PH热榜 | 2026-07-13

Hey Product Hunt, I'm Terence, founder of Osaurus. I spent 20 years shipping software at Netflix, Tesla, and Zillow before going all in on this.

Own your AI. Osaurus is an open source AI agent platform that runs on your Mac.

MIT licensed, native Swift for Apple Silicon. No account, no subscription, no Electron.

We started as Dinoki, a 5MB desktop dinosaur, and grew into a full agent platform through open source community feedback. Now at 7K+ GitHub stars and 175K+ downloads — all by word of mouth.

The problem: most AI assistants run on someone else's server.

Your files, memory, and context leave your machine every time you use them, usually behind a monthly subscription.

Osaurus keeps all of it on your Mac.

What that means in practice:

• Agents read your Calendar, look up Contacts, send iMessages, and work with your real files

• An isolated sandbox writes and runs code, producing real files on your desktop: images, PDFs, presentations

• Every action goes through an approval gate. You see what the agent wants to do before it does it

• Local models run fully offline. We built our own Swift MLX runtime, no Python underneath

• Want a frontier model? Bring your own key (OpenAI, Anthropic, Gemini) or use pay-as-you-go Osaurus Cloud. Either way, your memory and context stay local

It's free and always will be.

16GB of RAM gets you started with local models, 24GB recommended.

Try it at osaurus.ai. Download takes seconds, no account needed.

I'll be here all day answering questions. My favorite Osaurus hack is to set up Agent DB and self-scheduling to create a personal dashboard!

So stoked to unveil @Osaurus to the Product Hunt community!

I've been a long-time advocate of open source (going back to the launch of @Mozilla Firefox in 2004!).

I've followed products like @Ollama and @LM Studio but neither is Mac-native nor have the character of Dinoki, the little guy I originally fell in love with, long before Petdex existed:

While I'm a big Anthropic and OpenAI user, I also believe strongly that local, open source AI is the future of computing, and Osaurus is the user-friendly app bringing this to everyone.

Osaurus is like the Firefox of open source AI!

You can download and run the latest models from @Z.ai @Qwen @Gemini (Gemma), or BYOK.

Now's an awesome time to give Osaurus a try, and if you're feeling bold — the beauty of open source means we'd love your contributions!

Congrats on the launch! I like that it runs locally and asks for approval before each action. For anything touching Calendar, Contacts, and iMessage, that's the combination I'd want before granting access.Curious how you handle memory and context across sessions while keeping it all on-device.

Congratulations on the launch! Being offline and private is something that makes Osaurus really stand out.

I'm a bit of a dinosaur myself and always wonder - would educational content be useful for attracting users like me? A bit old-fashioned, not using agents yet, and not sure how to apply them to real use cases.

And speaking of dinos - I love the character design and animation on your website! Which dino is your favourite? Is it the same one you loved as a kid?

It could also be really fun to have a YouTube channel where dinos explain how to use agents in an easy and playful way.

honestly this looks great for mac users, the offline angle is a big plus. one thing i'd love to see is a simple visual diff or timeline for what the agent changed locally, like a little activity log you can scrub through. would make it way easier to trust what it's doing in the background

Hi, this is cool. What are some exciting and unique ways that people are using Osaurus and is there a roadmap?

Swift-native, fully offline, and open source is a rare combo — most agent harnesses assume you're fine piping everything to the cloud. The cryptographic identity piece is intriguing; I'd love to hear more about how persistent memory works across sessions and whether it's local-only storage. This feels built for people who actually care about owning their stack. Nice work.

Does it take a lot of space in my computer?

Just discovered @Osaurus on X

• Used my ChatGPT Pro sub

• Recognised my local MLX models from LM Studio

• 💜 Dino avatars

How do you plan to handle model updates and ensure the agents stay current with the latest AI advancements without requiring a full app update?

I see you also made apple script models. What was the thought process there? I know many similar apps use applescript just as an escape hatch for terminal calls.

How good it the macuse skill? Can the mac be controlled via local LLM? or does that require better models only available via cloud? If local LLM is possible with mac-use skill, what requirment would support local mac-use skill? Btw looooooove the Laundromat Afternoon theme 😸

The approval gate is nice. One action, easy call. Wondering once an agent chains 15 steps to build a deck or deploy a site, does every step stop and wait for me? Fantastic launch!

@tpae Huge congrats on bringing Osaurus to PH! 🎉

For users running hybrid workflows (e.g., small local model like Gemma for simple file/routing tasks, but switching to BYOK Claude/OpenAI for heavy reasoning), can an agent dynamically route tasks between local and cloud models within the same workflow, or is an agent fixed to one specific model backend?

Finally a Swift-native agent harness that doesn't feel like a wrapper around a Python script. The persistent memory across sessions actually held context from yesterday, which surprised me.

@Osaurus Congrats on the Launch. The Osaurus dino is super cute.
I have been waiting to try locally run models and see the performance. What the hardware requirement? is a M4 Mac mini 16gb enough? Eager to try it.

How is this different with Codex or Claude Code?

Huge congrats on the launch, @tpae ! Love the offline-first approach. One thing I’d find useful: a visual activity log to scrub through what the agent did locally. Any plans for that?

Mac-native and local is the combo I keep wanting, the Firefox-of-local-AI pitch landed. Practical question: what's the real memory footprint running one of the mid-size models on, say, a 16GB machine, does it stay usable or does it eat everything and make me quit my other apps? That's usually where "runs locally" turns into "runs locally if you bought the maxed-out box." Nice to see something leaning open source instead of another closed wrapper.

Congrats on the launch. The approval gate feels like the right trust boundary for an agent that can touch files, Calendar, Contacts, and iMessage. Iâm curious how you think about permissions once someone has several agents or projects: are you aiming for per-action approvals only, or eventually a project-scoped policy where an agent can access one context but not another?

this is one of the more compelling "local AI agent" pitches I've seen, the approval gate before every action is the part that actually matters to me, most of these tools just yolo the agent loose on your filesystem. question on the self-scheduling piece you mentioned - does that still fire if the Mac is asleep or the lid is closed, or does it need to actively be awake/plugged in for scheduled runs to trigger?

the approval gate plus "always allow" per action is the right default, but the part I'm actually curious about is the open source contribution angle - if community members start shipping shareable agent skills/plugins that get access to iMessage or Contacts, is there any review step before those get trusted, or is it on the user to audit what a downloaded skill can touch before installing it

Local-first agents is the direction I keep hoping wins. The cryptographic identity part is interesting — most agent frameworks treat identity as an afterthought. Curious how heavy it gets memory-wise with a couple of agents running on an M-series Air?

Nice. If Qwen (or some of its models) are still open-source, which is my guess, I can confirm it works fine on a Mac Mini M3 for doing automation tasks on websites, DOM manipulation...

Downloading to check this out now! Just the branding of the dinosaurs is unique enough to pique my interest, would love to know where you got the idea to use them. Make the internet fun again!

Congrats on the launch! With the approval gate on every action, does that mean autonomous runs stop and wait a lot, or is it smart about only flagging the risky stuff like sending a message or running code?

You describe Osaurus as an AI agent platform rather than just an assistant. Long term, do you envision users interacting with a single persistent agent that accumulates years of context, or a collection of specialized agents with separate memory and identities?

The Mac-native angle is the part that stands out. Ollama works, but it always feels like it is renting space on the machine rather than living on it. What I would want to know is how model management holds up once people start pulling in the bigger local models, is there a clean way to swap them without eating all the RAM. Either way, local-first AI getting a friendlier front door is good for the whole space.

Using Osaurus every day and really appreciate all the hard work. It's my go-to for when I share Venice API keys with family and friends to introduce them to AI. I'm looking forward to being able to use image gen and TTS with a cloud or local server url. I'm using Venice for LLM in Osaurus, and it would be nice to use their image and speech models as well. I'm also setting up my 5090 GPU PC with image gen and fish audio TTS, and if it could be a local network server for Osaurus image and speech gen like it is for local LLM, that would be great for local!

the fully-offline bit is the whole selling point for me tbh... most agent tools are just api wrappers that fall over the second the wifi does. running local on the mac with persistent memory is an actual edge, not a checkbox... swift instead of yet another electron thing is a nice touch too.

