PH热榜 | 2026-07-05

Hey Product Hunt,

I'm part of the team that built WorkBuddy.

AI made the thinking part faster, but you still spend hours turning "the response" into an actual file. WorkBuddy closes the execution gap with a team of AI experts that helps produce sharpened, ready-to-deliver results.

Here's how it works:

1. Pick an expert team.

2. Describe what you need in plain language.

3. The team gets to work in parallel, dividing the task, cross-checking each other, and synthesizing everything into sharpened, ready-to-deliver results.

Not sure which way to go? You can also bring in another expert opinion to guide the direction.

Why you want WorkBuddy:

• A team, not a bot: experts cross-check each other into sharpened results

• 100+ Pre-built Expert Teams: across every domain, call them like a colleague, not a tool

• You stay in control: bring in another opinion and guide the direction, at any point

• Parallel by default: multiple agents run in parallel, no waiting in line

• Real deliverables: finished files in your folders, not trapped in a chat

To celebrate our launch, we're giving the first 300 users who come from Product Hunt an extra 500 Credits. First come, first served. Claim by July 20 🎉 Claim here → [link]

Try WorkBuddy today 👉 workbuddy.ai

Really like finished files in folders instead of answers trapped in chat! The last mile is where agent tools usually stall. QQ - when 2 experts genuinely disagree at synthesis, who wins? Congrats on the launch!

Congrats on the launch! This looks like it could save a lot of time for managers. Do you have examples of tasks where WorkBuddy performs better than a single-agent workflow?

100+ pre-built expert teams is a lot, BTW how do you actually find the right one fast? Is there search, or do you browse by category?

the cross-check between experts is the part i'm most curious about - when one expert flags something, does that actually change the final synthesized output, or does the merge step just blend everyone's take together and the disagreement quietly disappears? in my own multi-agent setups the hard part was never getting different perspectives, it was making sure a real objection survives the final combine step instead of getting smoothed over

For enterprise teams, governance will matter. Are there admin controls for which files WorkBuddy can access, where outputs are saved, and what context is remembered?

This looks like a huge time-saver! I've been struggling to coordinate different AI tools for my projects, so having a unified team of AI experts sounds ideal. How exactly do the different AI "experts" communicate with each other within the platform?

The disagreement-survival question Omri raised is the one that bit me hardest building this stuff. What worked for me wasn't a smarter merge prompt, it was giving each expert a flag field the synthesizer literally can't drop, so a minority-of-one objection still renders in the final output. The moment I let the combine step weigh objections by confidence, the useful contrarian take got averaged into mush. Do your experts write to a shared structured object, or hand back prose the synthesizer re-reads and re-summarizes?

Congrats team! The expert team concept feels much closer to how people actually collaborate at work.

I like that you're showing how people can guide the AI instead of pretending it always knows the right answer. What happens if two of the AI experts recommend different approaches? Does the system try to resolve that on its own, or does the user decide which direction to take?

I’d like to have Workboddy take over my browser to perform repetitive tasks, but I’m very concerned about the potential leakage of my account details, payment information, and personal data. How do you handle the uploading and protection of personal information?

@sherina_chen Congrats to the entire team on this success. Looking forward to see how impactful the product will be as time goes on. how does your product deal with agents working cross function? in a situation where 2 different agents could possibly do different task but have 2 identical task they can both do, which of the agents will take up the task?

I hope you do understand my question?

Have been kicking the tires on this one today and the multi-agent setup is honestly pretty clever. Got a messy brief turned into a clean summary and follow-up email in one go. Curious how it handles more technical docs.

How does the pricing work for the AI expert team — is it a flat subscription or do you pay per agent or per task?

Used it for a quick report draft and the multi-agent angle actually helped, gave me three different angles in like a minute instead of one generic answer. Slightly clunky UI but the results felt noticeably more useful than my usual single-prompt tools.

Congrats on the launch. The part I'm most curious about is data boundaries rather than the synthesis logic - if I spin up a team with a research expert and a legal expert on the same brief, do they see each other's full working context and any files I uploaded, or is each expert scoped to only what it needs for its part? Multi-agent setups are great until one expert quietly has more access than the task calling it actually required.

The part I'd poke at: asking in the brief is prompt-level, and in my runs a plain-language 'keep the dissent' held up fine until the context got long, then the minority take quietly vanished maybe 1 in 5. Is preservation actually enforced in the synthesis step, or is it the model choosing to honor the instruction each time? That gap is what decides whether I'd trust it on anything with a real decision attached.

