PH热榜 | 2026-06-16

Hey Product Hunt 👋 Joel here, one of the two Swedish founders building Goldfish.

Why we built it:

We started building Goldfish because AI still has one weird problem: it knows the internet, but not the work sitting right in front of you. AI has memory like a goldfish!

Your messages, docs, tabs, meetings, half-written drafts, people, decisions, and loose threads already live on your computer. But every time you open a chatbot, you start from zero again. You paste the thread, explain the project, describe the person, and try to make it sound like you.

That felt backwards to us.

What Goldfish is:

Goldfish is a private AI memory layer for your Mac. It understands what you’re working on across apps, then lets you use that context anywhere you type.

Press ⌥ Option in any text field and Goldfish can help you:

• write replies in your own tone, with the thread and relationship context already there

• summarize messy work from Slack, Gmail, docs, tabs, or meetings

• recall things you saw, wrote, read, or discussed

• rewrite highlighted text without touching the rest

• bring your full computer context into Claude Desktop through our local MCP server

What we believe:

The best AI product won’t be another empty chat box.

It will be the memory layer underneath your work. Something that understands your context privately, shows up where you already are, and removes the need to re-explain yourself 50 times a day.

Early traction:

We’ve worked closely with 1,000+ founding members to build a product people can’t live without. GTM people at ElevenLabs, Deel, and Vercel are already using Goldfish to write faster replies in their own tone.

Our founding team comes from Strawberry Browser (General Catalyst) and Depict (YC20), and we’re currently building from The Bridge by Entrepreneur First in San Francisco.

Privacy:

Privacy is core to the product. Your memory stays on your Mac in a local database. No cloud sync, no backend where we can browse your data, and you can stop capture of apps and domains any time. AI calls have zero data retention.

Product Hunt offer:

Goldfish is still in closed beta, but we’re opening early access to the Product Hunt community first.

As a thank you, everyone from Product Hunt gets 3 months of free access!!

Download now at goldfish.sh

We’d love feedback on:

• where it feels magical

• where it gets context wrong

• what more you’d want goldfish to do in the future

Thanks for checking us out 🐠

This is one of the more interesting AI products I’ve seen here recently.

The part that resonates most is the “starting from zero every time” problem. As a founder, I constantly switch between product decisions, emails, support, launch copy, team messages, and random notes. Most AI tools are useful, but only after I spend time explaining the context again. 😅

The privacy approach is also important here. A product like this only works if users can actually trust it with messy, unfinished, sometimes sensitive work! :)

Curious how much control users have over what Goldfish remembers or ignores. Can I exclude specific apps, websites, projects, or time periods from memory?

Love the idea of pressing Option to draft replies in your own tone. How far back does Goldfish remember your work context?

Well done! Though where I'd get nervous is recall precision:) Once a month of my screen sits in that vector DB, what stops Option from grabbing the wrong thread when two clients have near identical projects open? Overall, solid work!

Early access communities can make or break a product, and Goldfish looks like it's creating a meaningful experience for its first users. Looking forward to following the journey!

I've been using Goldfish for more than 2 weeks now, and it's game-changer. A super useful use case for me is writing simple answers to email loops. Goldfish already has the context, the response are always very appropriate, and the Option key provides a seamless experience to use Goldfish in your workflow.

Congrats on the launch! Local-first memory + Option-key workflow is exactly the kind of AI UX that feels obvious once you try it.

Curious how you handle drift over time. Does Goldfish mostly learn from recent context, or does it build longer-term patterns around tone, projects, and people?

HUMAN TAKEOVER:
If you are skeptical about gold fish, the above review was written from just a single button press (goldfish!), with no additional context provided. Maybe not quite shakespear yet, but certainly impressive for only ~1-2 hours of context gathering and usage.

Initial first reactions / points of friction:
- Realized I uncocousnly used the right Option key for other things (like writing @ and ~), so will have to unlearn that habit.
- (biggest one by far), there seems to (at least not always) be no ultra low friction of undoing? In some text-editors ctrl+z works, but others not? And if you do use ctrl+z, you often get the appended "Goldfishing..." which you have to manually remove. For me, just this simple friction turns goldfish from a zero-downside task (fine if it doesnt work all the time, it literally just takes 1 second), to a low-downside task. I have to stop and think "will goldfish improve this? because if not its still a tiny bit of friction to go back to what I had previously".

Anyways! Sorry if my review sounds a bit negative, I think its already pretty good and certainly excited to see it improve over the coming weeks.

Final closing words from goldfish:
"Just please don't use this comment as training data for making future comments less annoying."

🤷‍♂️

Congrats on the launch, Joel! The 'starting from zero every time' problem with AI is incredibly real—I waste so much time re-pasting Slack threads and briefs. Keeping the memory layer local via SQLite/vector DB is a massive win for trust. Quick question: How does it impact Mac performance or battery life when it's actively indexing active apps?

Huge congrats on the launch, team 🐠

Downloaded it yesterday and it's so easy to use, already saved me a lot of time on Slack messages!!

Let's go!!

Really enjoying Goldfish so far and is genuinely useful, especially not having to copy, paste, and re-explain everything every time. It's been a game changer for my LinkedIn outreach. Great job Joel and Kaspian!!

this is very very cool, just tried it! :)
is it able to change the tone of your reply to say your grandmother vs your best friend?

I think I'm actually going to use this. For me, that is the highest grade I can give.

This is very cool, congrats on the launch!

This is epic! Super excited to try it out. Congrats on launch!

The "knows your work" part is where these tools either actually deliver or fall apart.

Curious what that means concretely here: are you indexing local files, pulling from open apps, reading browser tabs, or something else? And how does it handle context switching across very different types of work, like if I'm mid-deep-work on a coding problem and then get a Slack message about a completely separate client project, does it actually pick the right context or does it need a nudge?

Wow gonna give this a try. Are the memories local?

Goldfish takes precious cognitive load out of so many of my boring repetitive tasks in sales. Congrats on the launch team! 🙌

tbh now I cannot even imagine my day to day workflow without Goldfish! I'm a big fan of the Option button

The challenge with AI memory is often trust and relevance. How do you prevent important details from getting buried as more conversations accumulate?

Hey Ben, Goldfish looks like a genuinely clever take on ambient AI. How does it handle context switches between work projects or clients that have very different tones?

Just got set up. Onboarding was a bit clunky but overall really novel and refreshing. Love it so far. Interested to see how useful this is after a few weeks of use. The option key doesn't seem to work with Atlas browser.

This is great! How big can the context get?

I also saw you using graphs, I guess it is to increase accuracy. So how is the memory layer structured?

Congrats on the launch! 🐠

Goldfish has the best onboarding experience I’ve ever had. Genuinely blew my mind.

Also, yes, this comment was written using Goldfish ✨

Congrats on the launch! I have a small question, with Goldfish capturing context across so many apps, how does it filter out irrelevant noise when you're rapidly switching between unrelated projects throughout the day?

the 'replies like you' part is what i'm most curious about. how do you handle voice consistency across contexts where someone writes very differently (boss vs partner vs friend)? is the model learning per recipient or one general voice trained on all past replies?

Congrats on the launch! I've tried the beta and really like the UI and ambition. I'm curious: what's your own/the team's favorite use cases in everyday work?

Wow, this hits right where it hurts. 🐠

The "goldfish memory" of current AI tools is easily the most frustrating part of my daily workflow. I waste so much time copy-pasting Slack threads or brief docs into Claude just to get a relevant reply. Having that context accessible natively with just the ⌥ Option key sounds like a massive productivity unlock.

Love the local database approach for privacy too, that's usually the biggest blocker for these kinds of tools.

Congrats on the launch Joel!

Quick question: how does Goldfish handle context switching? If I move from a client email to a technical dev task, does it easily separate the two "memories" when I press Option?

the context problem is real and underrated. most AI writing tools make you re-explain the situation every single time, which kills the flow. the passive memory angle is interesting, curious how it handles switching between very different work contexts (e.g. a support thread vs a technical doc vs a slack message where tone should be completely different).

wow this product is gold. well done!