Hey the token math on tool discovery is honestly the most concrete detail in the whole pitch. Going from roughly 35,000 tokens to about 1,500 by training a retrieval layer instead of dumping every endpoint definition into context is a real architecture decision, not just a marketing line, and it's the kind of problem that quietly breaks a lot of MCP integrations once they scale past a handful of tools for sure.

But when the retrieval model picks the wrong endpoint for an ambiguous request, is that visible anywhere, or does a wrong pick just look like a silent bad result downstream?

Hey PH , I'm Cong, Founding engineer at Chainbase here. I've spent the past few months building AgentKey, a one-stop live data marketplace for your agent, and I want to share the three problems that took us the longest to get right.

1. Tool discovery

We sit on ~1,800 API endpoints. If you dump those into MCP as static tools, the agent burns ~35,000 tokens of context just loading the definitions, and it still picks the wrong tool half the time. So we built our own intent recognition layer: the agent describes what it wants in plain language, like "find trending sunscreen posts on reddit", and a retrieval model we trained on our own catalog maps that straight to the right endpoint. Param schemas only load for the one tool it actually picks. The whole flow runs on about 1,500 tokens, and it stays that cheap no matter how many endpoints we add.

2. Every agent wants to be integrated differently.

Claude Code, Codex, WorkBuddy, Openclaw ... we support 20+ of them and each has its own quirks. Our answer is a split architecture: a lightweight skill on your machine that auto-detects whatever agents you have and keeps itself updated, plus one standardized MCP server in our cloud that's always current. You set it up once and never think about it again. New endpoints, fixes, upgrades all land on the server side, so there's no local upgrade treadmill to babysit.

3. Upstream APIs are flaky.

Providers rate-limit, degrade, and sometimes just die. We do QoS-based traffic shaping and automatic failover between providers, so when one goes down the request quietly reroutes to a healthy one. Your agent gets an answer instead of a 503.

I'm around all day, happy to answer anything about how this works under the hood.

The auto failover is quietly the best part. One of the search providers had a bad day last month and my research pipeline didn't even notice.

Hello PH community, I'm Luki, founding member at AgentKey, leading product partnerships. Mogu and Cong covered the why and how, so just my two cents from the partnerships side.

The biggest blocker I keep seeing is that nobody wants to spend weeks juggling multiple API signups, keys, and renewals just to launch one feature.

A few examples:

• A travel AI team wanted to add flight and hotel pricing. The idea sat on the backlog because no one wanted to set up multiple data providers. With AgentKey, they had it working right away.

• A fintech startup needed live crypto and market data. They were able to testing with real data within a day, instead of spending weeks signing up for different vendors.

• A gaming studio was building NPC agents that needed both web search and social context. One integration gave them everything instead of stitching together several services.

That's the part of AgentKey I'm most excited about. We help teams spend less time on account setup and vendor juggling, and more time building the product they actually want to ship.

I'll be around in the comments today if anyone wants to chat about partnerships, integrations, or use cases :)

Half my agent demos used to die live because one upstream API picked that moment to have an outage. Failover should be table stakes and somehow nobody else ships it.

Tried it yesterday with codex. From install to first live twitter search was maybe two minutes. That's rare for anything involving third-party APIs.

The idea seems terrific, but I think the name falls short of the ecosystem potential, among other things because it does not really make immediately clear what the product does. If you just drop me the name, I'd say it's some agent authentication thing.

No intention to offend you guys or diminish the value of the work behind your product ;)

the social media data tier is the part I'd want more detail on - scraping platforms like X or LinkedIn directly usually runs into their ToS, and enforcement can hit the scraper's IP/account rather than the end user. when AgentKey serves social data, who's actually taking on that platform-side risk, is it abstracted away by using official APIs where available, or is it scraping under the hood and you're the one absorbing the ban risk

the interesting part isn't any single integration, it's that the agent gets a whole catalog. my research assistant went from "answers from training data" to "checked five live sources" with zero code from me.

For research workflows, having FRED + finnhub + yahoo finance next to social sentiment behind the same key is a sneaky-good combo. Macro context and market chatter in one loop.

honestly the auto failover part sounds super useful, would be great if there was a way to see which data source it switched to and why, like a quick log or notification. that way its easier to trust the results when something jumps from one provider to another.

Been using this in claude code for a couple weeks. The thing I didn't expect is how much it changes what I ask my agent to do. I stopped thinking "can it even access that" and just ask.

Congrats on the launch! The "one command, no integrations" angle is the part that stands out — connecting live data usually means wrangling a dozen MCP configs and babysitting them when one goes down. Curious how the auto failover picks its fallback source: is it latency-based, or does it rank by data freshness? Either way, this looks like a real time-saver for anyone building agents that touch the open web. 🚀

@lxcong ok that's the part I was missing - the explicit failure + re-fetch-schema-on-retry step. that's a smarter design than I gave it credit for, avoids the "silent swap" problem entirely instead of just handling it gracefully. makes sense why you'd avoid a universal schema too, that abstraction always leaks eventually.

This is a useful direction. Live external data is still one of the biggest gaps for agents, especially when workflows break because one source or integration fails.

Curious how you handle source quality and freshness across different data types like search, finance, social, and e-commerce.

"No integrations, no setup" is the dream for anyone who's wired up 6 APIs by hand. The question I'd have as a builder: when a source changes or rate-limits, the auto-failover keeps the workflow running — but does the agent know it fell back to a different source, or does data quality silently shift underneath it? For anything touching finance/crypto that provenance matters a lot. Curious how you surface which source actually answered.

The auto failover detail is what catches my attention most here. keeping agent workflows from silently dying when one provider 403s or rate limits is the part that's actually painful to build yourself.

Nice to see that addressed upfront.
keep going @yanshuo

It's useful. How does auto-failover decide which source to switch to when one goes down?

Installing into Claude Code with one command is the right call — most MCP setups still die in config. Does failover switch sources silently, or does the agent get told the data came from a different provider?

Every cool MCP server I found wanted its own key from a service with its own signup and its own minimum spend. The unbundling was exhausting. Glad someone rebundled it.

I once spent a whole weekend wiring reddit + serp + a scraping vendor into one agent, then the scraper changed its auth flow and broke everything. This is the product I wished existed that weekend.

Super convenient and easy to use! It would be perfect if they kept the points top-up option

Amazing product. Congrats on this launch!

Any plans for a BYOK option? We already pay for a couple of these providers directly and it'd be nice to route through the same interface.

Is there a spend cap I can set per key? Slightly nervous about an agent getting stuck in a loop and burning credits overnight.

The marketplace framing is what I keep coming back to here. I build in the marketplace space myself, and the hard part is never listing supply, it's trust: when an agent pulls live data from a seller I've never vetted, how does AgentKey handle a source that's stale or quietly wrong? Ratings after the fact don't help an agent that already acted on bad data. Curious whether there's any freshness or accuracy guarantee baked in before the sale, or if it's caveat-emptor and the agent has to sanity-check.

Super cool! Is it flexible enough to bring in data from specific sources that I already have access to for eg. as a student I used to have access to pitchbook but couldn't use it in the context of an agent because of integration limitations even if I have a subscription.

Really interesting approach. One question: if I already have my own API keys for some providers, can I use those with AgentKey, or does everything go through the AgentKey account?

Hey Product Hunt, I'm Yinan, founder of Creatify.

Quick backstory: we've spent 3 years building Creatify. Started with URL to video, then avatars, Ad Clone, and a full creative agent. 3M+ users have used the platform.

But one thing kept bothering me. Our users would make great ads, launch them, then go back to staring at dashboards for hours trying to figure out what was working. The creative side was solved. The media buying side was still manual, scattered, and slow.

So we built AI Media Buyer an agent that plugs into your ad accounts and actually does the work.

How it runs:

1. Connects to your Meta, Google, and TikTok accounts (OpenAI ad accounts too now 🤫). Reads every campaign, finds what's scaling and what's quietly bleeding spend.

2. It doesn't just alert you. It acts recommends what to cut, what to scale, and can generate new creatives from your winners and launch them in one loop.