Tencent's name on this raises an immediate data question for anyone considering it for actual office work. Where is the processing happening, what data residency options exist, and is there an enterprise tier with clear answers to those questions before someone puts sensitive work documents through it?

Poked around with it for a bit and the "second opinion" angle genuinely caught me off guard in a good way, it actually feels like having a teammate double-check your work instead of just another chatbot agreeing with everything you say. Results came back clean enough to drop straight into a doc.

The "finished files in your folders, not trapped in chat" line is the strongest part for me. That last mile is where a lot of AI tools still feel unfinished.

Curious how much control users have over the final deliverable format. Can someone define a preferred structure or template once and have future expert teams follow it?

Interesting idea! I'm pretty set with using Claude Code for everything, and I think most people by now use what they are used to. How will you convince users to switch and try WorkBuddy instead of what they are used to?

Also cool logo, here's a free static QR code that goes to your website that I made for you to use if you want:

Love the idea of experts cross checking each other instead of just one model guessing alone, that debate before delivery feels like the real innovation here.

Congrats on the launch @sherina_chen

When two agents are running in parallel and one goes off track, does anything catch that before it bleeds into the final output? Or do you only find out at the end?

Congrats WorkBuddy team.

the "bring in a second opinion" framing is a nice touch. most AI-work tools stop at one answer, and leaning into an expert-panel idea instead is more how real teams actually decide things. curious how you keep the experts from just agreeing with each other. congrats on the launch.

I like that this is built around everyday office work instead of abstract agent demos. The positioning feels grounded.

This feels very relevant right now.

Docs used to be "just" onboarding and support, but now they also decide what AI tools and agents understand about your product. If the docs are stale, the agent context is stale too.

We’re working on our own help center and llms.txt setup for our product, so I really like the idea of treating documentation as a maintained knowledge layer, not a side project someone updates when they remember :)

Curious how DocsAlot handles product changes over time. does it detect when docs are outdated from changelogs/GitHub/product updates, or is the maintenance workflow more manual right now?

Hi Product Hunt, I’m Faizan, founder of DocsAlot.

We built DocsAlot because more software is now being discovered, evaluated, and used by AI agents, but most products are still documented in a way that only really works for humans.

That creates a real adoption problem. If an agent cannot understand your docs, find the right setup path, or use your product without guessing, it becomes much harder for that product to show up in AI workflows and actually get adopted.

DocsAlot helps teams create, clean up, maintain, and distribute documentation that works for both humans and AI systems. That includes help centers, knowledge bases, developer docs, API docs, CLI docs, and AI-readable outputs like llms.txt, skill.md, and hosted MCP access.

What makes us different is that we are not trying to be just another docs editor. We are building the maintained knowledge layer that helps products become easier for agents to understand, recommend, and use, so documentation becomes part of agent adoption instead of just a support artifact.

Happy to answer questions all day. Thanks for checking out DocsAlot.

Faizan, this lands at the right time :) But my honest first thought is the one you'll probably hear a lot: llms.txt, skill.md and hosted MCP are becoming checkboxes, Mintlify and GitBook are already bolting them on. Emitting an AI-readable format won't stay a differentiator for long.

The line that actually caught my eye is the one you dropped to Andras: "data on how agents traverse help-centers." That feels like the real moat, using how agents actually read docs to restructure the content for them, not just expose it in a format everyone will have in 6 months. Is that where you're heading (scoring and reshaping docs for agent comprehension), or is the core bet still the unified output layer?

Congrats on the launch! ;)

Connected my Notion help center and the hosted MCP endpoint worked first try, which never happens for me. The llms.txt output was surprisingly clean compared to what I had hacked together before.

Finally, someone is tackling this! Writing docs that LLMs can easily parse while keeping them readable for human developers is such a tricky balance right now. Does this integrate directly with GitHub repos to keep the docs synced with the codebase?