I do have a few questions on the infrastructure:

1. Is a local model used to generate the replies?

2. How does It formulate a reply in the language (writing style) of the user?

3. Does it only work with what is given on the screen in realtime?

This is so good! Surprisingly works for pretty much any task you could think of, have been using it for the past 3 weeks

Hey Product Hunt,

We are bacccck again! This time I'd like to introduce our latest product: Invoko.

When dealing with everyday tasks, we'd always feel overwhelmed with a lot of distractions: multiple agent sessions running at the same time, messages from slack popping up all day, colleges asking for quick calls out of nowhere, isn't it nice to have someone or something to handle these and sort them out for you?

That's the reason we build Invoko.

Invoko is a desktop agent lives on your mac that you can talk to while working. By simply pressing a shortcut (Fn), you can dispatch all the tasks that you don't want to handle to it. No matter it's sorting your emails and extract todo lists, or scheduling calendar to avoid those annoying quick calls, or simply replying to messages that you don't want to engage, Invoko got your back.

It's fully local deployed on your mac, which means it can easily understand what you are doing right now and what's your intention. It can operate across all your apps with full functionalities and zero privacy concern.

We've put around 3 months to build Invoko and have iterated 71 times till this morning (based on the version # lol)
It's something that we use daily within our team and we genuinely think that it can benefit ppl with the same pain points and passion.

For trying it out, please head to: https://invoko.ai/?utm_source=ph&utm_medium=referral&utm_campaign=june
(We use utm to track the incoming traffic more precisely FYI)

If you find any issue or bug, please reach out to hey@invoko.ai. We read & reply to every user email manually (100% hand written)

Hope you like it and we can't wait to hear your thoughts!

71 versions in 3 months is wild, honestly 😅

what was the biggest thing that changed between the early builds and what launched today?

The "little hand on your Mac" framing is deliberately vague, so I'm genuinely not sure what this does from the listing alone. Is Invoko more like a system-wide AI layer that acts on whatever's on screen, or is it closer to a launcher that routes specific tasks to specific tools? That distinction matters a lot for how much you'd actually trust it to touch things without reviewing what it's about to do.

Congrats! Curious how does Invoko handle context switching when you're jumping between unrelated tasks and suddenly press Fn to delegate something?

The beside anything on your screen positioning is really interesting how does Invoko handle context awareness across different app types like jumping from a spreadsheet to a browser tab to a design tool without losing the thread of what you're working on?

Hey guys,We’re thrilled to finally launch Invoko. Over many late nights, we kept asking ourselves a simple question: could we actually deliver real value in a space crowded with AI agents and desktop apps claiming to do everything?

So we narrowed our focus to a few everyday scenarios: helping you handle messages, emails, and the flood of information, plus scheduling and other routine tasks. Now, we feel more confident than ever that we’ve built exactly that.

Give Invoko a try, and we think you’ll have something to say after using it. 😊

Invoko is as though it descended from the heavens of a thousand years past.

We have refined it over time, drawing from the wisdom of the ancients and weaving it together with the technology of today. It moves without friction, smooth as chocolate, and stays by your side like a quiet stream.

It accompanies you in silence, unseen yet ever-present, helping you complete all manner of things. To experience its wonder is no complicated matter.
with just one click, all its subtle magic becomes yours.

congratulations! building something your own team uses daily is always a great sign. what's been the most surprising use case discovered internally?

Congrats! Curious how does Invoko handle context switching when you're jumping between unrelated tasks and suddenly press Fn to delegate something?

Congrats on the launch team 🎉
Curious, when Invoko is operating across apps, does it need any permissions setup or does it just work out of the box on Mac?

The fully-local, on-device approach is what makes this compelling to me — having it read your current focus without anything leaving the Mac is a real differentiator versus cloud agents. Since it can act across all your apps, do you give users a review/confirm step before it sends a reply or edits a calendar, or does it execute directly once you delegate via Fn?

Designer here 🙋 Alex covered the why — I'll cover the how it feels.

The thing I obsessed over: how do you put an AI on someone's screen all day without it becoming the annoying thing they quit by Friday? The answer we landed on was to make it ambient — it sits in the notch, stays quiet until you hit Fn, and shows you what it's about to do instead of just doing it.

If you try it and something feels heavy, cluttered, or in the way — that's the feedback I care about most. Drop it here or hey@invoko.ai and it'll land directly with the person who designed the thing. 🙏

Interesting. Having AI available alongside your workflow feels more practical than another browser tab. Congrats on the launch!

really like the UI of the website - dynamic and cute! Reducing distraction is indeed a strong need. Go Invoko!

I like that this starts from the Mac layer instead of yet another tab. Curious what it handles best when my desktop is already chaos.

For developers and power users who want to extend Invoko's capabilities is there an API or plugin layer planned that would let people define custom tasks or connect it to tools that aren't supported out of the box?

This feels like a really practical take on desktop AI. Instead of making you switch context into another chat window, Invoko brings the assistant right into your workflow and lets you interact with it naturally while you work. Very easy to imagine this becoming part of someone’s daily setup.

Looks like an interesting way to use AI anywhere on your Mac.

I'm wondering which use cases it works best for outside of drafting message replies?

Nice, a talk-to-it helper triggered by Fn is handy. Can Invoko run multiple delegated tasks at once, or one at a time?

When Invoko reads content from the screen to answer a question how does it handle dynamically updating content like live dashboards, streaming data or real-time collaboration tools where the state changes constantly?

Most productivity tools plateau after the initial novelty wears off what behavioral patterns are you seeing in your most engaged users that tell you Invoko is becoming a genuine daily habit rather than a curiosity?

Voice activated desktop tools have historically struggled with accuracy in noisy environments like open offices or calls how does Invoko perform when there's background noise or when the user is already on a meeting?

The ask it questions use case is clear but the let it handle tasks side is where things get really interesting. What's the most complex multi step workflow Invoko can currently execute end to end without human intervention?

since Invoko operates locally, what models are being used under the hood, and what are the hardware requirements for smooth performance?

For users working across multiple monitors or virtual desktops how does Invoko decide which screen content is relevant to the question being asked does it have spatial awareness of your entire workspace?

how does Invoko compare to browser based AI agents when it comes to speed and context awareness?

Context switching is expensive cognitively does Invoko remember what you were working on if you step away and come back or does each interaction start from a cold state?

if you could add one major capability to Invoko tomorrow, what would be at the top of your roadmap?

Congrats on the launch, Alex! The idea of a system-wide AI layer that automatically picks up on your current Mac focus is exactly how an assistant should work

how does Invoko manage permissions across different apps while maintaining privacy and security?

Hey PH 👋 Sachin here, jumping in with the engineering side for anyone curious.

Each employee is a Kubernetes pod with its own filesystem and its own postgres-backed memory. State survives restarts and channel disconnects. Even a full redeploy. We chose this over a serverless function model because we wanted the agent to be a process you can talk to at 3 AM and have it remember the conversation from yesterday morning. Cold starts and stateless containers kill that.

For app integrations we run a hosted MCP layer at the workspace level. You OAuth once per app (GitHub, HubSpot, Zendesk, Jira, Asana, Airtable, Gmail, Outlook, Calendar, more) and any employee in your workspace can use the integration after that. Each MCP server is managed on our side, so there's no JSON config or token paste on yours. No re-auth per agent.

For configuration we built a chat-driven onboarding flow. Instead of filling forms to set up the employee's role, tone, and context, you talk to it. It asks the questions, you answer, it writes its own config record. The mental model is onboarding a remote hire rather than setting up software.

Skills are modular blocks of context the agent retrieves dynamically when a task needs them. Private skills stay scoped to your workspace. Public ones get installable by any maker. So the agent isn't carrying the whole brain on every call. It pulls the right context for the work at hand.

Two runtimes:
- PicoClaw runs on Python. Lighter, supports email channel and cron scheduling.
- Moltis runs on Rust. Heavier. Web dashboard, browser automation, voice (15+ TTS/STT providers), CalDAV.

Happy to go deeper on any of the architecture. Ask me anything.

AI employees living right in Slack and Teams sounds really useful. Can each AI employee be customized to a specific role or task?

the 'ai employee' framing is getting crowded fast - so many tools launched this month doing the same slack/teams bot thing. the per-call pricing on tools is the bit that'll catch people off guard when an agent loops or retries unexpectedly. what actually differentiates the memory layer here vs just wiring up a standard agent with a slack connector?