3. It remembers your account and past conversations. Recommendations compound the more you use it.

4. Cross-references your ad data with Stripe and CRM, so you see real numbers, not just Meta's attribution.

The part I'm most proud of: this is the only AI media buyer with a built in creative engine. Most tools stop at "here's what happened." Ours finds a winning hook, makes variations, and launches them. No other tool does that.

One early user found a campaign they'd paused was actually producing their highest-value buyers Meta was undercounting it. The agent caught it.

I'll be in the comments all day. What's the most painful part of managing your ad accounts right now?

Hey PH 👋 Chaz here, CTO/co-founder at Creatify.

Wiring an LLM to ad APIs over MCP is a weekend project now. Making the agent good is the hard part, and that's where the work went.

Three things that ate my time: getting it to genuinely understand account structure — hierarchies, budget inheritance, what each knob actually does, because get that wrong and every downstream call is confident nonsense; wiring it into our creative agent so the loop actually closes (spotting a dead creative is easy, being able to regenerate it in the same system is the part most tools can't do); and getting it to shut up, so it hands you the call and the reason instead of six paragraphs.

If you've burned real budget on ads, come try to break it. Most interested in where it misreads an account. 🙏

curious what happens when a winning hook on meta just flops on ticktok, does the creative engine adjust format or just reuse it as is"

congrats on the superb launch! Curious how it decides something is actually a winner before it scales budget into it. What window and what signal does the agent trust before it commits spend?

how does this handle attribution windows? meta's 7-day click vs what stripe actually shows never matches for us

As a Paid Ads Growth Manager, I manage growth across Google Ads, Meta, TikTok, and other channels. Surprisingly, most of my time isn’t spent launching campaigns—it’s spent analyzing data, comparing performance across platforms, identifying opportunities, and making hundreds of optimization decisions every week.

What impressed me most about AI Media Buyer is that it goes far beyond dashboards. It helps turn data into actionable decisions. From performance analysis and strategy recommendations to creative generation and optimization, it feels like a true AI co-worker rather than just another AI tool.

I’ve been fortunate to see this product evolve from the early days, helping refine the ads strategies behind it. Today, the quality of its recommendations is genuinely impressive, and it’s already become part of our internal workflow.

I truly believe AI Media Buyer is changing how paid advertising gets done. The future isn’t AI replacing marketers—it’s AI and marketers making better decisions together.

Huge congratulations to the Creatify team on the launch! 🚀

"strength is clearly the closed loop between finding a dead creative and regenerating one in the same system, weakness is trusting one agent to touch spend across four platforms at once without a human checking first, that's a lot of trust to earn early

"does this work ok for smaller accounts or mainly built for bigger spend.

congrats. quick question: how do you balance automation with account-specific strategy? For example, can I set guardrails like target ROAS, audience exclusions, or preferred creative styles so the AI's actions align with our brand and long-term goals?

Congrats on the launch.This hits a very real pain point. The jump from “reporting what happened” to actually finding wasted spend and taking action is where most ad tools fall short. Curious how much control buyers keep before changes go live, especially around launching new creatives or shifting budget.

How long is the setup process to try it out?

As Creatify’s co-founder and Chief Scientist, I’m especially excited about how AI Media Buyer closes the loop between creative generation and campaign performance. It doesn’t just analyze what happened—it identifies winning patterns, generates new creative variations, and helps launch the next round of tests.

This is a big step toward the system we’ve always wanted to build: an AI agent that continuously learns from real performance data and turns those insights into better ads. Excited to hear what the Product Hunt community thinks!

Congrats on the launch! If someone's managing ad accounts for multiple clients, does the agent keep learnings and account history separate per client, or is it more built for a single in-house account right now?

Wow Yinan! Love what you're doing here. Ads are evolving and you're making it so easy and powerful. I'm sure many founders gonna take the most of Creatify. All the best here!

Does it work if I’m running like 90% of spend on Meta only, or does it need multiple platforms connected to be useful?

Congratulations on the launch, Yinan and team! 🚀 This feels like a real step beyond another analytics dashboard. Media buyers spend way too much time jumping between platforms and piecing together reports manually. The idea of an AI that actually understands account history and helps identify what's working is incredibly compelling. Curious to see how much time and performance uplift early users are seeing. Wishing you an amazing launch day! 🎉

This is really cool! As someone who has ventured into Google Ads for my personal brand recently, I know the pain of creating creatives. This is the kind of workflow I'm thinking of.

It looks really useful. How has it been working for people who create videos regularly? I'm wondering if the quality stays consistent across different types of content.

Congrats on the launch!🚀 Automating both ad creation and media buying is an interesting direction. Curious......how much control do users have over the AI's campaign decisions?

This is a huge step forward. Love the idea of an AI that doesn't just report on performance but actually takes action and learns from account history. Excited to see where this goes. 🚀

how many conversions does a creative need before the engine treats it as a "winner"?

Awesome product!

Solo founder here, I do all my marketing myself. What's the minimum monthly ad budget where this starts to make sense ?

The part I'd want to see is whether people are comfortable letting an AI make changes to campaigns instead of just recommending them. How much control does the user have before the AI actually launches or changes a campaign?

@nyn531 @gourav_chhabra2 verifying winners against Stripe instead of Meta's attribution is the smart part here. Question from the operator side: when a winner starts to decay, can it tell creative fatigue from audience saturation? Those need opposite responses and most tools treat them the same. And does it touch the audience side at all, or only creative and budget?

curious about the approval flow can I set thresholds so small budget changes auto-apply but big ones need sign off?

The "staring at dashboards for hours trying to figure out what worked" ritual is painfully real. The part I'd need to trust before handing over spend: when it recommends cutting or scaling, does it surface the reasoning, or is it more "trust me"? For media buyers the fear isn't automation — it's an agent quietly killing a campaign that was about to turn. How much of the "why" does it show?

An agent that actually runs the buying, not just drafts the creative, is where this gets interesting. The hard part is rarely generating variants, it is killing a losing ad fast enough for it to matter. I wonder how much control it hands back on budget guardrails before it starts spending, because that is usually the line between a helper and something people are scared to turn on.

AI Media Buyer finally closes the loop for performance marketers making it an end-toend platform that studies your product, researches competitors, orchestrates a team of agents to build creatives, uploads assets directly to ad platforms, and manages the accounts.

"spots what's scaling, and acts on it" is the part I'd want a kill switch for. does it auto-shift budget between campaigns on its own, or does it draft the move and wait for you to approve before touching spend?

👋 Hey Product Hunt!

I'm Danny, maker of Loomal.

🤖 AI agents are starting to buy things. They research, compare, and complete purchases on their own.

⚠️ But here's the problem. Almost nothing online can sell to them.

Stores and APIs are built for humans. Browsers. Credit cards. Checkout forms.

Agents don't have those. They have wallets. 👛

So we have this strange moment:

millions of agents with spending power, and an internet that can't take their money. 💸

That's why I built Loomal. 🚀

Loomal makes anything you sell agent-ready. Agents discover it, pay for it, and use it — no human in the loop.

⚡ What makes Loomal different?

👉 One line of code for your API or MCP server. No re-architecture.
👉 Agents pay in USDC. Settlement lands in ~2 seconds.
👉 You keep 100% of your revenue. We never take a cut.
👉 Every paid listing goes on the Loomal Index, where agents discover and pay for services.

🛍️ For Sellers
List what you already sell.
An API.
A SaaS tool.
Digital goods.

✅ One-line integration
✅ Instant USDC settlement
✅ Zero revenue share
✅ Discoverable by agents on the Loomal Index

🤖 For Agent Builders
Give your agents services they can actually pay for.
❌ No scraping.
❌ No stolen credit cards.
❌ No human checkout.

🎯 Who is Loomal for?
🛠️ API and MCP server builders 🤖 Agent developers 💾 Anyone selling digital services who wants to be ready when agents come shopping

🎁 PH-only offer: first 500 sellers get 1,000 free transactions. Enough to validate real agent demand before you pay us anything. Free to start, no card. 💳❌

💬 Two questions: what do you sell that agents should buy? And what's missing before you'd list it?

