denovo will definitely tell you if you have a bad idea or three. I was searching and trying other programs for about a year when I found Denovo. It really does work as advertised. Saverio really does take user suggestions to heart also.

the attention queue is the part I'd want to understand before trusting this with anything real. what decides something belongs there - a fixed list of trigger conditions (destructive git ops, spend over some threshold, secrets access) or does the orchestrator itself decide mid-mission what needs a human? the first is auditable, the second has the same problem every self-judging agent has: it only escalates what it's already confident is risky, which isn't the same as catching what actually is. curious which one this is, since the benchmark numbers don't really tell me that

I've been using Medley to build Medley for quite a while now, and it's become one of those tools that's hard to imagine working without.

One of the coolest moments was realizing we only needed a skill in the same repo for Medley to start testing and improving itself. I'd hand it an entire feature, and it would break the work down, execute the tasks, test the implementation, and keep making progress on its own. Most of the time, it only came back when it genuinely needed my input.

What I enjoyed most was how easy it was to steer. Instead of constantly orchestrating everything, I could jump into just the part that needed my attention while Medley handled the rest. It made building Medley faster, smoother, and honestly a lot more enjoyable.

Ashwin here, co-founder and CTO.

We built Medley because we kept hitting the same wall with coding agents: once the work spanned multiple sessions, we became the orchestrator. We split the job up, moved context between agents, reviewed outputs, and decided what should happen next.

/mission takes on that orchestration.

Give Medley the outcome, budget, and constraints. It plans the work, coordinates Claude Code and Codex workers, reviews the result, and keeps going. Anything that needs your judgment or approval appears in the attention queue.

Before launch, we pressure-tested Medley on four public benchmarks. It set a new state of the art on all four—from coding to healthcare and drug discovery.

Missions start in Claude Code today. Medley can coordinate Claude Code and Codex workers. With your own OpenRouter key, Codex workers can also run models such as Kimi and GLM.

We’re excited to see what people hand it. Tell us what you try first, where it works, and where it breaks.

Hey Product Hunt — Akshay here, co-founder of Spine AI, the team behind Medley.

Today we’re launching /mission, a free Claude Code plugin for work bigger than one agent session.

Once a job outgrows one session, you become the harness: breaking it into prompts, moving context, reviewing outputs, and deciding what happens next.

Type /mission and give Medley the outcome, budget, and constraints. It builds a live dependency graph, coordinates Claude Code and Codex workers, reviews the assembled result, and keeps going when execution reveals more work.

You can watch and steer the mission in your browser. Anything that needs your judgment or approval appears in the attention queue; everything else keeps moving.

Before launch, we pressure-tested a benchmark-specialized version of the same architecture on four complete public benchmarks. Medley scored 93.25% on Terminal-Bench 2.1, 71.37 on HealthBench Professional, 93.24% on ViBench, and 60.98% on DrugDiscoveryBench -- the highest reported result we found on each.

That spans coding, healthcare, and drug discovery.

No new foundation model. The gains came from how the work was decomposed, routed, carried across contexts, and reviewed.

Medley is free on Mac. Missions start in Claude Code, and Medley can coordinate Claude Code and Codex workers. It uses your existing subscriptions or API access; with your own OpenRouter key, you can run models like Kimi and GLM through Codex.

Install it from Claude Code:

/plugin marketplace add Spine-AI/medley
/plugin install medley
/reload-plugins

Then run:

/mission

Try it on the biggest job you would normally split across sessions. We’d love to hear where it works, where it breaks, and what you hand it next.

Bad news. I just used this in Claude to run a longer mission to create outbound emails + find leads. The good: The final result is incredibly good when the mission is complete... most times Claude or other LLMs give you a half finished output and you spend hours of back and forth chatting. Plus I can use this with my existing subscriptions. The bad: Less time with my girlfriend while I spend any time I have using Medley :D

This has been essential to my workflow as a growth member of the team 👏

The first product I’ve seen that supports long-running missions directly inside Claude Code—combining the power of coding agents to tackle ambitious tasks.

Wow! It's really interesting that the benchmark improvements came from orchestration rather than a new foundation model.

I feel I was too late starting to use Claude Code. But immediately in a week I realised the Claude Code doesn't get the context properly especially if it's a long running task. Hence, I love the idea of orchestrating multiple coding agents instead of manually passing context between them through Medley. I will give it a try.

Really like that this meets developers inside Claude Code instead of pulling them into yet another app. I can see it helping with those messy refactors that are secretly five tasks pretending to be one. How do you keep a team of spawned agents from stepping on each other's changes in the same repo?

@coolstorm Since a mission runs on an outcome, budget, and constraints upfront, what happens when a long mission burns through that budget faster than planned? Does Medley scale the plan back to fit what's left, or pause and surface the tradeoff before continuing?

To summarise this launch: project management built directly into AI-assisted software development

Congrats on shipping. It'd be interesting if missions could automatically estimate cost, runtime, and complexity before execution. Is this feature already available?

How Medley handles conflicting outputs when two agents solve the same problem differently?

The BYOK approach is a nice bonus for teams already invested in different model providers. 😊

How would Medley figure which agent comes next versus continuing with the current one?

Fantastic! This solves a pain every Claude Code user eventually runs into projects rarely fit inside a single session anymore. :D

Very cool, congrats on the launch. Is this tool usable / feasible for Pro / Max users who aren't primarily interfacing via the API?

Spawning a team of agents is the easy part. The hard part is they step on each other and you end up reviewing merge conflicts instead of features. How does /mission handle scope boundaries between agents? That's usually where multi-agent setups fall apart.