The per-call model is a smart call, pay for what your agent actually does is way easier to justify than another monthly subscription. Personal assistant template first for me. Would love to see a research analyst role next

Giving each employee its own pod with persistent postgres-backed memory that survives restarts is the detail that won me over, "remembers yesterday's conversation at 3 AM" is exactly where most chat-widget agents fall apart. On the pay-per-call model: do users get spend caps or alerts per employee, so a retry loop on a tool can't quietly rack up cost overnight?

Congrats on the launch!

The "AI employee in Slack/Teams" framing has been tried a few times but the pricing always trips it up — per-seat feels wrong for a non-human. How did you land on your model?

(Indie maker here, curious about the pricing call more than the product itself.)

This looks really cool, congratulations on launching! what's next for your roadmap?

Congrats! I'm a little curious how does MakersClaw handle context continuity when the same user interacts across Slack and Telegram simultaneously?

the framing of 'AI employees that live in slack' is sharp. the part i'm most curious about: how does an AI employee accumulate a track record? a human gets references from former teammates, a paper trail of what they shipped. when this AI moves between teams, what travels with it?

Super cool idea. I think the pay-per-call tool model is great.

AI employees living where the team already talks is the right move. The hard part is making them helpful without becoming one more coworker to manage.

AI teammates living where you already work, rather than another dashboard to check, feels like the right direction. Bookmarking this one.

The "configure by chatting with the employee" concept really caught my attention. In practice, do users prefer conversational setup over traditional forms and settings?

Congrats on the launch! The chat-driven onboarding is interesting. How are you validating role configuration flows when users give ambiguous or conflicting instructions to an AI employee? I imagine those edge cases could be challenging to test across multiple integrations.

Hey Product Hunt 👋 — I'm Siddhant, co-founder of PeakRoutine.

Two years ago I was pre-diabetic at 32, staring at an Apple Watch full of data and still had no idea what was actually making me sick. Turns out it was my sleep. No app told me that. That's the thing nobody tells you: most health apps are expensive mirrors. They track everything and change nothing.

So we built one that would.
That gap became PeakRoutine. We pull in everything — sleep, sunlight, exercise, calories, nutrition, mood, hydration, HRV, stress — and instead of handing you another dashboard, our AI correlation engine finds the patterns that matter for your body specifically. Not a generic plan. Your biology, learned over time.

Here's what that looks like in practice:

1. You sync through Apple Health, so your Apple Watch, Oura, WHOOP, or whatever you use all flows in automatically.

2. We correlate your biomarkers against each other and surface the relationships you'd never catch manually — the ones that explain why you've been dragging this week.

3. Six AI health coaches answer anything about your sleep, recovery, stress, nutrition, or energy, grounded in your full biological picture.

4. A biology-aware habit engine builds your daily routine around what your body can actually handle right now — not what a template says you should do.

We recently became new parents — which has a way of making you ruthless about long-term health. You stop optimizing for today and start thinking in decades. It also showed us how differently health shows up depending on who you are and what season of life you're in. Women's health and parent-specific experiences are next on our roadmap — not as an afterthought, but because we've lived the gap firsthand.

So we built around your biology, not a dashboard.

If you've ever looked at your health app and thought "so what?" — this is built for you.

We're in early access on TestFlight right now — a small, intentional rollout before our full App Store launch. Install takes 2 minutes. Honest feedback from this community means everything at this stage — the good, the bad, and especially the brutal.

One question to kick things off: do you actually act on your wearable data, or mostly just look at it?

For the launch day we are offering 35% - come and try it out and I am sure you will love it.

AMA in the comments. Let's talk. 🙏

In my experience every other app guilt-trips you over one bad night and kills the streak, so building around what your body can actually handle that day is a real wedge, congrats! I was wondering at n=1 with noisy wearable data, how do you stop it surfacing a coincidence as a pattern, i.e. telling me morning sunlight fixed my sleep when really I just skipped the late glass of wine? Is there a confidence threshold before a correlation becomes advice?

Hey Product Hunt 👋 Maker here.

PeakRoutine started from a simple frustration: health apps drown you in data and starve you of direction. Most of them hand you dashboards and leave you to play scientist with your own body. The goal here was the opposite: less charts, more "here's what to actually do tomorrow."

So it's built around the things the category mostly ignores:

• Per-biomarker AI insights — every metric (sleep, HRV, RHR, sunlight, activity…) gets a plain-English read on how you're doing and exactly what to improve, not just a number on a chart.

• Correlation, not dashboards — it connects sleep, HRV, mood, activity and sunlight and surfaces the hidden links (the most common "whoa" moment: morning light before 10am quietly driving both mood and deep sleep).

• 6 named specialist coaches with memory — Health, Sleep, Fitness, Nutrition, Recovery and Mental Health — all reading the same biomarker data. No generic, one-size AI.

• Sunlight / circadian tracking — genuinely rare in health apps, and a real wedge rather than a footnote.

v1 is biomarker insight + correlation, all pulled from Apple Health (hardware-agnostic). Habit execution, direct wearable integration, and parent mode come next.

Would genuinely love for you to tear it apart: what's the one thing you'd need to see before you'd switch from whatever you use today?

Health coaching based on your actual biomarkers is a great angle. Which biomarkers does PeakRoutine track, and how do you input them?

I wear a whoop band. Whoop already has coaching feature. Would Peak Routine still be useful?

Other than the Apple health data, does it also sync with other wearables?

Congrats on the launch. Is this more of a preventive care app or it's meant for people with chronic conditions or both? How do you envision it?

Congrats on the launch @siddhant_gupta12 !

Why would someone use this instead of Apple Health? Is the biggest differentiator the unified dashboard, the AI insights, or the habit-building system?

Hey Product Hunt! 👋 I’m Sakshi, one of the makers of PeakRoutine.

We built PeakRoutine because most health apps felt like expensive mirrors.

They showed us sleep scores, steps, HRV, stress, and trends — but still left us asking:

“Okay… what should I actually do today?”

PeakRoutine turns your wearable and lifestyle data into simple, personalized weekly routines across sleep, movement, sunlight, mood, energy, and recovery.

No overwhelming dashboards.
No generic habit lists.

Just small, practical actions that help your body feel better over time.

Our mission is simple:

Turn health data into direction.

We’re launching our first version today and would love your honest feedback:

What’s one thing you wish your wearable or health app helped you improve automatically?

Excited to hear what the PH community thinks 🙏

The recovery-aware habit engine feels like the strongest part. Wearables already show what happened, but turning that into small actions your body can actually handle is much harder. Curious how quickly the routine adapts after a few bad nights.

The correlation engine is the part that stands out to me, surfacing relationships across sleep, HRV and sunlight is genuinely hard to do well at n=1. How many days of synced history do you typically need before the AI starts surfacing confident patterns rather than noise?

Finally an app that turns my wearable data into something I can actually act on instead of just staring at. Does the habit engine adapt on low-recovery days, or hold me to the same plan?

How does PeakRoutine distinguish between a real causal signal and a random correlation? What evidence shows the recommendations actually improve outcomes?

Which wearables and data sources do you support right now, and do you rely on Apple Health as the hub?

personalized health coaching from biomarkers is the kind of thing that only works if the recommendations stay coherent over time. how do you handle the case where a biomarker shifts (sleep, HRV, fasting glucose) and your last week's plan suddenly contradicts this week's? does it explain the why behind the change or just hand you the new plan?

Biomarkers plus coaching is way more useful than another dashboard yelling at me. How do you keep the advice simple enough to actually follow?

The biomarker-driven angle is what makes this stand out from generic coaching apps — most personalize on goals, not actual physiology. Curious which wearables/inputs it pulls from out of the box?

Looks really interesting Sakshi - using biomarker data to drive coaching recommendations is a much more honest approach than generic advice. What biomarker sources are you currently integrating with? (Oura, Apple Health, blood panels?) - I have a friend of mine that was working in sth similar...

Hey Product Hunt 👋

Sacha here, co founder of Edgee.