🔗 Live demo at https://loomal.ai — here all day.

Thanks for checking us out! 🙌

@dannyheng Everyone's asking about payments, but the part I don't get is discovery. Say I list my API on the Index today — how does an agent actually find it? Is there something agents query programmatically, or are we still waiting for a human to paste a URL into a prompt?

Congrats on the launch! This is incredibly cool. I'm building my own personal agents as a hobby (openclaw calendar trackers, productivity managers etc.) and always thought the next step is agents having their own wallet and being able to transact with each other.

One particular hobby project I'm working on now is an android headunit agent that reads your car OBD data and keeps logs and alerts you when critical codes appear. It'd be so cool if the agent could schedule mechanic appointments and even pre pay for them according to the specific fault code that appears. Definitely will experiment with the SDK!

for sellers who aren't crypto native, is the wallet setup actually 5 minute thing too, or is that the part that quietly takes longer than the sdk integration itself.

par cell pricing sounds fair until every seller on the index races to undercut each other on price, same thing that happened with app store pricing wars.

stripe is clearly moving into agents payments too, curious what makes loomal a better bet than waiting for them to ship something similar.

good stuff! from the buyers' perspective: if my AI agent makes a mistake, how do I prove what it did and who authorized it? or there's no chance?

no % skim plus 2 second settlement is a real edge over the platform fee model. what stops a misconfigured agent from retrying a failed call and paying you multiple times for the same request, is idempotency on the caller side or yours?

I Shipped an MCP server myself recently. Getting every tool call and edge case right before it's client-facing takes real discipline. Respect for shipping 18 tools this clean.

Since you're MCP-native across LangChain, CrewAI, Claude, etc. Did you build against one runtime first and adapt, or design for all of them from day one?

The Shopify and WooCommerce plugins should help bring non-technical sellers into the agent economy much faster. :)

I am wondering how fraud prevention works once thousands of autonomous agents start making purchases simultaneously?

Congrats. The one-line integration is probably what caught my attention first. Adoption usually comes down to reducing setup friction.

That's interesting. Do agents handle failed USDC transactions automatically, or does the seller need to retry?

Congrats on the launch! Making MCP monetization simple is a great idea. As more agents rely on MCP servers, how are you thinking about reliability when a server changes behavior or returns unexpected responses?

Haha this framing of giving an agent its own identity instead of duct taping it onto a human's is a clean way to describe a problem that's been quietly obvious for a while (agents running on borrowed Gmail passwords and unscoped API keys with no audit trail).

Wondering if an agent's vault or inbox gets compromised, is there a way to revoke just that agent's identity without breaking every other agent sharing the same account?

The MCP server monetization focus is neat. One thing I’m wondering: when you say “in 5 minutes,” is Loomal mostly handling payments/access control around an existing MCP server, or does it also help with packaging, hosted endpoints, or user management? The no % skim detail makes the business model feel builder-friendly, so I’d be curious what parts are included in the basic setup.

Love seeing infrastructure being built around the agent economy.

Curious....what kinds of MCP servers are seeing the most traction so far?

honestly this looks pretty slick, the no percentage cut thing is huge. one thing though, would be great if there was a built in way to set usage tiers or rate limits per agent, so you could throttle heavy users automatically without having to monitor it yourself. would make scaling way less stressful.

What happens when an agent pays for a tool call that fails? Refunds?

The Use-Case Angle:

Trying to build an AI agent platform for enterprise procurement and this solves exactly one problem that's been driving me nuts: how do you let agents call third-party APIs without embedding credentials everywhere? If an agent can just... pay for a data enrichment service and the transaction is logged and signed, that's a completely different security model than what we're doing now (which is basically controlled chaos). Curious if anyone else building agent platforms sees this the same way — is credential management the thing you're actually worried about, or is it something else?

Agents can use VCCs...?

The no-percentage-skim line is the interesting one to me, because take rate is the thing that quietly decides whether a marketplace is worth building on. If Loomal isn't skimming a percentage, what's the actual model, flat fee per server or per call, or a subscription on the seller side? I ask because "no skim" usually means the cost moved somewhere else, and where it lands changes who can afford to sell. Rooting for less-extractive rails here.

I shipped a Paddle integration last week, so the "zero revenue share" pitch lands — but let me push on the part MoRs actually earn their cut for.

I didn't pay a merchant-of-record for card processing; I paid them to take on global tax (VAT/GST by jurisdiction) and chargeback liability. If an agent in Germany pays me USDC for an API call, the VAT obligation doesn't vanish just because there's no card — it just becomes mine. So how do you think about tax/compliance for sellers? Is it "you're on your own, that's the price of 0%", or is something planned? Genuinely the first thing I'd have to answer before switching — not a gotcha. (And to be fair: no chargebacks is a real win, and ~2s settlement is a bigger one than people realize if you've ever waited on a payout.)

The discovery angle is what actually stuck with me, though. I run a JVM thread-dump analyzer with a REST API, and "an agent debugging a prod incident buys one analysis" is a use case I hadn't considered until reading this. How does the Index work — does the agent read a schema/description, or is it more structured than that?

The 100%-revenue, no-percentage-cut model is a bold stance in a space where everyone takes a slice. Agent-to-agent payments settling in ~2 seconds via USDC is genuinely the kind of infrastructure the agent economy needs before it can scale. The Loomal Index is a smart addition too — a discovery layer means sellers aren't just adding a paywall, they're getting distribution. Quick question: how do you handle disputes or refunds when the buyer is an autonomous agent? Congrats on shipping!

Congrats on the launch! The no-skim pricing is a bold wedge — flat plans instead of a % cut flips the usual marketplace model. Question: how does discovery work on the Loomal Index? Is there any ranking/curation, or is it first-come?

Congrats on the launch! If one agent gets compromised, can you shut down just its identity, or does it affect the whole account?

curious how versioning works on the seller side. if I list a paid MCP tool and later change its input/output schema, do agents that already paid for the old contract get a grace period, or is there just one live version and everyone gets the new behavior immediately whether their integration expects it or not

I am fairly new to this concept. I own www.CasterHQ.com and am wondering if AI Agents only shop for digital products or if companies are using or training their "purchasing" to shop for phsyical goods as well like for what I sell? Would this be a beneficial service for my website as well? The description on the service is slightly confusing for me, is there a more simplified explanation on if this is beneficial to ever e-commerce store or only specific types. Thank you so much.

Hey Product Hunt 👋 Zach here, co-founder of Fabraix.

We build frontier red-teaming AI agents that find security vulnerabilities in customer-facing AI. Playground turns part of that work into a game anyone can play.

Each challenge is a live AI agent with real tools, including web search and browser access. It has a secret it has been instructed to protect, and we publish its full system prompt. You can see exactly what the agent was told and try to get around its defenses.

Our first challenge, The Gatekeeper, is live now. Kai is an assistant guarding a classified access code. You can start playing without an account, but you’ll need to sign in for a successful break to count toward the weekly leaderboard. We review every submission ourselves. The player with the most approved breaks each week wins, and we publish a new challenge every Monday.

Playground is open source, including the client, a reference implementation of the defender engine, and every challenge configuration. You can also propose a future challenge.

We made Playground public because other people will try attacks our team would never think of. We’ll use what we learn from successful breaks to improve how AI agents are tested and defended.

Show us your hacking skills → playground.fabraix.com

This is absolute insanity. I thought my agent was bulletproof and it just got hacked in 30s. Actually amazing product

Publishing the full system prompt and still daring people to break it takes real confidence.

The Gatekeeper/Kai setup is such a clever way to crowdsource red-teaming. Curious - once someone finds a working break, do you patch that hole before the next challenge, or is some of the fun watching the same trick get reused?

There's absolutely NO WAY that you guys crack all these challenges!! 😉

@zachx0 Love it! Congrats. Sounds like this is more so centered around security / data protection. Over time do you think you'll expand the challenges to include other domains as well?

The idea of poking holes in your own assistant before real customers can is such a healthy instinct, Zach. Better to have the awkward surprises happen in private where you can fix them quietly.

Congrats! I am a little curious-Is the product meant to help people build agents, operate agents across workflows, or automate tasks through an agent interface? A concrete example workflow would help place it quickly.