Love how DocsAlot bakes llms.txt and skill.md right in instead of treating them as afterthoughts, that little detail shows the team actually understands how AI agents consume docs in practice.

the "detect outdated docs from source-code" part is the piece i'd want to stress test before trusting it on a real repo. false negatives are one thing, but false positives (flagging a doc as stale when the underlying behavior didn't actually change) seem like the bigger risk since that's what erodes trust in the tool and gets people ignoring the suggestions after a few bad flags. how does it decide a doc is stale, does it diff behavior or just correlate with commit/PR activity touching related files?

qq does this support openapi/swagger imports or is it text only for now? congrats for shipping today @haya_jawed

The MCP + llms.txt combo makes sense for the AI-answers use case. One thing I'd want to know before pointing this at our repo: since it's reading source code to detect stale docs, how does it handle internal-only context, things like code comments referencing unreleased features or internal tooling that shouldn't end up summarized into a public-facing llms.txt. Is there a way to mark certain source paths as off-limits for the AI-facing output, or is that on the roadmap?

The "agents stop reading stale context" line is the real problem statement here — most doc tools optimize for human search, but agent context freshness is a different failure mode entirely. How are you handling versioning when the underlying docs change — does the MCP endpoint serve live content, or is there a sync/caching layer in between?

how does docsabot actually keep things in sync when a source article changes, is it just polling or is there some kind of webhook setup

AWESOME!

The failure mode I keep hitting wiring agent docs through MCP: the agent grabs a confident, plausible snippet that's one version stale and never flags it, because retrieval has no notion of 'this section is old.' Emitting llms.txt is the easy part. The valuable bit is a freshness signal inside the MCP response itself, so an agent can tell a current param from a deprecated one. Does your outdated-doc detection surface a staleness marker in the MCP payload, or only in the dashboard?

I like the focus on keeping human facing docs and agent readable documentation tied to the same source

how do you handle versioning and change tracking when product docs or API references are updated frequently?

"Your docs show up in AI answers" depends entirely on how the underlying models are trained and updated, which DocsAlot doesn't control. The llms.txt standard is still not universally respected across major models and crawlers. What's the realistic expectation for how quickly and consistently docs actually surface in AI answers after setting this up, or is that more of a long-term bet on the standard gaining broader adoption?

FYI, I should mention that as part of this launch we are offering 50% off on all pricing tiers, for 3 months. But the pricing is only valid today.

docs used to be written for humans who were already lost. now they're also the thing that decides whether an AI agent gives your users correct information or confidently wrong information. the MCP + llms.txt combo is the right bet, that's basically the emerging standard for "here's how to actually understand my product." curious how you're handling versioning, when the product changes does the knowledge base propagate updates automatically or is that still a manual step?

This is really amazing. Do you have a plugin to embed in docs or we need to pass the doc url to generate the context

Love the concept—making docs dynamically ready for AI agents and managing llms.txt is brilliant.

From a quick CRO Audit lens on your mobile landing page, here are 3 quick conversion leaks you can patch today:

Flipped CTA Priority (1000017242.png vs 1000017252.png): Near the top, "Try DocsAlot" is primary (Blue). At the bottom, "Request AI audit" suddenly becomes the primary blue button. This role reversal confuses the user's visual habit.

Microcopy Friction: Mixing "Get AI visibility audit" and "Request AI audit" creates minor text inconsistency. Standardize the verb to smooth the funnel.

The Non-Technical Barrier (1000017245.png): The terminal CLI code block is great for devs, but it creates friction for non-technical buyers (CEOs/PMs) who hold the credit card.

Fixing these visual bugs could easily boost your signups from today's traffic. Best of luck! 🙌

This feels timely. Docs are no longer just for users and support teams. AI agents also need a reliable source of truth now.

Curious how DocsAlot handles drift over time. If the product changes, does it detect outdated docs automatically, or do teams still need to manually trigger updates?

Been watching this space for a while and the "banking at home" framing nails the actual pain. The five-day wait and hidden FX are exactly what kills margin for anyone selling across borders. Congrats on the launch.

Hey Product Hunt 👋

I'm Ashita, co-founder of Endl.

We built Endl because every founder outside the US knows this pain: you earn in dollars, but spending them is somehow still a problem.

You're paying for Claude, Cursor, AWS, Notion, Linear. Your card declines. Or you're converting local currency, paying FX fees, and losing money just to access tools your US counterparts use without thinking about it.

We built the account we wished we had:

• Receive payments from global platforms in stablecoins

• Hold your balance in digital dollars, no conversion needed

• Spend on any SaaS, ads, or tools with a global card

• Earn yield on idle balances

• No US company required

Built for founders and businesses operating across Southeast Asia, MENA, and Latin America.