Story time. A few weeks ago I was working with Claude Code on a refactor with Opus. The model knew exactly what to do, but I sat there watching a 500-line file crawl out one token at a time. Two minutes for one file. Multiply that by every step of the agent loop and you realize: speed is the silent tax on every coding session.

Around the same time I started testing open-source models like GLM and Kimi K2.7. The quality on coding tasks was honestly impressive.

But the speed on standard endpoints was even slower than the closed models. And the setup was painful: API keys, code changes, CLAUDE md to rewrite, MCP servers to reconfigure.

That's the problem we built Edgee Turbo Models to solve.

What it does:

→ Run frontier open-source models (GLM 5.1, Kimi K2.7 Code, Kimi K2.6, MiniMax 2.7) directly in Claude Code.

→ At up to 4x the speed of standard endpoints (~200 tok/s vs ~50).

→ Flat $29/month. No metered token bill that climbs as your agents work harder.

→ Setup in 2 minutes. Your CLAUDE md, MCP servers, and entire setup stay exactly where they are.

Important point I want to get out front because it'll come up:

Turbo is NOT a smaller or quantized version of these models. They are the full open-weight checkpoints. Turbo only changes how they are served, on dedicated high-throughput inference infrastructure built for raw speed, not a shared best-effort endpoint. Same outputs, just faster.

How this fits with our previous launches:

- Compression: use fewer tokens per request

- Teams: see who uses what, per repo, per PR

- Fallback Models: keep working when Claude or Copilot hit limits

- Turbo Models: run open-source models at premium speed, for flat pricing

Together that is the Route + Compress + Observe stack of our Agent Gateway. Today we're shipping the speed layer.

Why now: The Economist published a piece this week confirming that "token-maxxing is over" and that companies are routing to cheaper models. Open-source models are clearly part of the answer. Turbo

makes them actually usable.

A few questions I'd love your feedback on:

→ Which open-source coding model are you most curious to try?

→ Is flat $29/month the right price point, or would you prefer usage-based?

→ What other models should we add to the Turbo lineup?

Will be in comments all day. Thanks for checking it out 🙏

Being able to run different models through Claude Code is really cool. Can you switch between models mid-session, or is it set per project?

Are you planning to add more models to the Turbo lineup? Curious specifically about DeepSeek V4 and Qwen 3.5 Coder. Also, will Turbo ever work with Codex or just Claude Code?

flat $29/month instead of usage-based is the right call for anyone running agents that loop unpredictably. the worst part of token-based pricing is never knowing what the bill will be until it's too late. also being able to swap in open-source models without changing any code or rewriting configs removes the biggest barrier to actually trying them. most people stick with what they know because switching is painful, not because alternatives aren't good enough

Proxying Claude Code's API calls through a gateway to route to Kimi K2.7 or MiniMax without code changes is clean architecture. We've hit throughput ceilings in agentic workflows where task latency compounds fast, so the 4x speed claim is interesting. Does Edgee handle automatic fallback if a model hits rate limits mid-session?

Love the flat rate approach for unpredictable agent loops, excited to test Kimi K2.7 Code with this kind of speed. Huge congrats on shipping this, @sachamorard

Really great product. Our team can have a good usecase with the product. Already filled the contact us button. (poko.video email ID). Lets schedule a meet!

Using Edgee with Kimi K2.7, huge savings comparing to OpenRouter. Keep going guy!

This is exactly the kind of model switching devs pretend they do not need and then use 12 times a day. Love the practical angle.

The edge angle I always want clarity on: cold-start and state. For these turbo models running at the edge, are you keeping them warm across regions or is there a first-hit penalty when a PoP hasn't served the model recently? And for anything stateful, how do you reconcile across PoPs without round-tripping to a central region — or is the model purely stateless inference?

How do GLM 5.1 and Kimi K2.7 Code actually compare to Sonnet 4.6 on real coding tasks? I've tried open-source coding models a few times over the past year and they always felt one tier below the closed ones. Has that gap really closed?

Congrats on the launch, Sacha! Curious about how does Edgee handle consistency during peak demand when multiple teams are hitting the same Turbo endpoints simultaneously?

token reduction at the gateway is the kind of infra that quietly changes the economics of agent workflows. curious about the tradeoff curve: at what compression ratio do you start seeing measurable accuracy loss on downstream tasks? and is that loss uniform across model families or does it hit some harder than others?

So I can create apps with this new Edge Turbo Model?? An also a standard fee 29.99 ? No more token manipulation?

Hey Product Hunt! 👋 Excited to finally share this with the world.

Honest origin story: we used other support platforms for years, and we were constantly frustrated - bloated UIs, slow as hell, features buried under features, and pricing that made no sense for a startup.

So we built SparrowDesk the way we always wished those tools worked.

The centerpiece is Zoona, our AI agent that actually handles support end-to-end - not just suggests replies, but resolves tickets. Across Slack, WhatsApp, chat, email, and your help center. One setup, everywhere your customers are.

Here's what makes Zoona different:

🧠 Train from anything

Connect your help articles, URLs, documents, Notion pages — Zoona learns from all of it and stays up to date as your product evolves.

⚡ Commands that take action

Zoona doesn't just answer questions. Use commands to make your AI agent trigger API calls, update records, create tickets, and take real actions inside your stack.

✍️ Copilot for your team

Get AI-suggested replies, instant conversation summaries, and a copilot that helps your agents close tickets faster — without burning out.

📚 Self-healing knowledge base

Our copilot analyzes resolved conversations and automatically rewrites + improves your knowledge base. Your docs get smarter every day without anyone lifting a finger.

We built this for startups who want enterprise-grade support without the enterprise-grade pain.

And if you're a startup - we have a special program just for you: sparrowdesk.com/startup

Would love your honest feedback — the good and the brutal. Ask me anything! 🙏

The UI looks great and the timing is interesting too. With Fin heading into Salesforce, a lot of teams are about to re-evaluate their support stack, and a focused agent like Zoona is well placed to catch that wave. Strong first impression. All the best with the launch.

Spent some time this week red teaming support chatbots,

Sparrow's ZoonaAI held up to everything. Asking it to repeat its instructions, rephrasing the request a dozen ways, sneaking it through a translation, getting it to "autocomplete" its own prompt , nothing worked. Genuinely well built. Congrats to the @SparrowDesk team, and @ayesha_kulsum_s_j . 👏

Then there was Zomato's bot, which made me laugh.

I'd spent ages trying to reach a human for a delivery issue ,it kept getting told all agents were busy. So on a whim I asked for its system prompt.

It transferred me to a human instantly.

It's funny that the magic word for reaching a human wasn't "human" five times over. It was "what's your system prompt."

Support bots are getting harder to crack than they used to be.

@georgejustin22 A lot of support tools promise faster responses.

What’s more interesting here is the claim that customers never have to repeat themselves when a human takes over.

In practice, that handoff is where many AI support experiences fall apart.

How much context is actually transferred to the human agent, and have you measured whether customers feel the transition is seamless versus just receiving a conversation summary?

An AI-first help desk that cuts down on searching sounds great for support teams. Does SparrowDesk pull answers from your existing docs and tickets?

Tried Zoona for a bit and the first thing that stood out was how polished the experience feels. The inbox workflow is intuitive, and the AI seems focused on actually helping resolve tickets rather than just being a flashy add on. Curious to see how teams adopt it over the next few months. Congrats on the launch!!

Been using SparrowDesk and it's been impressive so far. Clean UX, useful AI capabilities, and a strong focus on productivity. Excited to see where SparrowDesk goes!

@kevin The product makes sense, but I wonder if the future customer support team even

looks like today's support team.

If Al ends up handling most inbound requests, do you think support eventually becomes a product and operations function rather than a customer service function?

Curious how your customers think about that shift.

Just checked out Zoona AI and the setup experience genuinely surprised me! No complex configurations, no developer dependency- it felt like something any support team could get running on day one. This is the kind of simplicity that support teams would require. Great work, team!

Congrats on the launch. The past-conversation learning part is the interesting bit for support teams. How are you handling the boundary between what Zoona can resolve automatically and what needs human approval or owner sign-off?

I tried SparrowDesk and was impressed with the overall experience. The design is clean, intuitive, and easy to use, making ticket management simple and efficient. Beyond the great user experience, the product feels fast and responsive, even when handling a large volume of tickets.