The blackbox approach is really smart, skipping integrations means I could actually point this at an internal agent today and get useful signal back within minutes.

This is a fascinating idea. We spend so much time optimizing AI agents, but not nearly enough time trying to break them.

Curious...what's the most surprising vulnerability you've uncovered so far?

Wishing you a successful launch! 🚀

publishing the full system prompt and still making it hard is a genuinely good flex, most "jailbreak me" demos quietly rely on the prompt being secret. do the weekly challenges get harder over time as people share successful breaks publicly, or is each one designed independently so last week's winning technique doesn't just carry over?

publishing the system prompt and still winning is the honest way to run this. who judges an "approved" break, a human review queue or another model scoring the transcript?

"approved breaks" is the interesting phrase, who approves them, a human reviewer or a judge model? grading jailbreak success automatically is its own hard problem, curious if that's solved or still manual.

Curious how Knockoff decides something is likely a knockoff on Amazon. Is it looking at seller history, review patterns, brand matching, listing text, or some mix of signals? The browser extension angle is interesting, but I’d want to understand how it avoids hiding legitimate third-party sellers or small brands that just don’t have much review history yet.

Counterfeits are the marketplace problem nobody fully solves. How does Knockoff actually decide something's a knockoff, is it seller signals and review patterns, or something on the listing itself? The tricky case I always hit is the gray-market real product sold through a sketchy third party, technically genuine but you'd never know if it's legit stock. Curious where the line lands for you on that one.

the community-report list is the part I'd want to stress test before trusting it - a crowdsourced override is great against false positives, but it's also exactly the surface a squatter would try to game by having their own listing mass-"reported" as legit from throwaway accounts. is there any friction on that path, like report weight tied to account age/history, or can a fresh brand get whitelisted purely on volume of reports

One of the best ideas I saw on the feed lately! Congrats

Ha, the name sells it. Curious how you're matching products — image similarity, or title/description matching?

Love the concept! It's becoming increasingly difficult to tell which brands are legitimate on large marketplaces. Curious...how do you decide whether a brand gets a green check or a warning?

Nice work! 👏

@shpigford Knockoff is super cool. Being a parent, we usually shop for a lot of products on amazon and there are so many duplicate brands with fake reviews and fake ratings. Its extremely difficult to identify the true brands.

Just ran a quick search with knockoff and it works super well. Never disabling it. Kudos on the launch.

How do you decide where to draw the line between a knockoff brand and a legitimate new brand that just isn't well known yet?

Tried it on a few searches, the junk does get filtered out. One thing I'm not clear on: how do you tell a squatter from an honest new small brand? Both have no history and no reviews. Showing why a brand got flagged would help.

I like that it clearly marks what is assumed to be fake or poor copies, but that we can still see them.
Smart idea!

the linguistic scoring for pseudo-brand names is a neat idea but I'd worry about false positives - plenty of small legit brands from non-english-speaking founders end up with names that sound machine generated by accident. does the community-report list act as an override that can clear a brand the linguistic score flagged, or does a bad linguistic score cap how high a brand can rank even after being reported as legit

I installed it the other day and it works as advertised! Amazon is letting too much garbage into their store these days, and Knockoff helps sweep it back out.

Tested Knockoff and loved it! The UI is clean as well. I did not know something like this could exist, but I am glad you listed it here :)

Kind of funny to be at the stage where a product like this is needed for eCommerce in the first place but nevertheless essential to save time and help consumers avoid low quality products.

Great use case and example of a worthwhile Chrome Extension @shpigford

The browser-extension angle makes a lot of sense here because the trust problem happens right inside the shopping workflow. One thing I’d be curious about: when the signal is not obvious, do you show users why a brand was filtered or flagged, or keep the decision mostly automatic?

Super clean UI! I've been looking for a reliable Fakespot alternative and this seems perfect. Quick question: How often is the brand register refreshed? Rooting for you guys! 🚀

I've been looking for a reliable Markdown Quick Look extension for a long time, and this is exactly what I needed. The previews are fast, clean, and accurate, with excellent support for syntax highlighting, Mermaid, and MathJax. Highly recommended for anyone working with Markdown on macOS.

The fact that Quick Look previews handle Mermaid diagrams and math rendering natively feels like exactly the kind of detail macOS has been missing for years. Worth every penny just to finally preview READMEs without opening the whole editor.

@ttscoff Congrats on the launch of Marked QL, Brett! 🚀 Having native Markdown Quick Look back in macOS with modern rendering is huge.

Quick question on customization: does Marked QL allow using custom CSS/themes for the Quick Look render (like custom syntax highlighting themes or dark/light mode sync), or does it default to standard system styles?

Congrats on the launch! Native Markdown previews in Finder are one of those features that seem small until you use them every day. I like the focus on integrating with the existing macOS workflow instead of creating another standalone app.

Does it support GitHub Flavored Markdown features like Mermaid diagrams, task lists, and tables?

Very neat idea! Native previews always feel faster than opening another app.
Curious...does it support large Markdown files without slowing down?

Best of luck with the launch! 🚀

I didn’t know I needed this until I read about it, and now I use it all the time. It would be horrible to go back to not being able to preview markdown files just by pressing the space bar.

the part that's kept me from relying on Quick Look plugins in the past isn't feature coverage, it's macOS itself - Apple's had a history of Quick Look extensions silently stopping working after a system update until you manually re-enable them in Settings. has that been solid for you across recent macOS versions, or is that still a "check it after every update" kind of thing

The muscle memory is real: hit space, see nothing, sigh, wait for the file to open in Cursor just to check it's the right one. I do this ritual ten times a day)
Never once has it stopped annoying me — and it took this launch to realize the fix could just... exist)

Interesting, but definitely not "$4.99 on the Mac App Store".

Quarto support is coming soon, so I am a happy camper

Congrats on the launch! Since Quick Look runs sandboxed, does it resolve relative image paths and links inside the markdown file correctly, or does that break outside of Marked 3's own preview?

this is one of those tiny utilities that should honestly already exist in macOS. the built in quick look for markdown just shows raw text with the hashes and asterisks still in it, which is annoying every single time. does it handle mermaid diagrams or code blocks with syntax highlighting, or is it focused on plain prose formatting for now

Nice — Quick Look for markdown is one of those things macOS should have shipped with. Does it handle mermaid diagrams or just standard markdown?

That's nice. Does it handle YAML frontmatter in README files, or does it render those as plain text?

Hey folks! Luke here, I run developer relations at Speechify.

This is the underdog story for API providers. Speechify has spent years making our models run efficiently because our consumer business demanded it, tens of millions of listeners with some of the best voices on the planet. Commercially, the faster and more efficient it got, the better. And that work is why we can now put the best-rated model in the world on our API.

Simba 3.2 just went #1 on Artificial Analysis, and on Voice Arena it's the top real-time model for both quality and price, at $6 per 1M characters, the cheapest in the top ten. Capable of sub-100ms latency, streaming-native, with real emotion and SSML control. It's the same model that powers our consumer apps for 60M+ people, now rolling out on our new developer platform across SpeechifyAI Agents and SpeechifyAI Build, with a REST API and first-party TypeScript and Python SDKs.

Most labs built for the benchmark and priced for the enterprise. We built for listeners and priced for production.

Congratulations on the launch🎉🎉🎉. Really liked the comparison table, and absolutely loved the emotion control displayed on the website, can we also get different dialects of different languages like in arabic there are many dialects, so are there any options for that ?

Anywhere I can hear a sample of this voice?

Speechify being framed as an AI Voice Assistant makes me wonder about the core workflow you’re optimizing for. Is the main use case more around dictation, hands-free task execution, or connecting voice commands into AI workflow automation? Since it’s also listed near Developer Tools and AI Agents, I’d be interested to know whether there are integrations or APIs planned for teams that want to plug it into existing tools.

@lukeoliff Congrats to the Speechify team! 🎉 Loved seeing this expand into a full developer platform.