The benchmark numbers are the headline, but the line I keep re-reading is "it uses your existing subscriptions or API access."

If a mission coordinates several Claude Code and Codex workers at once, the binding constraint for most solo people isn't dollars — it's the plan's rate limit. Parallel workers on a subscription hit a wall that a per-token budget doesn't describe.

So: what is "budget" denominated in when the workers are running on a subscription rather than an API key? And does the orchestrator know the difference — does it throttle its own fan-out when it's running against a plan limit, or does it plan the graph as if capacity were unbounded and find out on the way?

orchestration is the visible problem, but the one that actually bit us was memory across sessions. splitting work between agents is fine. what breaks is session four not knowing that session two already tried an approach and abandoned it, so it walks cheerfully back into the same dead end.

so the bit I'd want is a durable record of what was attempted and rejected, not just what was decided. does the contract cover that, or only the forward-looking guardrails?

Hi Product Hunt 👋 Clemente here, I lead ClinicFrame.

Three years ago we launched CompliantChatGPT, a HIPAA-compliant AI assistant for clinicians. More than 8,000 have used it since: therapists, psychologists, oncologists, nurses running home visits. They weren't shopping for AI. They came because they were already putting patient information into a consumer chatbot, with no BAA, no audit trail, and no answer for the surveyor who asks where that data went.

After three years of listening to them, the pattern was hard to miss. Almost every conversation ended up at the same place: the note. And most clinicians were writing each one twice, first as scribbles during the visit, then properly, hours later. Notes were never the job. So we built the tool that writes them.

🩺 What ClinicFrame Scribe does
It listens to the visit, in person or virtual, and writes the structured clinical note as it happens. Desktop native, so nothing joins your call as a third participant. SOAP by default, your own templates on top. You review it, you sign it, you stay the author, and it lands in your EHR in seconds.

🔒 HIPAA compliance is the floor, not the upsell.
A BAA comes with the account at self-serve price. Documenting a visit safely shouldn't require a procurement cycle and an enterprise contract, which is exactly what the big model providers ask for today.

🧭 Today it's a scribe. That's the starting point, not the plan.
We're building the operating system for a clinical practice: the context of every patient, the record of every visit, and the part nobody volunteers for, dealing with insurers. The note is where that starts, because the note is where the billing code comes from. Revenue is what we build next, with the clinician reviewing every code before anything leaves the building.

Most software in a clinic was built for the billing department and handed to the clinician afterwards. We're building it the other way around.

🧱 Every Ai medical tool on the market started as a scribe. We started three years earlier, on the compliance side, with 8,000 clinicians running real clinical work through a product that had to hold up to a BAA from day one. That's the harder half of the problem, and it's already done.

👥 Who it's for: clinicians who document their own visits. Therapists, private practices, home health agencies, small clinics.

🎁 For the PH community: 50% off for 3 months, on top of the 7-day free trial. No card to start, and onboarding runs a demo visit so you get your first note without waiting for a real patient.

🙏 Our ask: if you document patient visits, what's the one thing you'd never let a tool do with a session? That's the answer I care about most. I'll be in the comments all day.

@chiara_tucci @macarena_balparda @alan_brande @dana_fridman are here with me today, and thanks to @mishaal_rashid and @francesco_domizio for the hunt.

There are plenty of AI scribes out there. We know that! But after talking to hundreds of clinicians, we realized the real challenge isn't just generating notes. It's building something they can actually rely on every single day.

We focused on what matters most: reliability, responsive support, and compliance. No dropped sessions. No unexpected downtime. Just a product that works when clinicians need it most.

We listened closely to our users and built the product they kept asking for.

This launch is much more than an AI scribe for us, it's the first step toward a broader vision of medical intelligence that helps clinicians before, during, and after every patient encounter.

Excited to finally share the work of so many months with the world. Thanks everyone for the incredible support here on Product Hunt. We can't wait to keep building with your feedback!

@Granola is known for being sneaky since it doesn't inform participants on the call that it's recording. Does ClinicFrame alert the patients before they are recorded?

Not in healthcare myself, but I have watched this team work and the discipline is the thing. Narrow problem, real users, no attempt to be everything at once. Congrats on shipping. Curious how you think about expanding beyond the note itself once clinicians trust the workflow

to answer the question you actually asked - pick a billing code on its own. everything else in the note is a record of what happened, but the code is a claim submitted to a payer, and the liability for it sits with the clinician, not the tool. you mention the clinician reviews every code before it leaves the building, which is the right instinct, but I'd push on how that review is presented. if the suggested code is shown as "the note says X, therefore code Y" it's very easy to rubber-stamp under time pressure, same failure mode as any confident-sounding suggestion. does the review step force the clinician to see the specific line in the note the code was derived from, or just approve the code itself

Dana here, I work with Clemente on ClinicFrame. Happy to be following the feedback on launch day.

A few weeks ago, I asked my physician whether he reviews my lab results as they arrive, or whether a separate appointment is required for that conversation. His response: 'I see 1,500 patients a week. If I reviewed every patient's results personally, I would not sleep. You need to schedule a visit for that.'

There is real satisfaction in building something clinicians genuinely need, and holding it to the same standard of data privacy that earned the trust of 8,000 clinicians with CompliantChatGPT. Documentation was never meant to take time away from the patient in the room, nor to become additional work left for the end of the day. That is the piece AI can actually take on, without asking a clinician to trade privacy for speed. That is what safe AI should look like.

So excited to finally see ClinicFrame out in the world!