It’s clear that a lot of thought has gone into both the design and the underlying performance.

How do you handle rollback or recovery when an AI-initiated action goes wrong mid-ticket?

the help desk space is brutal because the value isn't in answering the easy 80%, it's in not breaking the hard 20%. curious how you decide when to confidently resolve vs route to a human. is the threshold the same per category (billing vs technical vs account) or are you tuning it per tenant once you learn their support team's tolerance?

How would this build on top of existing support platform data? Can you import from zendesk etc. to bring in historical cases and how they were closed? Also can it use a human's response as info to automate any future tickets with similar issues?

I tried SparrowDesk and I fell in love with it, intuitive UI and clean layout, fast and responsive. The AI assistant is very powerful and it makes it easier to be productive

@georgejustin22 The self-healing knowledge base sounds powerful, but also slightly risky.

If Zoona learns from thousands of customer interactions and continuously updates documentation, who ultimately owns the truth: the documentation team or the Al?

Feels like there could be interesting situations where the Al discovers patterns and solutions before the company officially documents them.

J'ai pu tester Zoona récemment et franchement, ça m'a agréablement surpris. L'interface est agréable à utiliser, la gestion des tickets ne demande pas de prise en main particulière, et l'IA est vraiment là où on en a besoin pas juste pour faire bien sur la fiche produit. Vivement la suite pour voir comment les équipes vont s'en emparer. Beau travail pour ce lancement !

The "learns from past conversations" angle is underrated. Most support bots reset context every time. Does this work across channels like WhatsApp or just web widget? Congrats on the launch!

Learning from past conversations is the bit that matters. Docs alone miss all the weird customer phrasing where support actually happens.

Learning from past conversations is the key part here. Docs alone rarely cover the messy edge cases in support. Curious how much review control teams get before Zoona starts replying live.

@georgejustin22 Resolving 60%+ of tickets automatically sounds impressive, but I’ve always found the hardest support tickets aren’t the repetitive ones.

They’re the edge cases where a customer explains something in an unexpected way, where the issue spans multiple systems, or where policy and judgment matter.

What types of tickets have you found Zoona consistently struggles with today, and where do you still believe humans are significantly better?

Gave Zoona AI a try and it really stands out. The UI/UX is clean, intuitive, and genuinely easy to use from the first click. No fumbling around to figure things out, which is rare. The AI responses also felt useful right out of the gate and the setup was way easier than I expected. Looking forward to exploring more. Excited to see how Zoona evolves. Congrats on the launch!

Hi Product Hunt! 👋

I’m Olli-Pekka, one of the co-founders of GitHits.

I've been a member of the open-source community for 15 years. I created opencv-python, which got 100M+ downloads while I maintained it as a side hustle. Fun fact: I’m from 🇫🇮, just like Linus Torvalds. 🙂

I noticed I kept giving the same advice to colleagues and friends when the docs were missing something. My go-to hack was simple: use GitHub search to find code that already solves the problem. It’s powerful, even though it only returns raw results rather than the answer in context.

I started building GitHits to bring that workflow to coding agents.

GitHits complements tools like Claude Code, Codex, Cursor, and other AI coding agents.

Those agents are great at navigating your local codebase. They can grep, search, and read files to understand how your application works.

The problem is that modern software doesn't stop at the repository boundary.

A large part of the system lives in frameworks, libraries, SDKs, and other open-source dependencies. Agents can usually see where your code calls into those dependencies, but they often can't navigate and inspect them in the same way. And even when an agent reads the docs, they only tell it what to call, not how it actually behaves. For that, you need the source.

GitHits gives agents access to:

• Code examples based on real implementations from repositories, issues, discussions, and pull requests, linked back to the implementation code

• Code navigation across packages and repositories: search, grep, file listing, and exact line reads without cloning

• Package inspection for dependencies, vulnerabilities, changelogs, and upgrade changes

• Documentation access across hosted docs and repository-backed docs

GitHits does this by building a version-aware index of open-source code on demand, usually in 10-20 seconds for an average repository.

GitHits is useful when an agent reaches the limits of the local repository.

That might happen during planning and research, when it needs to understand how a dependency works, what changed between versions, or how something has been implemented elsewhere. It also happens during implementation, when the agent starts retrying variations and exploring dead ends because the answer isn't in the local repository.

As developers, that's usually the point where we leave our own repository and start reading somebody else's.

What users say about GitHits:

• Onni: Helping Claude Code to find undocumented APIs from the code

• Rob: had a problem in my openusage v2 app - boom, perfect fix without guesses

Forever free tier available.

No trial period. Just create your account and connect your agent to GitHits.

Launch day special.

Everyone who signs up today gets 3x credits for 6 months. No strings, just more GitHits.

Setup is one command:

npx githits@latest init

It installs the CLI and connects GitHits to Claude Code, Codex, Cursor, or any MCP-compatible agent. Or sign up using this link.

If you’re already using GitHits, let me know what you use it for, and how we can make GitHits even better.

We look forward to your brutally honest feedback and sincerely appreciate the support!

Hi Product Hunt! 👋

I'm Juha, co-founder and chief architect of GitHits. I'm responsible for the indexing engine underneath, so I want to talk about the part you don't see.

The question I get more than any other is some version of this: my coding agent can already clone a repo, grep it, and search GitHub, so what is GitHits actually adding?

It's a fair question, and the honest answer is that scraping GitHub gives your agent raw results, not understanding. It clones whatever is at the top of the default branch and greps one repo at a time, treating code as text with no sense of what is worth trusting. And here's the part that surprises a lot of people: GitHub's own code search only covers a repository's default branch, so the older version you actually depend on usually isn't even searchable. That's fine for a quick lookup, but it falls apart the moment the problem gets specific.

What GitHits does instead is build a real, version-aware index. We fetch the actual source, parse it, turn it into structure, and keep that structure version by version, so your agent reads from something we already built rather than scraping and guessing in the moment.

People are usually surprised how much work that is. One small example: just figuring out which commit a version like 1.2.3 actually points to. There's no standard for how projects tag releases, so the part of our indexer whose only job is that is more than 700 lines, and we've rewritten it six times as we keep hitting new conventions in the wild. And that's before we've parsed a single line of the actual code.

And it isn't only code. We index documentation the same way, version by version, and combine it with the rest so your agent grounds itself in a balanced mix of code, tests, examples, and docs, instead of one slice it has to guess from.

I wrote up the full under-the-hood walkthrough here if you want the long version: https://githits.com/blog/what-it-actually-takes-to-index-open-source/

The best way to judge it is to just test it out. Setup is one command:

npx githits@latest init

It connects GitHits to Claude Code, Codex, Cursor, or any MCP-compatible agent. Or sign up here: https://app.githits.com/signup

I built the indexing engine, so I'd genuinely love to hear where it works or breaks for you. Brutally honest feedback is the most useful kind.

This app is seriously helpful and turns your coding agent in a top tier senior engineer. When I am introducing a new pattern or library in my code, I always think to find real battle tested implementations first from GitHits. Eventually you realize just letting a coding agent vibe out consequential and high leverage code is just plain dumb and irresponsible.

Really impressive backend work. Are you guys running the parsing/indexing entirely on your infra and serving it via API, or is there a local caching layer involved when we link it to Claude Code via MCP? Happy to support 🙌

Building a version-aware index per release rather than scraping GitHub's default branch is the right call. Agents that can't inspect the exact version they depend on just hallucinate API signatures. We've hit this when agent-generated code breaks because docs described the wrong version. How does GitHits handle monorepos with independently versioned packages?

Works great, I've been using Githits to explore implementation details from different libraries to ground my coding agent with real world examples. It's cool to point your coding agent towards an open source reference implementation that you know has already implemented what you want to implement.

giving the agent access to open source code is the easy half. the hard half is helping it pick the right code. github has incredible code and also a long tail of broken half written experiments. how do you weight signals like maintained recently, used by many, tests pass, vs raw similarity to what the agent is trying to write?

Giving agents real OSS context feels like a cheat code. Curious how you rank which repos are signal vs random GitHub soup.

The dependency-context angle makes a lot of sense. Docs explain the API, but source code and real examples show how it actually behaves. I’d be interested to see how GitHits weighs tests vs production implementations when finding examples.