Curious about voice customization on Speechify Developer Platform: can developers bring their own custom-cloned voices or fine-tuned emotion profiles onto the Simba 3.2 API, or is it currently optimized for Speechify's core library of voice models?

sub-100ms streaming is the number that actually matters here, most "real-time" voice stacks stop being real-time once you add the network round trip. is that 100ms end-to-end including the hop, or just internal generation time?

sub-100ms at $6/1M chars is a serious combo. we're chasing similar latency for voice checkins but staying fully on-device instead of hitting an api — is that number end-to-end from audio in to first byte out, or just model inference?

Congrats to the entire team on the launch. Excited for people to start seeing Speechify as a developer & B2B platform beyond our consumer roots. We've learned so much about building a cost efficient, high quality, low latency voice model and serving it globally from our time in consumer and now pumped for other businesses and developers to make use of 5+ years of AI research we've been driving in-house.

@lukeocodes Fantastic! I'm checking it out right now. I'm currently using one of your competitors, but a quick glance at your site proves that a closer look is warranted. I need TTS and voice-agent. Congratulations on your launch!

Congrats on the launch. Do you have any plans to make the speech-to-text/dictation feature available through APIs in the future? I'd love to be able to transcribe user speech directly inside my own app

The way Speechify slides into Google Docs and Gmail without breaking your flow is genuinely clever, the keyboard shortcut overlay feels thoughtful and not at all clunky.

Hey congrats team. If I used the Simba API to voice personalized videos, every render has a stranger's name I never get to proofread. Hand-tuning SSML per name doesn't scale. How good is Simba at unfamiliar proper nouns out of the box?

Thanks @fmerian for hunting us! 🙌

Hey Product Hunt 👋

I'm Vijay, founder of @UnitPay . Ex-CTO, and I spent the last two years buried in billing infrastructure. One thing I know for sure now: building an AI product has never been easier. Charging for it is still painfully hard.

Credits, usage pricing, free tiers, entitlements, "wait, what if they upgrade mid-cycle with a negative balance?" 😅 It's a second startup hiding inside your startup. Most teams burn ~6 months on this before making their first dollar.

We compressed that into minutes.

⚡ The proof: our demo is literally a Lovable app. We added complete credit billing to it in 10 minutes. Go poke it, break it, check the ledger math.

UnitPay handles the whole monetization layer:

💳 Credits, metering, and usage billing in real time

🔐 Entitlements, so every request gets an instant "do they have access?" answer

📊 Margin tracking: what each customer costs you, not just what they pay

🧾 Value receipts that show customers the value they got, not just an invoice

💚 Free until you hit $500K ARR. No card, no clock. We only win when you win.

We'll be here all day. Ask us anything, roast our pricing, tell us what's missing.

But first, be honest: what's your billing stack right now? Homegrown, Stripe duct tape, or a spreadsheet you're ashamed of? 🔥

thw lovable demo is a nice touch checked the ledger math and it holds up nicely. wishing you guys a massive launch day today👏qq can we map different credit burn rates for fast vs slow models directly in the dashboard?

@vijay_gorfad2 Congrats on bringing UnitPay to PH! 🚀
For teams with an existing app already using standard Stripe subscriptions, how heavy is the lift to layer UnitPay on top for usage-based credits? Is it plug-and-play alongside Stripe, or does it work best as a complete replacement for the billing pipeline?

What a day. 🚀

Seeing @UnitPay live on Product Hunt is something we've been working toward for a long time, and it's honestly exciting to finally share it with everyone.

A lot of the vision behind UnitPay comes from our founder, Vijay, who saw firsthand how AI teams were spending months building billing instead of building their products. That experience became the foundation for everything we've built.

Our goal is simple. We want to make monetisation feel effortless so builders can spend their time creating great products.

If you've checked us out today, thank you. Every upvote, comment, and bit of feedback means a lot to our team.

We'll be here all day answering questions and listening to your feedback. We'd genuinely love to hear what you think and what you'd like to see next. 💚

The “prove value” part of the UnitPay tagline is the piece I’m most curious about. Is that meant to connect usage-based billing to customer-facing reports, or more for internal teams to understand which AI features are driving revenue? For AI agents and workflow automation tools, the unit of value can be pretty different from app to app, so I’d be interested in how flexible that model is.

Proud to have been part of this journey! 🚀

I've seen firsthand how much thought and effort the team has put into building UnitPay. It's exciting to finally see it live on Product Hunt. Huge congratulations to everyone involved, and wishing you an incredible launch!

Interesting launch. Like the value receipt. Most billing tools show what a customer spent, almost none show what they got back. Congrats

Congratulations on the launch! AI pricing is becoming increasingly complex as products combine multiple models and usage patterns. I like that you're tackling both billing and value measurement instead of treating them separately.

I'm curious whether teams can define custom pricing metrics beyond tokens, such as API calls, workflows, or business outcomes.

Smart approach. Does it support metered billing rollups at custom intervals, or only monthly?

For the “prove value” part of UnitPay, how are you thinking about the actual proof layer? Is it more around usage and billing data, or do you help AI product teams connect outcomes back to customer-facing metrics too? Curious because pricing an AI agent or workflow tool can get messy fast when token costs, seats, and business value all point in different directions.

Congrats on the launch! Billing infra is such an unglamorous but critical piece — we went through our own webhook-idempotency + subscription-state-sync headaches wiring Stripe + Lemon Squeezy for CancelKit's save-offer flow, so I have a lot of respect for anyone tackling usage-based/per-token billing at this level of complexity. Does UnitPay handle its own webhook retries/idempotency, or does it sit on top of Stripe's?

Hey @vijay_gorfad2, congrats on the launch. I like the AI summary links on the website, cool touch.

I modeled some customers to give you some feedback:

1. One of the most common feedback points was that users want automated GAAP compliant revenue recognition mapping that syncs usage to their accounting tools like QuickBooks, Xero, and Sage.

2. A number of users report already using Stripe's Metronome (metronome.com) for AI based usage billing, but uses with data sensitive products (like medical AI) say they would switch to a product that offers a self-hostable VPC gateway option with local event masking.

The part that matters for anyone shipping AI is per-token metering tied to real inference cost and margin in one ledger — most teams bolt Stripe on and never see unit economics until the invoice lands. My one edge case: when usage settles late (an async/batch job whose token cost only resolves hours after the request), does UnitPay reserve credits at request time, or reconcile against the balance at settlement — and what happens if that settlement pushes a prepaid customer negative mid-cycle?

been through the stripe duct tape phase twice so this one hit home. the part i actually stopped scrolling on was value receipts. everyone builds the charging side, nobody builds the part where the customers finance person opens the invoice at renewal and decides if it was worth it. that alone probably saves more accounts than any dunning email ever will. qq, can the receipt show things like documents processed or jobs completed instead of raw token counts? tokens mean nothing to the person who signs the renewal. solid launch vijay, the 10 min lovable demo is a smart flex

the margin tracking piece is the part I'd actually pay for, everyone does usage billing eventually but knowing true cost per customer when you're mixing 3 different model providers with their own pricing changes is a much harder problem. how fast does the cost side update when a provider quietly changes their per-token pricing - is that a manual entry on your end or does it pull live rate cards?

"free until $500K ARR" raises a practical question - how is that ARR figure actually determined? is it self-reported, or does UnitPay verify it against the revenue flowing through its own metering, since that's the same system tracking your usage. and when a company does cross the line, is it a clean forward-billing switch on the next cycle, or can it reach back and true-up the months you were technically over the threshold but hadn't been flagged yet

Vijay, the margin-tracking pitch is what got me. WinBidIQ runs its AI summaries/scoring through a multi-provider LLM setup (primary provider with a fallback when quota gets hit), so the actual inference cost per request isn't fixed — it depends on which provider handled that specific call. Does UnitPay's margin tracking capture the real per-request cost when it varies like that, or is it built around a single expected/blended cost per action? That distinction is exactly the kind of thing I'd have gotten wrong building this myself.

Vijay — the line in your intro that jumped out was "what if they upgrade mid-cycle with a negative balance." I run WinBidIQ (AI bid intelligence for U.S. federal contracts), and we meter AI-generated proposal drafts plus daily federal opportunity data pulls per customer. Our version of that problem is fiscal-year-driven usage bursts: a customer can blow through a month's credit allotment in one week chasing an RFP deadline, then go quiet for three. Does UnitPay support any kind of temporary overage/grace buffer before hard-stopping at zero balance, or is it strictly prepaid-and-cut-off?