We've processed $100M+ in volume across 500+ businesses.

If you're a founder outside the US who's ever had a payment fail on a tool you needed, I'd love to hear what's breaking for you.

— Ashita

Congrats on the launch @ashita_batra1

Where is Endl based out of? And is it part of any regulated authority?

the "no US company required, hold digital dollars" pitch is appealing but it's also the part i'd want to poke at hardest before actually moving real revenue through it. stablecoin regulation is still shifting country by country, and if a jurisdiction you operate in tightens rules on holding or spending stablecoins, what's the fallback for a business that's already routed its whole payment stack through Endl? is there an easy off-ramp back to plain fiat rails, or does that become its own migration project

How are you handling compliance across 160+ countries without making onboarding painful?

Congrats on the launch 🚀

This is a space that still feels much more fragmented than it should be.

As someone building a startup from Europe, I've definitely felt how quickly global finance becomes messy once you have international customers, contractors, different currencies, and eventually stablecoins entering the picture. Having one operating account instead of stitching together five different services is a compelling direction.

The stablecoin support is especially interesting. it feels like we're finally reaching the point where it's becoming practical infrastructure rather than just a crypto feature :)

Curious which type of company has been the biggest early adopter so far. SaaS startups, agencies, marketplaces, or something completely different?

I like the unified workflow concept. can finance teams set different permissions and approval flows for multiple team members?

For product teams specifically, is there an API-first setup here or is this more of a dashboard tool that finance ops people live in day to day?

Reconciliation is usually the most painful part of cross-border ops honestly, it just breaks everything downstream. quick Q, does it handle multi-currency reconciliation automatically or is there still some manual matching involved?

most fintech tools solve one piece (payouts OR FX OR reconciliation).
Having all of it in one place sounds great but how hard is the initial setup for a team already using 3 different tools for this?

How does the onboarding process differ for freelancers and registered businesses? Are there any countries or business types that are currently not supported?

How quickly would a freelancer normally receive a payment through Endl? Does the speed depend on whether they choose a bank transfer, local currency, or a stablecoin?

Great Launch! What would you say is the biggest advantage of using Endl instead of Wise or Airwallex for a small business with a distributed international team?

Congrats, team! How do you compare with Mercury for non-US founders who need USD banking details but also regularly make international payouts?

Do the corporate cards support real-time controls such as freezing a card, changing its limit, or restricting certain types of merchants?

Is there an API available for businesses that want to automate payouts to contractors or integrate Endl into their own finance workflow

I like the idea of combining collection, holding, cards, and payouts in one platform. Which part of this workflow do your current customers find most valuable compared with using several separate providers?

That's clever. Does the card support international payroll or is it mainly for vendor payments?

Great work! What BIN country do your cards carry? If US, that single detail does more for a founder in MENA or LatAm than the rest of the stack combined!

Congrats to the Endl team on the launch.

For companies dealing with USD collections, contractor payouts, stablecoins, and cards across markets, the real value here is less about payments alone and more about reducing operational drag in global finance workflows.

What happens if the value of the stablecoin used for the balance temporarily moves away from $1? Does Endl provide any protection or automatic conversion options?

How does pricing actually work when you combine fiat accounts with crypto payouts in the same flow, is it per-transaction across both rails or broken out separately?

Could Endl also work as a treasury tool for startups that raise or earn revenue in stablecoins but still need to pay most of their expenses through traditional banking rails?

How do you handle compliance and regulatory requirements across different countries, especially with crypto involved?

Curious how reconciliation actually handles mismatches when crypto and fiat settlements hit at different times across chains and bank rails. Is it fully automated or do teams still need to babysit edge cases?

Congrats on the launch! Can a company create approval workflows for larger payments, so that one team member prepares a transaction and another person approves it?

Hey Product Hunt,

I’m Ben, and I’m building TryCase.

This came from my own workflow. I’ll often have a bunch of agents running at once across different worktrees, each trying changes, spinning up the app, testing, iterating, taking screenshots, or recording proof.

That gets messy quickly. My laptop becomes the bottleneck, ports collide, installs overlap, browser sessions get reused in weird ways, and I still end up doing a lot of the final verification myself.

So I built TryCase to give each agent its own disposable Linux environment in the cloud. The agent can run the app, test the change end to end, capture screenshots or video, and come back with proof instead of just code.