Congrats on the launch!

The "AI agent + OSS context" gap is real — most agents either hallucinate APIs or refuse to commit when they're unsure. Giving them grounded access flips that.

Question for you: how do you handle licensing exposure when the agent pulls from copyleft repos and the user is on a proprietary codebase? That's the piece I'd worry about if I were shipping this into a regulated team.

Giving a coding agent access to real open-source code is a smart idea. How does GitHits pick which repos or snippets are most relevant?

Congrats!

Hey Product Hunt 👋

I'm Kunal, the founder of Stride.
https://www.stride.page/

Here's the moment that made me build it. I was "planning a feature" and counted the tabs open to do one job: a board for the tickets, a whiteboard for the diagram, a doc for the spec, a tracker for status, and three AI chats I kept re-explaining my project to from scratch. None of them talked to each other. I was the integration layer. And I was exhausted.

So we built Stride: one AI-native workspace for the whole journey from idea to shipped.

📋 Plan — a flexible board with custom stages, WIP limits, and issue tracking that bends to how your team actually works (not the other way around)
🎨 Design — architecture diagrams, solution design, and PRDs, drafted and refined with AI right next to the work
⚙️ Optimize — map, model, and mine your processes to see how work really flows and where it gets stuck
✅ Verify — close the loop on quality: define acceptance criteria, build test plans, validate that what ships actually matches what you planned, and catch gaps before they reach users
🤖 Agent — an AI teammate that lives inside your real project. It creates, updates, and moves work for you, and plugs into Claude Code and Codex over MCP
⚡ Ship — go from idea to PRD to shipped without ever leaving the app

The thing I'm proudest of: the AI isn't a bolt-on chatbot staring at a blank box. It sits inside your actual project data, so it already knows your tickets, your stages, and your context. It does the work instead of just talking about it. Less "write me a prompt," more "handle this."

We're a small team and every comment today genuinely shapes what we build next, so I'm parked in the thread all day. One question I'd love your honest answer to: what's the one tool-switch in your workflow that makes you sigh every single time? Plan to design? Spec to tickets? Reply and I'll tell you exactly how (or honestly, whether) Stride kills it for you.

Thank you for being here. It means a lot. 🙏

An AI workspace that goes from planning all the way to shipping is a neat all-in-one approach. Does Stride integrate with tools like GitHub or Figma?

What used to take days of planning, grooming, and coordination can now be completed in under a minute. 👏

I’ve been using Stride heavily for AI-assisted full-stack development, especially with Codex and Claude Code through MCP, and this is where it really clicks.

The biggest value for me is the flow from product thinking to execution: PRD → epics → stories → acceptance criteria → test cases → implementation.

I can then hand that work off to an AI coding agent that has the full project context through MCP. That means the agent is not starting from a vague prompt. It knows the ticket, expected behavior, acceptance criteria, related test cases, and the current workflow state.

In practice, this makes the dev loop much tighter. Claude Code or Codex can pick up a ticket, work through the implementation, use the acceptance criteria as the target, run the relevant test cases, and move the ticket from To Do → In Progress → Ready for Review.

What I also like is that the workflow does not stop when implementation is done. Once a ticket moves into review, Codex can continue updating the work through MCP: adding comments, attaching relevant context or outputs, recording what changed, and keeping the ticket useful for the next person reviewing it.

For anyone building with AI coding agents, this solves a very real problem: the gap between “we planned the work” and “the agent actually has enough structured context to build the right thing.” Stride gives that context a home, and MCP makes it usable directly inside the development workflow.

If you’re doing AI-assisted development with Claude Code, Codex, or similar tools, I’d definitely recommend trying Stride.

🎁 Early-bird discount for the Product Hunt community: use code AAYUSH10

The "plans, designs, and ships" framing covers a lot of ground, and the interesting question is where the handoffs happen. Most tools like this are solid at one of those three and then quietly hand you back the wheel for the others. Curious whether Stride is actually driving the design-to-code transition itself, or whether "designs with you" means something closer to a Figma-adjacent whiteboard that you then feed into the build step. Also wondering how the Vercel tie-in works in practice: is deployment genuinely wired into the workspace so shipping is one action, or is it more of a pre-configured export target?

The interesting part is keeping planning, design, and shipping in one loop. Most tools nail one of those then lose the thread.

I've had a front-row seat watching Stride come to life, and one thing that constantly stood out was how much time teams lose simply moving context between tools.

One doc becomes five tabs. One requirement becomes ten conversations. And before you know it, half the effort is spent reconnecting information instead of building.

Seeing Stride evolve from an idea into something that actually keeps planning, design, engineering, and QA connected has been incredibly rewarding.

I'd genuinely love to know: what's the one context switch in your daily workflow that frustrates you the most?

Every answer helps us build a better product. ❤️

plans, designs, and ships are three very different modes. plan is about exploring options, design is about constraint, ship is about commit. curious how Stride keeps context coherent across these without flattening them. does it know when to push back on something planned because the design phase showed it's not feasible, or does it just keep building?

Great platform! The ability to generate test cases directly from requirements and maintain traceability throughout the development process has been particularly valuable for our team. 👏

I like the framing that founders shouldn’t just become faster at doing email. The real unlock is probably seeing only what actually needs a decision, and letting the rest move without constantly pulling you away from building.

Curious how Dirac decides what needs the founder’s attention and what can safely be handled in the background. Is that something users train manually over time, or does it learn mostly from inbox and sent email behavior?

Also, the “what would you always keep for yourself?” question is a good one. For me, anything related to partnerships, sensitive customer issues, or important founder-to-founder communication should probably stay human.

A morning briefing from your inbox is such a useful idea for busy founders. Can you customize what the daily brief focuses on?

I like the framing of “only what needs your decision.” Most inbox tools still assume you want to process everything, while in reality most emails are just noise

#9 and 82 points on launch day — how'd you pull that off straight out of the gate? Pre-warmed audience, community seeding, a strong hunter? Genuinely trying to learn what a clean PH launch takes these days.

Hey Product Hunt!! 👋 I'm Peter, the Maker of Dirac.

On my last startup we were a tiny team with barely any traction. And I was still losing about a fifth of my working time to email. Cold outreach, notifications, prospects, internal threads. Volume wasn't the real issue.Everything funneled through me, and most of it didn't need me. But I had to read every message just to know which was which, or risk something important slipping through.

That's the part I kept coming back to. Founders aren't bad at email. They're stuck being their own assistant.
Every tool out there makes you faster at doing email. Superhuman, Hey, Gmail, all of them. But being faster at the wrong job is still the wrong job. Email pulls you out of building all day, and the best tools only make that a little more bearable.

🚀 🚀
So I built Dirac. An AI-native inbox where the AI does the busywork, not you.
Every morning it hands you a brief with only the emails that actually need you. Everything else gets sorted, filed, and handled in the background. Tell it to reply and it drafts in your voice. It already learned your tone from your sent mail. You show up to email already knowing what matters and what to say.
🎯
That's the difference. Every competitor is AI bolted onto a traditional inbox. Dirac is agentic. You make the calls, it does the work.


It's live at https://dirac.app if you want to try it. 14 days free, no card needed. But mostly I want to hear from you. What's the one email task you'd kill first? And what would you always keep for yourself?

morning founder brief is a beloved category that everyone underdelivers on because importance is so personal. curious how Dirac learns my actual signal of importance. is it from explicit feedback (i marked this as relevant), implicit signals (i replied within an hour), or a blend? and how long until the brief stops feeling generic?

An inbox that briefs me before I open the inbox is the dream. Founder email is basically a boss fight with subject lines.

Congrats on the launch!

The "morning brief" framing makes more sense to me than "inbox AI" — most founders I know don't have an inbox problem, they have a "what should I care about first" problem.

Does Dirac pull from anywhere outside email (Linear, GitHub, Sentry, etc.) for the brief, or is email the only signal source for now?

As someone who spends way too much of her morning triaging email instead of actually building, this hits close to home. The "drafts replies in your voice" part is what I'm most curious about. Does it actually learn your tone over time or is it more of a one-time setup?

Congrats on the launch, rooting for this one!

Update: We have a yearly coupon too:
TRYDIRAC25

use wisely...