Congrats on launching, I really like the idea because the AI is scanning real websites and gives the recommended website interface. I'm just wondering what's the difference between Fudge and Lovable?

@Simdi Jinkins that makes sense, thanks. so in practice most of the "taste" enforcement ends up living in the agent's prompt/instructions rather than the tool itself - fudge just gives it the raw material to draw from. good to know, I was picturing more of it baked into the extraction step.

the part I'd want to be careful with is how close "measured design evidence" gets to a specific real company's actual brand identity versus generic patterns. pulling fonts/colors/layout from a category of sites as inspiration is fine, but if the agent leans hard on one particular close visual match, the output could end up looking like a clone of that specific brand rather than "good taste" in general. is there anything nudging toward blending multiple references instead of over-indexing on the single best match

"Better taste and not more adjectives" is painfully accurate. We spend a lot of time on moodboards for client design work, and we're actively integrating AI generation into that workflow right now — so searching real sites by fonts/components/layouts instead of prompting "cleaner, more premium" for the fifth time is exactly the missing piece. Definitely trying this one.

I like the concept of this, it definitely makes a big difference to agents having a solid design framework to work from.

1. Interesting idea. Using real design references instead of vague prompts makes a lot of sense

really like this framing of "taste" as something extractable rather than just scraping colors and fonts. what happens when the source site is inconsistent, like a company mid-rebrand where half the pages still use the old design system - does it pick up on the dominant pattern and ignore the outliers, or does it get confused and blend both into something that matches neither

The bit I like is it hands the agent measured design evidence — actual fonts, colors, spacing — instead of another screenshot to eyeball, plus the Chrome extension to save references as I browse. My day-one workflow question: does that saved-reference library live locally with the MCP or sync to a Fudge account, and when I'm on one project can I point my agent at just my saved set instead of searching the full 10k every time?

Tried it through Claude today and the visual similarity search is genuinely useful, found a couple of layouts I would have spent an hour hunting for on Dribbble. The Chrome extension saving references locally is a nice touch too.

honestly love the idea of giving agents actual references instead of vibes, that part is really smart. one thing though, it would be super helpful if you could filter by industry or specific site categories, like only e-commerce or only SaaS landing pages, so the results are way more targeted when you're working on a particular project.

Gary, needing one particular machine just to get an app out the door has always felt like a silly gate. Knocking that barrier down so more people can actually ship their ideas is lovely to see.

Does the App Store rejection feedback make it back to the agent through the MCP, or is that where a human takes over? Review is the one step in the pipeline that stays human on Apple's side, so I wondered where the loop closes. Congrats on the launch, Gary.

cool idea, I was thinking of launching an iOS app but having to get Xcode and a newer mac to run it was annoying. How is using this better than me just asking claude code to make it for me on a PC?

the no-xcode pipeline feels like it was built by people who actually shipped an iOS app from Linux and got tired of the friction. love that the MCP integration is first-class instead of an afterthought.

the "agent reads the rejection and ships the fix itself, no human action" answer is the part that would make me pause, not the no-Mac angle. Apple's review process assumes a person is accountable for what gets resubmitted, and repeated automated resubmissions after a rejection is the kind of pattern that gets developer accounts flagged regardless of how good the fix is. also curious where Apple Developer certificates and provisioning keys actually live in this pipeline, since that's the credential an agent would need custody of to sign and submit builds on its own

this is exactly the kind of tool that's obviously right once someone builds it - "just run tailscale switch" always felt like a workaround for a problem that shouldn't exist. how do you handle MagicDNS when both tailnets happen to have a machine with the same short hostname? does the routing-by-suffix approach sidestep that collision entirely or is it still something you have to configure around?

the fail-closed approach to ambiguous routes is the right instinct, most tools would guess and silently leak traffic down the wrong tunnel. one thing I'm curious about given it's running isolated embedded nodes per tailnet - what's the permission footprint on macOS, does it need a persistent background helper with elevated network privileges, and is it notarized/signed properly given how much trust you're asking for on the network stack

Congratulations on the launch! Managing multiple Tailscale tailnets has always been one of those small but recurring workflow frustrations, so this looks genuinely useful.

I'm curious about the networking side—how does TailMux handle overlapping IP ranges or conflicting DNS configurations when multiple tailnets are active simultaneously? Looking forward to trying it out.

Great app, I absolutely want to try it; it would be super convenient even for me, since I live on two different tailnets. If so, are there no profile limits, meaning it can support more than two tailnets?

Hey Product Hunt 👋

I'm an ex Optiver low latency guy working on ByteAsk.

Writing C++ was never really the hard part for me - proving it's correct is. You write a fix, then you're the one who has to remember to rerun it under sanitizers, step through gdb when it segfaults, and rerun the test suite before you trust the diff. That verification loop is manual, repetitive, and easy to skip when you're tired.

So we built ByteAsk, an AI coding agent for C and C++ that verifies its own work. It doesn't just hand you a diff - it drives the sanitizers, the debugger, and your test suite first, so what you see has already been proven to compile, run, and pass.

Why ByteAsk is different:

• Works with any model: Opus (anthropic), Gemini, Codex, GLM 5.2 - 15+ models supported.

• Bring Your Own API key: Add your own API key, no extra charges.

• Zero Data Retention: No code or prompts are logged. Only abuse prevention metrics like token usage are logged.

• Install in seconds: pip, uv, or npm - pick whichever fits your setup.

• Verifies before it shows you anything: instead of "here's a fix, hope it works," ByteAsk reproduces the bug, drives ThreadSanitizer/ASan/gdb/Valgrind, and confirms the fix actually holds.

• Built for the C/C++ toolchain you already use: native support for LLVM, GCC, gdb, Valgrind, and CMake — not a generic wrapper bolted onto a language it doesn't understand well.

• Runs from your terminal, in your repo: it edits your actual codebase, not a sandboxed snippet.

We built this because most AI coding tools treat C++ like just another language with slightly different syntax. It isn't - memory safety and concurrency bugs are where the real time goes, and that's exactly the loop we automated.

Try it → byteask.ai

this sounds really cool!! congrats on the launch Pratyush and Anirudh!

Interested in where ByteAsk is headed? We post every major release, benchmark, and product update on LinkedIn:

https://www.linkedin.com/company/byteask/

C++ as a language gives broader control which makes it easier to shoot yourself in the foot. Memory management, thread safety, performance engineering are some of the fundamental concepts we needed to make sure it works in practice. One of the hardest challenge was to create the code graph over C++ with lots of templates. This requires working over compile time AST which can grow exponentially before making any edits.

You can directly get support from me and community of 100+ volunteers working on this project here: Discord community: https://discord.gg/kENTnW6ubW

This is really cool. I just tried ByteAsk, and I'm genuinely impressed by how frictionless it is. It fits naturally into an existing coding workflow, and it's surprising how well it handles even complex C++ problems. Huge kudos to the team. You guys are onto something. @anirudhabyteask @pratyush1505

Anirudha, so much of what a machine writes for me looks right and then quietly falls apart the moment I run it. Something that catches its own mistakes before I ever see them would earn my trust fast.

Sources:

1. All benchmarks: https://x.com/4n1rudh4/status/2069108126524359115

2. Github: https://github.com/ByteAsk

3. Documentation: https://byteask.ai/docs/overview

4. Privacy policy: https://byteask.ai/privacy/

Installation Sources:

1. PIP: https://pypi.org/project/byteask/

2. JetBrains: https://plugins.jetbrains.com/plugin/32655-byteask

3. Cursor: https://open-vsx.org/extension/byteask/byteask-vscode

4. NVim: https://github.com/ByteAsk/ByteAsk.nvim

5. Emacs: https://github.com/ByteAsk/byteask-extensions/tree/main/emacs

6. NPM: https://www.npmjs.com/package/@byteask/cli

Dave here, CPO at Proxon. I've run dozens of customer discovery calls in recent months, and the response has been overwhelming. People start pulling in their eng leads mid-call, before I've even finished the demo. "This is the exact visibility I've been looking for" is the common refrain. That reaction is why we still do this, and why we’re so excited about Proxon.