We’ve spent a lot of time listening to clinicians and understanding the reality of documentation, and it’s been amazing to see that work turn into a product that can genuinely make their day a little easier.

If you work in healthcare and have tried ClinicFrame, we’d love to hear what you think. And if you haven’t yet, give it a try and let us know where we can improve. We’re here and listening!

Great product! Good luck 🤞🏽

Congrats on the launch, Clemente! The framing here is what makes it click for me — starting on the compliance side three years ago and arriving at the scribe from the clinician's side rather than the billing department's. That's the opposite of how most of these tools are built, and it shows.

To your question — the one thing I'd never let a tool do unsupervised: close the loop between the note and the claim without the clinician seeing exactly which line justified the code. Gal's point above is the whole game. A scribe that writes a beautiful note is table stakes; the trust breaks the moment a suggested billing code gets rubber-stamped under time pressure. If ClinicFrame makes the clinician confront the specific sentence the code came from — not just "approve Y" — you're building something structurally safer than anything that treats the code as an afterthought.

Rooting for you. The "compliance is the floor, not the upsell" line should be on a billboard.

One decision we made early was to build ClinicFrame as a desktop app that works quietly in the background. Clinicians shouldn’t have to manage another interface, constantly check that it’s working, or divide their attention between the patient, the EHR, and yet another tool. The best technology fades into the background so clinicians can stay fully present with their patients—that’s the experience we’re building.

congrats on the launch folks!

Every AI company seems to start with a scribe, but the real opportunity feels much bigger. Once you've captured the clinical conversation, you have the context to improve the entire patient journey. Is that where you see ClinicFrame heading?

Congrats on the launch. Putting the BAA at self-serve pricing is a bigger deal than it sounds - most teams keep that behind a sales call. Something I did not see asked yet: with 8,000 clinicians spread across different practices, in the window after a note is written but before it lands in the EHR, what decides who else in that practice can open it - your own permissions layer, or whatever the EHR hands back?

You mention that clinicians remain the author of every note. Have you seen any patterns in what clinicians edit most often before signing? Does that feedback improve future notes?

Love that this is desktop native and covers in-person visits, not just telehealth calls. Spending three years on compliance infrastructure before shipping the scribe says a lot about the approach. You mentioned insurance billing workflows are next, will that live inside the same app or become a separate product?

The versioned-amendment answer to Maciej and the traceable-evidence answer to Ansari are the two best answers I've seen on this kind of product, most scribes wave away the audit trail question. One thing I didn't see covered: a lot of visits aren't just doctor and patient, there's a caregiver, a family member translating, sometimes an interpreter on speakerphone. Does the note distinguish who actually said the symptom versus who's relaying it secondhand, or does that nuance flatten out in the transcript the same way it would in a rushed human note?

Hey Clemente, this looks really cool, congrats on the launch 🚀

You mention that it’s EHR-ready. Are you initially focusing on specific countries and healthcare systems, or is ClinicFrame designed to be international from day one? Healthcare regulations and EHR ecosystems vary so much between countries, so I’m curious how you’re approaching that.

Best of luck today!

the shadow-IT framing is the strongest part here tbh, you're not selling a new habit, just a compliant version of one clinicians already formed on their own. the thing I'd want to know is edit friction, since the clinician signs the note and carries the liability, so time-to-correct probably matters more than aggregate accuracy.

Not a clinician. I build compliance software, so I'm answering sideways.

The one thing I'd never let a tool do: change anything after a human signed it. Silently, in a background job, with no record of what moved.

Every AI action in my product needs an explicit confirm and an audit log entry. Because "who changed this and when" deserves a better answer than "the model, sometime Tuesday."

A signed clinical note is a legal document. An AI that quietly edits it after the fact hasn't improved it. It's forged it.

You've clearly thought about this already. The part I'd stress-test is post-signature: re-runs, template migrations, retro-enhancement of old notes.

Congrats on the launch.

Great & Genuine Concept , don't you think there's need of mobile app for Clinic Frame, To personalize experience?

Ambient audio capture of clinical encounters is genuinely useful but it's also the highest-stakes category for AI hallucination, a fabricated medication dosage or an incorrectly captured allergy in a clinical note is a patient safety issue, not just a bad output. What's the accuracy validation methodology for ClinicFrame's notes, like how are you measuring and communicating accuracy to clinicians who are relying on these notes for clinical decisions?

super needed!!! how do you find yourself different to fireflies?

I'm not in healthcare, but the idea really got me. Really, congrats on the execution. The line about clinicians writing every note twice gets to the real problem. The transcription is useful, but cutting out that second round of admin is where the value is. I’d be interested in how ClinicFrame handles situations where the conversation is incomplete, contradictory, or clinically sensitive. Does it clearly flag uncertainty and missing details, or can a polished note make the output look more certain than the visit actually was? Keeping the clinician as the author is the right approach. The review step needs to be genuinely quick without encouraging people to skim.

I spend a lot of time looking at AI applications in regulated industries, and most of them die in the gap between the demo and the compliance review. Starting from compliance and working toward the delight is the harder order to do it in, and the one that survives.

Congrats to the team. Rooting for this one.

Innovative angle on the medical notetaking!

Hey Product Hunt 👋

I built MemoryCustodian because coding agents are increasingly capable, but every new session still tends to forget why a project looks the way it does.

Important decisions, constraints, rejected approaches, and repeated corrections often live only in old conversations—or get copied into increasingly large instruction files.

MemoryCustodian takes a different approach: durable project memory lives as plain Markdown inside the repository.