One concern I have is overinterpretation. users might treat simulated neural outputs as scientific truth, so clear framing and limitations will be important.

the ositioning is bold, but also risky. When terms like “read minds” are used, expectations can easily go far beyond what simulated neuro-data can realistically provide.

Hey Product Hunt community! 👋

Thrilled to see this first launch from the Cassini Research collective hitting the top 10 (#9 right now! 🚀).

MindReader V1 is the result of deep research and countless hours collaborating with sales teams to solve real workflow bottlenecks. The best part? It’s already driving impact. We're currently being integrated into the evaluation pipelines of a marketing team and a YC-backed sales AI agent.

Huge shoutout to @ishita8088 for building something users are clearly loving. As a co-maker, I’d love to get your thoughts - where would you like to use it in your product or agent pipeline?

@ishita8088 Hi, Tried the product at the trial setup page and honestly the experience felt like stepping into a neuroscience lab 😄 The visualizations are fascinating and definitely spark curiosity. At the same time, I found myself wondering how much of the report reflects real cognitive signals versus an interpretive model. Either way, it's a very memorable experience and a fresh way to think about message analysis.

@jas_jaski @ishita8088 Bold vision here. The open-source angle makes it even more exciting for people who love tinkering and experimenting.

simulated fMRI is an interesting framing because it lowers the barrier from clinical setting to anywhere with a laptop. curious which use cases you're seeing pull on this first. is it more researchers doing prototype experiments before booking real scanner time, or builders putting brain inspired models inside consumer apps?

congrats on the launch! super interesting ideas. I wonder what you use as a proxy for attention. Do users need to give inputs or do you estimate where the users might be focusing on based on moment-to-moment 'salience'?

This is weird in the best Product Hunt way. Simulated mind reading for UX feels half research lab, half startup fever dream.

Simulating how a brain reacts to content is a fascinating concept, and love that it's open source. How accurate are the neuro-metrics compared to real fMRI studies?

Very interesting, and great that it's open source. But I'm not sure I understand it correctly. So the goal is to determine how a demographic will respond to certain sales call scripts or ad creatives?

The report is what makes this interesting to me. For mobile E2E, I care less about a bare pass/fail and more about seeing why a run failed. When writing natural-language steps, how specific should they be to avoid brittleness across small UI changes?

I literally use Revyl everyday to test regressions in our app as I'm adding new features everyday with Codex. It's such a good tool for peace of mind. Congrats on the launch :)

Have used Revyl to help me vibecode some iOS apps for fun, super fast and just makes testing so easy, the new maps feature highkey really helps for all the onboarding pages we have to build now.

Hey Product Hunt👋,

Revyl is the mobile source of truth.

There are 3 main parts of revyl:

- The platform captures the full run so you see why something failed

- CLI allows you (or your coding agent) to drive cloud device from terminal
- Atlas auto-maps every screen and flow in your app, a live source of truth instead of a stale diagram


Would love your feedback from anyone shipping mobile.

'the mobile source of truth' is a positioning that can mean a few different things. is the source of truth about app state in production, about which build is live across cohorts, or about user behavior aggregated across platforms? curious where you've landed and whether mobile teams are buying this as observability or as control plane.

This is super sick. Atlas is probably one of my favourite features

Driving real cloud devices from the CLI is the part that stands out to me. It feels especially useful for mobile teams trying to bring testing closer to the dev loop. How does it fit into CI today?

the file-system diffing across runs is the bit i haven't seen anywhere else - catching state that quietly accumulates between sessions is genuinely hard to debug otherwise. combining that with the full network waterfall on cloud devices is a strong combo for the 'only happens in prod' class of mobile bugs

Ran this on my 512GB MacBook and it dug up 110GB of dev junk. Android Studio alone was hoarding 52GB. On a machine that small with a dozen projects open, that's basically the gap between shipping and a "disk full" popup at the worst moment.

And with what SSD upgrades cost these days, clawing back 30-100GB whenever I need it beats paying Apple for the next storage tier. Good tool.

One bit of feedback: I'd love more granularity inside the categories. Ollama's the obvious one — it shows 22GB total, but I don't want to wipe all my models at once. Let me open it up, see each model that's installed, and drop just the ones I've stopped using. Per-model control instead of all-or-nothing would make me a lot more comfortable hitting Clean Now.

Wow, I had no idea AI apps were hoarding that much space! Cursor and Ollama have definitely been eating up my disk. Super timely tool, congrats on launching....

Fantastic job! I didn't know about this problem until I've try it. And its HUGE. Thank you that you built DevCleaner!

Every dev laptop eventually becomes a landfill with a keyboard. Love the idea of cleaning hidden GBs without me playing detective.

This is genuinely one of those problems everyone has but nobody thinks to solve. How much space does the average user recover?

This is super useful, as a software engineer (and a vibe coder :P) i'm ramping up so much GBs that sometimes my top tier macos pro even hangs and asks me to free up space, turns out alot of these GBs are build artifacts! thanks OP, upvoted.

AI for the boring YouTube ops is the right lane. Let creators keep the taste, let the machine fight the metadata swamp.

It hurts my heart to see ai + youtube...

But I mean... as long as the content itself is good, right? I hope the future isn't flooded with Ballerina Cappucina :)

earning more from existing creations is the part that flips the usual creator workflow on its head. most tools tell you what to make next. this looks at what's already there. curious how Lasso decides which old content has affiliate fit and which doesn't. is it semantic match against the affiliate catalog or are you scoring engagement patterns first?

Hey Product Hunt,

Over the last few years, we have focused on making Chatwoot a solid product for text-based customer conversations across live chat, email, WhatsApp, social channels, help center, automation, and AI. Once those workflows became mature, it made sense to bring voice into the same experience.

Today, we are introducing voice calls in Chatwoot.

You can receive and make calls from the browser, see customer context before picking up, and keep recordings, transcripts, summaries, and call history in the same conversation thread. It works with Twilio Voice, and we’re also working on WhatsApp calling.

This is part of a broader direction for Chatwoot: Chat, email, WhatsApp, voice, AI, and customer context should all live together.

We’d love your feedback, especially from teams that still rely on phone support or are trying to bring voice into a modern support workflow.

Happy to answer questions here.

Calls inside the same support inbox makes so much sense. The worst support tax is context living in five different places.

The voice layer inside the same conversation thread is the part I’d test first for support ops, especially with recordings, transcripts, and AI summaries attached to the existing customer context. For launch/community support, the edge case is a call that starts in WhatsApp and then needs a teammate to follow up by email. Does Chatwoot keep that handoff in one thread with the call summary visible to the next agent?

adding voice on top of a chat first support stack is exciting because it forces a question about session continuity. when a customer calls after starting in chat, does the voice agent inherit the full transcript context or start fresh, and how do you avoid the customer having to repeat the problem? curious how you've structured that handoff.

The thing I never solved is the alt-tab tax... I kick off a Claude Code run, switch windows, then keep flipping back to see if it's done or just stuck waiting on me. Does it flag when Claude's blocked on a question vs still working? That gap is what actually eats my evenings.

Sessions sitting blocked for 20 minutes because I forgot is exactly my problem too, except I'm on Windows so I can't try this. The "waiting for input" priority view is the feature I'd actually want most — that's the moment that costs the most time. Hope this comes to other platforms eventually.

Most of my claude code sessions run in the background while i'm doing something else. watching a video, handling emails, whatever. glint is the thing that actually tells you when it needs your attention without you having to go check. the subagent view is a nice bonus too - parallel sessions get chaotic fast. been using it every day since i stumbled on it

Glint roadmap already includes building support for other providers like Codex, Kiro and etc.

Claude Code activity surfaced where i'm already working is the kind of utility that becomes invisible in the best way once it's there. curious what the most common surfaces have been so far. is it editor adjacent (sidebar, status bar) or more in the team layer (slack, pr review). and is the activity feed read only or actionable from inside it?

Claude Code in the menu bar is exactly the tiny status layer I want. Half my agent anxiety is just wondering if it is cooking or asleep.