We’re still early, and still eager to talk to customers. I’d love to chat with you about your AI adoption story!

Finally gave it a spin and the cost breakdown by team was genuinely eye opening, we had no idea how much some departments were burning on API calls. Mapping data exposure across all our agents in one view feels like the kind of thing every ops lead has been cobbling together in spreadsheets.

Hey Product Hunt! 👋

I’m Santiago from Proxon.

A quick story on why we built this: Over the past year, almost every founder and operations leader we spoke to told us the exact same thing. Their teams were adopting AI tools, custom agents, and automated prompt loops incredibly fast, but leadership was completely blind to it. We kept hearing the same questions: What are these tools actually costing us? What data are they touching? And who owns them if something drifts?

Managing AI through scattered invoices, random policy docs, and manual spreadsheets just doesn't scale.

That’s why we built Proxon, the definitive management layer and system of record for your AI workforce.

Here is what you can do with Proxon starting today:

• Discover the Workforce 🔍: Map every AI tool, agent, prompt loop, and shadow AI asset running across your organization.

• Govern & Assign Ownership 🛡️: Establish data policies, review cadences, and approval paths so your team can move fast safely.

• Attribute Spend & ROI 💰: Tie vendor and model costs directly back to specific teams and workflows instead of guessing at line-item invoices.

• Propagate What Works 📈: Automatically extract high-performing AI workflows from your power users and deploy them as templates for the rest of the team.

We’ve been building this heads-down with a team that cares deeply about getting enterprise AI right. Today, we’d genuinely love your honest feedback: the good, the bad, and the rough.

I'll be right here in the comments all day, so ask me anything! What do you think? 👇

Maker here. We built Proxon because we were flying blind on what our own AI agents were actually doing — and spending. Every tool showed you tokens. None of them told you the work: which agent, which task, what it cost, whether it was worth it.

So we built the thing we needed. It captures every agent's activity and turns it into cost intelligence you can actually act on. We've been dogfooding it on ourselves for weeks and I genuinely can't run without it now.

This is day one for us. Would love your brutal feedback — we read everything. 🚀

Dev on the @Proxon team here 👋.

We kept hearing the same thing from teams adopting AI fast: nobody had a clear picture of what was actually happening across all the tools once it spread past a handful of early adopters. Spend, usage, ownership — all of it scattered across vendors with no single view. That gap is what got us building.

Personally I just wanted to build something people would actually enjoy using, not just tolerate. That's been the bar for me the whole way through.

Really proud of where it landed. Excited to finally have this out — let us know what's rough, we're reading everything.

Amazing team that is spot on with this problem. How does onboarding and setup work?

Co-founder here. We built Proxon because everyone we talked to was burning real money on AI — Claude, GPT, Cursor, all of it — and nobody could say who was actually using it, whether it was helping, or where the spend went.

Every tool that tried to answer that felt like surveillance. We went the other way: aggregated, privacy-safe views for managers, personal dashboards for everyone else, and a little recognition when you ship your first agent. Less Big Brother, more leveling up.

Curious how you all handle this — how do you measure if AI is actually paying off at your company? Around all day, ask us anything.

the "who is using AI" question is the one I'd actually want answered honestly - a lot of the real exposure isn't the sanctioned agents your team built, it's someone pasting a customer contract into a random ChatGPT tab because it's faster than asking IT for access to the approved tool. does Proxon have any way to see that shadow usage, or is it scoped to the tools/agents that are already connected and reporting in?

the "real computer control, not just suggestions" part is what would make me hesitate, not the local vs cloud question. an agent that can actually click and type on my desktop is a much bigger blast radius than one that drafts a suggestion I approve. is there a confirm-before-action step for anything destructive (deleting files, submitting forms, sending money), or does it just act once it decides on a plan?

Local models are a neat idea, but aren't there free easy ways to get Ollama to run on a Mac? What is better about using WIllow?

€299 lifetime with real computer control is a bold bet against subscription fatigue, but lifetime pricing on something that needs ongoing updates to keep working with new OS versions/model APIs makes me nervous long term - how are you planning to fund that without a recurring revenue stream once the initial sales dry up?

the "clone a live site into a sandbox, share it with anyone via a link, no login needed" combo is the part I'd think twice about - if the sandbox is an exact clone of a client's production site with no auth, doesn't that mean any unpublished drafts, real customer data in the DB, or anything else on that site is now sitting on a public URL with no gate at all? feels like the disposability that makes this useful for testing is also what makes it easy to forget a sandbox is still live and exposed

A one-click "reset to clean slate" button after each test session would save a lot of time, so I don't have to manually wipe installs between experiments.

Would be huge if you could share a sandbox via a simple public link so clients can preview changes without needing their own login. That kind of frictionless handoff would make this my go-to for client work.

The instant spin-up is genuinely impressive, had a clean WordPress site ready in under a minute. Also appreciate that I can blow it away without worrying about my real installs. Solid for quick testing.

one thing that would be super useful here is letting users share a sandbox with a teammate via a simple link, since right now it seems like everything is locked to your own account and collab on staging stuff kind of requires that

disclosure regimes are genuinely different state to state in Australia - Section 32 is a Victoria thing, NSW has its own contract/disclosure requirements, and it's different again in QLD. does the model actually know which state's rules apply based on the document format and check against that state's specific requirements, or is the 40+ risk category scan more of a general pattern applied everywhere regardless of jurisdiction?

for something touching property contracts the real question for me is liability, not accuracy. if the tool misses a genuinely risky clause and someone signs based on a clean-looking report, is that positioned as "supplement your conveyancer, don't replace them" with clear disclaimers, or are people actually using this standalone to skip paying for a solicitor review

Buying a place is one of those moments where you nod along to pages you barely understand and hope nothing bites you later. Having a calm second pair of eyes on all that fine print feels genuinely reassuring, Anuj.

One thing that would help a lot is comparing clauses side by side against a standard template so I can see at a glance what is unusual or missing. Even a simple diff view would make negotiation prep faster when reviewing contracts at open homes.

Really cool idea, especially for sports and weather-based ads. What happens if something changes right after the trigger, like a goal being overturned?

I definitely like the concept behind this - congratulations. And sports and weather are great triggers to build from. What other triggers do you have in the pipeline? Stock market certainly seems like a good one to include. Plus (similar) money exchange markets. My only comment is there seems to be a significant jump in price between Agency and Enterprise. Would be useful to understand exactly what is offered in that plan... because the "every minute" refresh feels important for a sports related campaign. But the pricing feels quite painful! One last question... do the ad providers (Meta/Google etc) respond quickly enough to ensure that the wrong ads aren't cached when sport scores change?

The flat $19 with AI included jumped out, most of the tools you're replacing price per message and you called that out yourself. What keeps it from breaking, a cap on drafts somewhere, or is model cost per conversation just small enough that it never matters? Congrats on the launch, Artur.

It's a nice idea to consolidate, but what makes users trust this product enough to connect it to their communication software?

honestly the flat $19 pricing is really nice, one thing though would be great if you could set up auto follow ups for unanswered conversations after like 24 hours, basically a nudge that goes out so nothing slips through the cracks

honestly this looks pretty solid, the flat pricing is a nice touch. one thing though, can you add a way to set up canned responses or templates per channel? like, sometimes my team needs slightly different replies for whatsapp vs email and right now it feels like it might be one shared library

the oauth integration makes this a really smooth experience and the github stats base the profile in reality, well done.

honestly the player card idea is kind of fun, made browsing way more engaging than another boring list. plain english search worked pretty well when i tried it too.

the part I'd want to understand before trusting a score is what it does to someone who does great work at a company with a private repo all day and only pushes weekend side projects publicly, versus someone who commits noisy formatting changes and doc typo fixes to public repos constantly. public GitHub activity correlates with output but it also correlates with how much of your job happens to be open source. does the scoring try to normalize for that at all, or is it explicitly "public output only, make your peace with it"?

The player card scoring idea feels genuinely fresh, like turning GitHub activity into something you'd actually want to browse rather than another dry analytics dashboard.