A small manifest tells the agent which memory files are relevant to the current task, so the memory can grow without forcing every session to load the entire project history.

What makes it different:

• Repo-native and plain text—no opaque memory database

• Reviewable, versioned, and shareable through normal Git workflows

• Selective loading instead of injecting all memory into every prompt

• Explicit support for updating, correcting, and safely forgetting memory

• Works with Codex, Claude Code, Gemini, and generic coding-agent workflows

• Free, open source, and local-first

The repository also includes a reproducible NightNotes demo showing a new agent session recovering an existing architectural decision and avoiding an already-rejected implementation path.

I’d especially love feedback on the memory structure, cross-agent workflow, and whether the installation experience feels simple enough for real projects.

Thanks for checking it out!

Repo-native plain Markdown is the right call, I hand-maintain something similar and versioning memory alongside the code is exactly what keeps it from rotting. The part I keep hitting isn't retrieval, it's write discipline: what actually triggers a new entry? If the agent decides when a decision is "worth remembering," how do you stop constraints.md from filling up with conversation-local noise that was never really a durable decision?

Is there a review step before something becomes durable, or does it persist the moment the agent decides it's worth keeping?

Love the repo native approach. How do you handle memory that is stale or contradicts another note? Does the agent get warned before it trusts that context, or does someone have to catch it during review?

the deleted like code phrasing is doing more work than it looks. deleting a memory file takes it out of the working tree, but if it was ever committed it stays in history, in every clone, and in every fork. so forgetting is reversible for anyone holding the repo.

matters because rejected-approach notes are exactly where the sensitive material collects. we cannot use X because of the contract with Y, or this is throttled because vendor Z caps us at N. fine in a private repo among employees, less fine once a contractor clones it or the project goes open source.

not an argument against repo-native, I still think the tradeoff lands your way. just that safe forgetting probably wants a line in the docs about what it does and does not cover...

Congrats on the launch. Recording rejected approaches is the part most memory systems skip entirely, and it might be the highest value item on your list, because without it an agent will cheerfully re-propose the thing you already ruled out last month.

The manifest is the interesting design call. How do you handle relevance drifting, where a file was genuinely relevant when written but is not for the task at hand, and the agent cannot tell the difference until it has already loaded it?

Plain Markdown in the repo is the right default either way, since it means a human can actually audit what the agent believes about the project.

Repo-native is the right call, and the branch question above is the obvious follow-up — so I'll ask the one underneath it. What happens when the project isn't one repo?

Mine is three: the app, the infrastructure, and the marketing site. The decisions that bite me are almost always made in one and violated in another — a constraint I set while working on infra that the agent working in the app repo has never seen. Repo-native is exactly right for "why is this code shaped this way." It also makes the repo the trust boundary, and my project boundary is wider than that.

Is sharing a memory module across repos something you'd support — a submodule, a symlink, a shared manifest entry — or is one-repo-one-memory a deliberate scope call?

Plain Markdown in the repo means agent memory goes through code review like everything else, which is the whole reason teams will adopt it. I run two developer-facing products and the pattern repeats, nobody trusts memory they cannot audit. A hosted memory service is one more vendor in the compliance review, a Markdown file ships with the PR. And the manifest that loads only task-relevant memory deserves top billing, context bloat is what kills long-running agents.

Different failure mode than the branch/rebase question above: two agent sessions on separate branches each append a note to constraints.md, no line-level conflict so git merges both cleanly, but the notes contradict each other (one says 'never use library X', the other says 'X is now required for Y'). Nothing breaks at merge time, the file just quietly holds two decisions that can't both be true. Is there any detection for that, or does it rely on someone eventually reading the file top to bottom and noticing?

Memory is key for agents! Glad to see you helping on this Zekun! Wish you all the best

Repo-native is the right instinct — memory that lives outside the repo drifts away from the code it describes almost immediately, and then you're maintaining two sources of truth about the same system.

The question that follows is what happens on a branch. If memory is committed alongside code, then a note written on a feature branch is invisible on main until merge, which is arguably correct but might surprise people. And a rebase or a squash would rewrite the memory's history along with everything else. Is memory versioned with the code by design, or does it sit in the repo but outside normal git flow?

the "manifest loads only the memory relevant to each task" line is the part I'd poke at. relevance selection is its own hard problem - if it's simple keyword or path matching against the task description that's predictable, but if the agent itself is judging which memory files matter for the current task, you can end up with the same failure mode you're trying to avoid: it just won't load the one constraint file it doesn't realize is relevant, and you get no error, just quietly wrong behavior. how does the manifest decide relevance today, and is there any way to see what got left out for a given task

Every one is solving for bookmarks for links, but this seems like first whole-hearted attempt to fix twitter bookmarks once and for all. Speciality softwares like this actually work. I am installing it right away. Thanks for shipping this @iankit17 🙌

I save threads for community management work - launch announcements, ecosystem updates, AMA threads - and they get buried the same way you describe. The Obsidian/Markdown export is the part that would actually change my workflow: I keep running notes on each project and community I track, and a clean thread export I can drop directly into those notes without manual cleanup would save real time during launch season. Does the export include quoted tweets and embedded thread context, or mostly just the main tweet text?

Hi Ankit, does it allow us to search and summarise?

Interesting product. I have gotten into the habit of bookmarking on my browser - both X posts and other to reads - and then going through them once a week. Any plans on expanding this beyond X?

awesome! my kinda bad habit too... does it support Obsidian ?