Core Features

• Brief-to-Creative Pipeline — input brand details, products, target audiences → generates on-brand ad variations

• Multi-channel Export — Meta, Google Ads, TikTok, Pinterest, email, landing pages (PNG/WebP)

• Visual Styling — Bold & Vibrant, Muted & Editorial, Soft & Minimal, Warm & Playful

• Performance Learning Loop — tracks ad performance, auto-shifts toward converting styles

• Product Integration — syncs with Shopify and WooCommerce

the 10x without a design team framing is what every growth team wishes for. the hard part is usually brand consistency once volume goes up. curious how Tadka holds the brand together across 100 variants. is there a human in the loop reviewing tone and color before they ship, or do you encode the brand once at setup and let the model self police inside those bounds?

Ad creative is one of those jobs where speed matters more than everyone admits. How do you keep outputs from looking like the same AI ad soup?

Congrats! What kind of creative can it creates? static visuals? videos?

I attempted to run the process expecting it to generate eight different variations of my image, but only one was created. I also encountered some bugs during generation. For instance, while I can upload multiple images, the server returns an error for more than 4 images I think. Additionally, I can add images but cannot remove them, after accidentally adding the wrong image, I was forced to restart the entire process to upload the correct ones. Also, review the payment modal the padding is missing.

The brief to creative gap is where most teams lose time. Does Tadka pull from competitor ads at all or purely original

generation?

Hey everyone, Nawneet here, maker of Tadka.

Real reason I built this is because I kept watching good campaigns stall. Not because the targeting was off, but because we could only ship 3 or 4 creatives a week. Meta and Google want volume, and a small team can't feed them fast enough. Hiring designers for that is slow and expensive.

So Tadka takes one brief and turns it into hundreds of on-brand variations in minutes. Then it tracks which ones convert and leans into those styles.

Would love your take. What's your biggest creative bottleneck right now: volume, testing speed, or staying on-brand?

Hi Product Hunt,

I’m Michail, a solo founder near Prague. I built Kraina because I wanted outdoor maps to feel more alive.

Most activity maps show where you went. Kraina shows what you reveal.

You move through the real world, clear the fog, close loops, and reveal the land inside. With Rolling Fog, places you stop visiting can fade back into fog, so the map gives you a reason to return instead of always repeating the same route.

Kraina is not really about performance tracking. It is more about exploration: what you opened, what you lost, and where you might go next.

It is now available on web and iPhone, with missions and challenge zones to help you explore with purpose.

I’d love to get your feedback:

• Is the loop-based territory mechanic clear?

• Does Rolling Fog feel motivating or too punishing?

• What kind of missions would make you try a new route?

Thanks for taking a look.

Loop based territory maps are a great way to make walking feel less like chores. Rolling Fog sounds evil in a useful way.

I will try you app.

turning outdoor activity into a territory game is a fun layer on something most people already track passively. curious about the local network effect. does a new player joining a city need other players around to claim territory from, or does the early experience hold up solo before the area gets dense? that early game design is usually what makes or breaks adoption.

is this tool can reuse my login session when open page? or it's like open in stateless

the framing of the web as a command line for an agent is powerful because it implies the agent thinks in verbs not pages. curious how auth and session state work in that model. if an agent needs to read something behind a login, does the user grant per site credentials, or is there a more general delegation primitive? that part feels like the unlock for real autonomy.

Designers needing leads is painfully real. I like that this points at the boring business part agencies avoid until pipeline gets quiet.

lead gen for designers and agencies is a tough category because the quality of the lead matters way more than the volume. one warm intro to the right kind of client beats 50 cold contacts. curious how ClientJam scores lead fit before it surfaces it. is it scraping signal from a designer's portfolio and matching it to a prospect's recent activity, or is the matching more rules based?

Curious whether you see this expanding beyond marketing sites to evaluate SaaS or native apps, scoring things like accessibility, information architecture, onboarding flows, and overall UX. The prospect company would obviously need to provide some level of access (a demo, free trial, etc.) for the agent to work with though. Maybe even analyzing customer reviews to surface common pain points if there is no product access? Asking because our agency is trying to shift away from brand design and toward more product design work, so that kind of analysis would be huge for us in lead gen on that side.

The cold email writing is a great feature. I'm a sales rep turned project manager but still carry a small level of sales outreach responsibilities at my new company, and a piece of my soul dies every time I have to write a cold outreach. If ClientJam can take that off my plate, that's a huge win. Have you thought about options for tailoring the tone to specific stakeholder roles? The messaging for a startup founder versus a CMO at an SMB versus a product director at enterprise level would probably vary pretty significantly. Could be a really compelling layer to add.

The basic CRM is a smart call too. We actually just cancelled our HubSpot account last week because we weren't using it and was a waste of money. Something simple to track outreach and manage follow-ups is all we really need, that's a great feature to include here.

If you could incorporate product analysis into ClientJam, I'd be very intrigued and would definitely give it a try. Best of luck on the launch, sending good vibes your way! Well done!

Agency lead gen is brutal. What channels are you finding work best for outreach right now?

Hey Product Hunt 👋 Kevin here, solo maker of ClientJam.

The hardest part of running a web design business was never the design — it was finding clients. My process was brutal and manual: scroll Google Maps looking for businesses with rough websites, open 30 tabs to hand-audit each one (is it on mobile? is it slow? no SSL?), then stare at a blank screen trying to write a cold email that didn't sound like every other cold email. Hours of work to send a handful of pitches.

I kept thinking: every step here is something a computer should do for me. I already know what a "ripe" lead looks like — a real local business losing customers to a broken website. I just needed something to find them, score them, and help me say the right thing.

So I built ClientJam to collapse that whole funnel into one pass: paste a URL (or pull a batch from any city) → technical audit → opportunity score → three outreach emails grounded in that site's actual problems.

How it evolved: I started thinking it was an "audit tool," but every test user kept asking the same thing — "okay, but what do I say to them?" That reframed the whole product. The audit isn't the point; it's the evidence. The real value is turning a score into a sentence you can actually send. That's when the outreach generator went from a nice-to-have to the core.

It's free to start (no card), and I built it solo on Next.js, Vercel, Clerk, Stripe, and Neon.

I'd genuinely love feedback from other designers, freelancers, and agency folks: when you're prospecting, what's the step that actually wastes your time? Happy to answer anything in the comments 🫐

Hi Product Hunt 👋

Avocado is an AI content operations platform and framework for Next.js websites.

🚀 Live Demo: https://avocado-editor.vercel.app/

🧩 Component catalogue used on the demo site: https://avocado-site.vercel.app/catalogue

Most AI-native content tools are bundled inside expensive DXPs, proprietary CMSs, or website builders.

Avocado takes a different approach. Instead of replacing your stack, it adds an AI content layer on top of your existing website.

With Avocado you can:

✅ Edit website content using natural language
✅ Manage pages, metadata, and structured content
✅ Work with real design-system components and content models
✅ Preview AI-generated changes instantly
✅ AI-powered content creation and updates
✅ Connect Contentful, Sanity, Strapi, or custom CMSs
✅ Bring your own OpenAI, Anthropic, or Gemini models
✅ Self-host and keep full control of your data

Think of it as an AI orchestration layer for content operations—not another CMS or website builder. Avocado works with structured content and real website components, enabling AI to understand and modify pages the same way your content and development teams do.

I built Avocado because I believe AI-native content management should be available to every team, not just enterprises with six-figure DXP budgets.

I'd love feedback from:

• Next.js developers
• Content teams
• Agencies
• Headless CMS users
• Composable architecture enthusiasts

What content workflows would you automate with AI?

Thanks for checking out Avocado Studio 🥑 !

Note: The GitHub repository isn't public yet. I'm using this launch to validate the product direction and gather feedback before open-sourcing Avocado Studio under the Apache 2.0 license. If you'd like early access or want to contribute, let me know in the comments.

content ops for Next.js sites is a sweet spot because the same team is usually shipping marketing pages and product copy from the same repo. curious how Avocado handles the boundary between AI generated drafts and the existing review workflow. does it open a PR like a teammate would, or does it write straight to a CMS layer that the team then approves?

Adding an AI content layer on top of an existing Next stack feels way more sane than replacing the whole CMS. BYO models is a nice trust point.

Congrats! Curious how does Avocado handle content model differences when connecting to multiple CMSs like Contentful and Sanity at the same time?