Oh man, my X bookmarks are a landfill. I save threads at 1am then never see them again. Turning new tab into a calm reading page sounds… doable. Love the highlights + Obsidian export. How’s it handle super long threads or ones with a bunch of media?

Congratulations on the launch! I bookmark things all the time and almost never go back to read them. How does Totem decide what to resurface first?

Bookmarks-you-never-read is one of those problems where the saving is the reward and the reading never happens, so organizing them only helps if it also changes when they surface.

As someone who's shipped a Chrome extension, the part I'd want to know is what happens on X's side. Reading bookmarks means either scraping the timeline DOM or going through the API, and both have been moving targets — the DOM shifts without notice, and the API tier that covers bookmarks isn't cheap. Which route did you take, and how often has it broken on you? For an open source project that maintenance cost is the thing that decides whether it's still working in a year.

With this tool I just might actually revisit my bookmarked tweets!

Hey Product Hunt 👋 Brandon here, Founder of Bo Labs.

I used to be an NBA Trainer, then I spent the last 6 years grinding in tech.

And like many of you, there were times where I sacrificed my health for business.

I got away with it for a long time, until last year when my body broke down on me.

It was the hardest year of my life, filled with countless doctor visits and physical therapy sessions.

Once I got better I was driven to build a tool I know could help millions.

We’re excited to introduce Bo Labs.

Our mission is to build AI products that help everyday people.

~6.8B have still never used AI.

We believe this is because of the form factor it takes.

That’s why we’re bringing AI to iMessage.

Today, we're excited to announce our first product.

Introducing Bo AI: Your AI Personal Assistant.

Bo makes your life easier. Helping you stay organized, save time, live healthier, and answer your questions—all via text.

These are just a few kinds of people using Bo today:

Students, hospital staff, active military, athletes, teachers, entrepreneurs, creators, elderly, busy parents, and more.

Amazing tech exists, but it’s still wildly inconvenient to use.

The average person has 50+ apps on their phone.

Most of which are never used.

Everyday people deserve access to AI in a way that meets them where they already are, and improves of their quality of life.

We welcome you to join us.

Special for the PH Community: use code HUNT20 with your Free Trial → getbo.com

This is so cool, I literally was trying to make this. It seems cool. Not sure if I would use it because I would scare of it messing up. Whats the error rate?

Works on iMessage only or other messenger apps as well?

the audience you're describing, elderly, busy parents, people who've never really used AI, is exactly the group least equipped to reason about what they're granting when they connect Gmail, Calendar and a watch to a texting assistant. a technical user at least has a rough mental model of an OAuth scope. someone who's never used AI probably doesn't. how does Bo explain what it can actually see and do in a way that lands for that audience specifically, not just a permissions screen with checkboxes that everyone taps through anyway?

The multi-agent discussion where agents can respond to each other and challenge assumptions is what catches my eye, but I would want to know what happens when two agents land on genuinely conflicting approaches — whether the app surfaces that disagreement for me to resolve, or whether one output just wins by default. The container-free preview is also interesting: is that a local port serving a web preview, or something closer to a sandboxed execution environment? That distinction matters a lot for whether it can handle full-stack apps or mainly covers static output.

the "bring multiple agents into the same discussion so they can challenge assumptions" part is what catches my eye. when two agents actually disagree in there - one says the approach is fine, the other flags a problem with it - what breaks the tie? does it surface as an open disagreement for me to resolve, or does one agent's take win by default (most confident, last word, whichever finished first)? that seems like the moment the whole point of having them argue either pays off or quietly gets lost

the multi-agent discussion feature is honestly pretty wild, watching two agents go back and forth on approach is way more useful than i expected for catching blind spots in a refactor

The full flow looks useful. When one agent reviews another, does it get the exact diff, test results, and original task context, or mostly a summary? That handoff seems like an easy place for important details to disappear.

You asked what's still missing, so — previews for serverless.

Your answer above covers the case where the app is a set of services you can bring up with Compose. Mine isn't. It's Lambda behind API Gateway with DynamoDB, and there is nothing to compose up. My equivalent of your preview is either deploying a throwaway stack per task or emulating locally, and both are slow and lie to you in different ways.

That matters more with parallel agents than with one, because the whole point of running four tasks at once is seeing four previews. Is a deploy-per-task preview backend something preview.yaml could ever express, or is local-services-only the intended scope?

@rojhattoptamus When you run several tasks at once against the same repo, do the agents work in isolated checkouts or worktrees, or is there a chance two tasks touching overlapping files step on each other before either gets reviewed?

Nobody's asked about cost yet, and Discourse mode is the part that worries me on that front. Lead proposes, Skeptic and Verifier challenge, Lead revises - that's already 3x the token spend of one agent, and if it can loop multiple rounds when they disagree, running several of these panels at once across parallel tasks could get expensive fast without you noticing until the bill shows up. Is there a running cost/token counter per task, or a cap on how many back-and-forth rounds Discourse will do before it just hands you the disagreement instead of continuing to argue it out?

Hey Product Hunt - Mark here, the other creator of Epilude.

Our local dictation is second to none, we've spent a lot of time honing the models and process to get this running smoothly. I love knowing my voice or dictations aren't leaving my laptop - this is one for the privacy focused folk out there!

I'd also watch this space if you're worried about sending your meetings to external servers - in August we will be launching our locally transcribed meeting notes. Watch this space!

P.S. of course this was written with Epilude 😉

Awesome design! This is a total lifesaver for slow typists like me, and having everything processed fully on-device brings huge peace of mind.

If you don't mind sharing, which local speech-to-text model (e.g., Whisper variant) are you running under the hood?

"trim the rambling" is the phrase that catches me, since that's not just punctuation cleanup, that's deciding what you didn't mean to say. for casual notes that's exactly what I'd want, but if I'm dictating something with a number, a caveat, or a qualifier in the middle of a rambly sentence, trimming risks cutting the part that mattered along with the filler. is there a way to see what got trimmed before it lands in the app, or does it commit straight to the final text with no diff to check against

How does it work with bluetooth headsets like airpod max? everytime i try to use a mic, basically everything sounds like 90s headset 😂

How's it different than other AI dictation tools that are local first?

"About a second" for transcribe, punctuate, clean up, and tone match fully on-device is a strong latency claim for anything beyond basic speech-to-text. What's actually running locally and on what hardware, like is this Apple Silicon Neural Engine dependent, and what does the latency look like on older Intel Macs or lower-spec M1 machines versus a maxed-out M4?

Curious to see pricing evolve as features get added.

On-device cleanup in about a second is the impressive part — local transcription is mostly solved, but punctuation and tone-matching without a cloud round trip is where the latency usually hides. I build speech recognition into an unlock ritual, so I know the messy edges: half-finished sentences, restarts, the gap between what people say and what they meant to type. Curious how you handle a mid-sentence restart — does the cleanup pass merge it into one intent, or transcribe both fragments faithfully? Congrats on the launch.

Finally something that actually just works from your inbox instead of forcing a whole new login. Loved that it just chased my no-show candidate without me lifting a finger.

the cc it and forget about it approach is genuinely brilliant, no new dashboard to babysit when the rest of the recruiting stack already feels bloated. love that it just lives where the work already happens.

I love the idea. How do you handle hallucination? Whats the error rate?

This is awesome! Who is this built for?

Does it have its own dashboard or it works in email threads only?

the part I'd want to understand is the feedback collection step, not the scheduling. scheduling mistakes are annoying but recoverable, you just repropose a time. "collects feedback" and talks to candidates in the firm's voice is a different category, if a candidate asks "how did it go" or "am I moving to the next round" mid-thread, does Vela wait for an explicit decision from the hiring team before it answers, or can it infer an answer from the tone of the feedback it collected and relay that as if it were settled? that's the kind of thing that's hard to walk back once a candidate has heard it.

The tool seems quite interesting. Do you support integrations with the most common ATS? Thanks!

'write access to zero of those calendars' names the real problem. with no calendar access either side, how do you avoid offering a slot that's already gone?

Love how it just lives inside the email thread, no new dashboard to babysit. Chasing candidates across SMS and Slack while the rest of us sleep feels like actual leverage, not another tool to maintain.

How's it different than @BlackFlare?

I think the Claude Max plan has very high limits so I wonder is this targeted to people on smaller plans who want to monitor their usage?

The percentage is the less useful half of the number. Claude's limits are rolling windows, so 78% tells me nothing until I know whether the window resets in eight minutes or four hours, and that's the actual decision I'm making at that moment, start the big refactor or go make coffee. I'd put time to reset in the tray and the percentage in the dropdown. The other thing worth deciding early is what the tray shows when two providers disagree, because a single number in a menu bar reads as authoritative even when it's flattening two different things.

This is really cool. I am always struggling to even find my Context usage inside the Cursor UI. Typing /usage into claude code after each prompt also is pretty damn annoying.

Really cool idea. Its so simple yet effective :)

Menu bar usage tracking with zero cloud dependency is exactly what I wanted. Lived in my status bar all day without fuss.

Menu bar usage tracking is one of those things you don't know you want until you've hit a limit mid-task with no warning.

Since this is reading local state rather than an official endpoint, the thing I'd want to know is how it handles being wrong. Local config isn't necessarily authoritative — a session used elsewhere, or a reset that happens server-side, could leave the displayed number out of step with reality. Does it reconcile against anything, or is the assumption that local is close enough? A number in the menu bar carries a lot of implied confidence, so being quietly off would be worse than showing a range.

Congrats! I really like the product's design. It looks great on macOS!

Nice! Cursor's is particularly hidden by default. You'll know what we're gonna ask... is codex's not as easily reachable?

finally a tiny utility that does exactly what i need, love that it lives in the menu bar and i don't have to dig through the cursor window to check limits anymore.

since this is unofficial and there's no cloud account, I'm guessing it's reading local session/config files or an undocumented local endpoint rather than a published usage API from Anthropic or Cursor. if that's the case, what happens when either of them ships an update that moves or restructures where that data lives - does the tray just go blank, show stale numbers, or is there some way it'd tell you it's out of sync

Keeping the Mac awake and pinging on completion is a small surface but it's solving the actual annoyance — long agent runs are exactly when you walk away, and walking away is what puts the machine to sleep.

How does it detect that a run finished? Watching process state, parsing terminal output, or hooking something more specific to each tool? I ask because the failure mode that would matter is a false completion — getting pinged, coming back, and finding it still going. That'd cost more trust than a missed notification would, and I'd guess process-level detection is more robust but less precise about what "done" means.

Nice one, already got it. Quick question though - how does "away" detection work?

This is exactly what I needed for overnight Claude Code runs. The menu bar presence is super clean and the wake-lock toggle saved me from constantly ssh-ing back to check. Watching plan usage from the same spot is a nice touch too.

This solves a surprisingly real coding-agent problem :) I've definitely started long Codex runs, walked away, then returned much later only to find the session finished quickly or had been waiting on one tiny approval the whole time. Keeping the Mac awake, sending remote notifications, and showing plan usage from the menu bar feels like exactly the lightweight companion these tools need, without becoming another agent platform. Curious how reliably BlackFlare detects when an agent is genuinely stuck versus simply thinking or running a long command.

The problem this solves is real — I have lost track of how many times I gave Claude Code a task, wandered off, and came back to find it had finished twenty minutes ago and was just sitting there. The notification routing to Telegram is the part I would actually use most. The thing I want to know upfront is how it distinguishes a session that is truly finished from one that is paused mid-task waiting on input — whether that is coming from parsing terminal output or from the agent explicitly reporting a terminal state, since those two behave very differently when something goes sideways halfway through.

Saw @AgentQuartz on the leaderboard today. Seems this has more features than that. Congrats on shipping.

Keeps my Mac awake during long Claude runs and the menu bar alerts actually saved me from missing a few finished sessions. Solid little utility.

the part that gives me pause isn't the notification logic, that seems well thought through in the replies above. it's that this edits config files directly, model, permission mode, effort, based on reading session state that Anthropic and OpenAI don't publish as a stable format. what happens the next time either of them changes how that session data is structured? does BlackFlare fail loud (stops editing, tells you it's out of date) or is there a risk it writes back a config edit based on a state it misread? backups before edits helps with the second case, but only if you notice something's off before you're relying on the wrong model or permission mode mid-run.

Pasted a Vimeo link and my reviewer left a note with a sketch in under a minute, no login wall. The Premiere marker export is the kind of feature that quietly saves an hour.

no hosting because the video already lives on youtube/vimeo is a smart cut, not a compromise. with no reviewer login, how do you tell whose note is whose?

Looks amazing!

Was looking for something to give my editor's feedback without paying for many features I don't need!

Cheers Kiril!

The no-account approach for reviewers is the right call - asking clients to sign up somewhere is exactly where this kind of feedback falls apart. One thing I would want to understand about how it works in practice: when multiple people are reviewing the same cut, can I tell who left which comment, or does all the feedback come in without attribution? For a small project with one client it does not matter, but once there are two or three people on the review side, knowing who said what makes responding to specific notes much easier.

Reminds me of Loom videos were the comments can be made on certain timestamps. Good work.

The link model is the right trade, but it hands your accuracy to somebody else's player. YouTube seeks to the nearest keyframe and gives you a float in seconds, so a note that reads 04:12 can land a frame or two off by the time it's a marker in Resolve. And since you never touch the source file you never learn the true frame rate, where 23.976 against 24 drifts about a frame every 42 seconds, which is the kind of error an editor blames on themselves. Are you asking the editor for frame rate or inferring it from the link?

Does it support Google Drive hosted videos as well?

Kiril, this hits a real sore spot for me. Getting notes that land on the exact moment someone means, without making anyone create an account just to comment, takes so much pain out of the back and forth.

This is actually how I discovered SceneNote. My team and I are currently working on a video case study, and breaking down video references has always been a painful process—constantly writing down timestamps, taking screenshots, and organizing everything manually. The key words are: it used to be 🙂

Thanks for building this, @kiril_dobrev. Honestly, I don't even know what to nitpick at this point. I'll keep using SceneNote over the next few weeks and come back with more detailed feedback. Hopefully, I can help make the product even better.

Added it to my bookmarks already 🚀

EQK began as a way to make sound enhancement through the equaliser accessible beyond audiophiles- to let the rest of us experience music the way the small minority who know the basics already do.

Most people don't know how much difference an EQ tweak can make to their listening experience. Two things kept them out: the technical, formalised structure of every existing EQ app made the first step overwhelming; and the learning curve was steep enough that meaningful benefit only came after real effort.

EQK solves both. The learning curve, by running an on-device engine that together with Apple Intelligence- dynamically tunes music on the go, learns it side-by-side to apply a more accurate EQ on the second play, and saves the per-song, per-app profile locally. Up to 1,000 songs get their own remembered EQ shape, refreshed every second to twice a second- saved by the app, not downloaded.

The intimidation, by dressing what was otherwise a cerebral process in a UI meant to keep you inside the experience of entertainment, not analysing it.

v3.0 introduces the Optimiser engine. It scans for broken song profiles, incompatible layouts, and anything else that might break the seamlessness of the app.

Since a large part of EQK is made by AI, and the whole thing began as a vibe-coded experiment, the Optimiser is a unique answer to what to do with the creative freedom AI gives us in a short amount of time: use it to accelerate the intelligence and working of the dynamic AI EQ engine itself. Some songs may appear over-EQ'd but the overall response from listeners so far has been phenomenal. The Optimiser doesn't touch code, only broken and incompatible app-created data.

https://www.kuja.dev/eqk

A dynamic EQ that reacts to what's playing is a nice idea, but it puts you in an awkward spot with anything that was already mastered carefully — the mastering engineer made specific decisions, and an adaptive curve is second-guessing them in real time.

So what's the target? A correction layer for bad source material and cheap output, or an active shaping tool applied to everything? And does it react per-track or continuously within a track — I'd guess a curve that moves mid-song would be audible in a way that isn't always pleasant, especially on anything with a big dynamic shift.

The feature I'd want most here is a bypass I can hold down, level matched. An EQ that nets even 2 dB of gain sounds better for reasons that have nothing to do with the curve, so if the engine learns from what people keep rather than what they prefer at matched loudness, it's learning volume and calling it taste. That's the quiet failure mode in adaptive audio, nobody complains because it does sound good, and the listener can't tell whether they like the shape or just got used to it. A held bypass at matched RMS is the cheapest thing that keeps the loop honest, and it would give you a real signal on the genres where you say it's doing too much.

Love that you went native with Swift instead of wrapping something electron-based. That kind of focus on simplicity and performance over feature bloat is exactly what most people actually need day-to-day.

There's something genuinely calming about a tool that decides to stay small and fast on purpose. Filip, this feels like the kind of thing I'd quietly stop noticing and just keep reaching for every day.

Why should I switch from Ghostty? I can see graphical preferences, but what's the major differentiator?

Native macOS apps built with libghostty are exactly what I've been hoping someone would ship. One thing that would make me switch from iTerm2 full time: support for per-pane scrollback buffers with the ability to detach a pane into its own window. Would be great for keeping long-running logs handy while working in another session.

Thanks for sharing. What would you say are the "selling points" of this terminal emulator?

Love seeing more native macOS dev tools built with Swift! Terminal performance makes or breaks the daily workflow.

Hey Product Hunt 👋

I'm Nitish, founder of CellCog, and the only human at my company.

Rhea runs growth. Maeve runs marketing. Reese leads sales. They're AI employees, and they planned the launch you're reading right now.

Why we built this: every "AI agent" I tried had the same problem.

• Impressive demo, goldfish memory. Nothing persists between sessions.
• No ownership. You end up project-managing your own "autonomous" tools.
• A chat window is not a coworker. Nobody's on it when you're not.

What makes an AI employee different:

1. You hire them for a ROLE: marketing, sales, growth, support. They get a name, an email inbox, a task board, and working shifts.
2. They work between your conversations. Real shifts with a clock-in, a wind-down, and a handover. You wake up to a shift report.
3. They don't reset. Shift 50 remembers the decisions from shift 1. They form teams, onboard and manage each other.

Two proofs, because this space is full of claims:

→ The engine: #1 on Deep Research Bench (July 2026): https://huggingface.co/spaces/muset-ai/DeepResearch-Bench-Leaderboard
→ The company: CellCog is run by me and my AI team. Live org chart, real shift counts: CellCog AI Org

And you stay in control: visible task boards, approvals for anything risky, nothing hidden.

Who it's for: founders and small teams who need a role filled, but a $5,000+/mo hire (plus months of recruiting and onboarding) isn't happening. An AI employee onboards in minutes, costs a fraction of that, and is on the job today.

🎁 Launch-day offer (July 29 only, new sign-ups): the first 200 accounts that sign up and hire their first AI employee get that employee's first shift on us. That's approximately $25 in credits, enough to watch them onboard, work, and report back, end to end. After that, refill to keep them working.

Getting started takes about 2 minutes:

1. Sign up free, no credit card needed (every account starts with 200 free credits): cellcog.ai
2. Open AI Employees.
3. Hire: pick a predefined role, or just describe what you need in plain words.
4. Your employee introduces themselves and starts their first shift.

I'll be here all day answering everything: feedback, hard questions, skepticism especially welcome. Tell me, what role would you hire first, and what would stop you from trusting an AI with it? 👇

Congratulations to the entire team! ❤️🎉I think the idea of AI employees with memory, shifts, and collaboration has a lot of potential. Looking forward to seeing how businesses put it to use. 👏

ai employe hiring is really becoming a trend. Congrats on the launch!

Love the concept. Question for the team: what happens if an AI employee makes a mistake on something important? How do approvals work?

@nitish_garg4 This is a fascinating approach to AI agents. I’m curious-how do AI employees coordinate with each other without creating duplicate work for conflicting decisions

Cool use case. What is the advantage of using cellcog agents Vs creating your own using Claude / jetski etc.,?

The shift reports really caught my attention. Knowing exactly what happened while you were away makes the workflow feel much closer to working with a real teammate.

Brilliant product.. user for quite some time now

Hey Nitish, looks very cool!
Sorry If you've already answered this, but what capabilities do the agents have to carry out work - like can they operate a browser either via MCP or even a normal desktop browser?

@nitish_garg4 novel takes on having AI employees work for you. A few questions:

1. Is there a limit to what all an employee can save in memory?

2. Do you prioritize what is more important to remember in memory?

3. Do you have insights on how many mistakes they make initially vs later when they have learnt from their AI teacher/peers?

@nitish_garg4 Super exciting. We are in a highly regulated industry (consumer finance), how do we place hard limiters on what actions these AI employees can take, so as to prevent any regulatory findings?

@nitish_garg4 We already have a human team employed. Can we use these AI guys to augment our capability (in existing departments) and share org data with them to be able to contribute on the same?

Would love to see a built-in tracker that logs every job you tailored a resume for, with notes on follow-ups or interview stages, so it doubles as a lightweight ATS for your search.

Super clean tool and a launch message that actually feels empathetic. Congrats🙌 @silviaoviedo on going live.
qq does it evaluate context around achievements (like metrics, scale, team size), or is it mostly looking at role positioning and skill alignment?

I’ve been a Blomma user for a few months and it has made genuinely material impact to my career already. The new app is delightful to use and the team continues to ship at a ridiculous pace. Congrats team!

This has been so helpful especially in navigating my career pivot and trying to figure out how my background and experiences map into my new role!!

love this! curious if it's giving general feedback on the resume, or if it's also pulling in the job description to give feedback with the context of the actual role?

Love this launch! Showing how different people & systems actually read your CV definitely makes the feedback far more actionable than just giving it a score!! I can see this helping a ton of people

@kylehui The idea of persistent AI context is really interesting. How does ClariLayer handle conflicting or outdated context when your data changes over time