It’s also useful for longer-running tasks. You can give an agent a goal, let it work inside a clean disposable environment, and ask it to come back with screenshots, logs, and recordings. Secrets can be passed in deliberately, and each run is isolated from your laptop and from other agents.

TryCase is still early, but the goal is simple: agents should only say “done” once they’ve actually run and verified the work.

It’s easy to try. Just ask your coding agent to use TryCase at trycase.dev:

- Fix this bug, test it end to end with TryCase, iterate until it works, and send me screenshots and a video recording as proof.

- Implement this feature, run the app in TryCase, iterate on any failures, and prove it’s working with screenshots, logs, and a recording.

- Use TryCase to run this repo in a clean environment, verify the main flow, and show me what the app looks like.

- Test this branch in TryCase, find anything that breaks, fix it, and prove the final version works.

- If manual login is needed, use desktop mode and give me the take-control link.

I’d love feedback from people using Codex, Claude Code, Cursor, or other coding agents. What would make you trust an agent’s “done” more?

Such a smart concept for vibe coding. Testing AI-generated code safely without messing up my local environment is always a massive concern. Can these disposable environments be spun up locally, or is the platform entirely cloud-based?

the 'done only after it actually ran and verified' bar is the right one. do you surface the failed attempts too, or just the final passing proof?

"agents should only say done once they've actually run and verified the work" is exactly right and it's the thing i keep running into with my own multi-agent setups - an agent will confidently report success based on its own transcript when the thing it was supposed to do never actually happened underneath. running multiple worktrees locally does turn into port collisions and reused browser sessions pretty fast like you said. does the screenshot/recording proof get attached anywhere the agent itself can't fake or reword, or is it still on me to actually look at the video rather than trust the agent's summary of it?

Love the framing around agents only saying done after screenshots/video proof in an isolated env — that is exactly the gap when running multiple worktrees locally. Good luck with the launch.

Congrats Ben — the part that resonates is “verified code instead of asking you to test manually.” In my own coding-agent loop, the weak point is not generation, it's proof quality: agents say “done” after lint passes, but the actual product flow is still broken.

A small thing I’d love to see in TryCase is a compact verification receipt: commands run, browser path tested, screenshots/video links, and what failed before the final pass. That would make it much easier to trust the result or review it later.

Curious if you're thinking about a standard “done packet” that agents can return to humans or CI?

@TryCase Giving coding agents an isolated, disposable Linux sandbox is exactly how we solve the local port collision and dependency nightmare. Since TryCase is exposed as a runtime tool layer for external agents, how are you handling base image caching? If an agent executes a massive npm install or updates a database schema across multiple iterative debugging runs, do you snapshot and delta-cache that specific container's filesystem state, or does it do a clean cold boot every time the agent invokes a test call?

As a solo dev shipping an iOS app, the final manual-verify step is exactly where my releases stall — I'm the QA team of one, and "done" from an agent usually just means "it compiled." The disposable box per agent is a smart way to move that check off my laptop. Does it handle mobile/app flows (simulator or a device) yet, or is it web + CLI apps only for now?

This maps to the messy part of agent coding for me: not the patch itself, but proving it ran in a clean environment. The useful constraint is keeping the proof lightweight enough that agents actually include it every time.

Love how the screenshots and recordings come back with the diff already attached, makes verifying what the agent actually did way less of a guessing game.

the disposable linux envs actually working from a single prompt blew me away, screenshots and all. finally lets the agent go end to end without me babysitting every shell command.

how does it handle cleanup of those disposable environments when an agent spins up a bunch in parallel, anything to worry about resource-wise?

finally something that lets my agent actually run the code instead of me playing QA. tried it on a small flask app and got back screenshots plus a clean diff, which honestly saved me a whole round trip.

Congrats on the launch, this is a real problem. My question is more about the iteration loop than the verification side - if an agent fails, tweaks the code, and needs to test again, does it get a brand new disposable box each time, or does it reuse the same one until the task is done? Fully fresh every retry sounds cleanest for isolation but on a repo with a slow install/build step that could add up fast if the agent is iterating 10+ times on one bug.

I’m curious about the differences between GPT-5.5 browsing and testing and TryCase. What makes TryCase unique, apart from just screenshots and videos? Thank you

How do you handle state between runs if the agent needs to verify something like a running database or queued background jobs from a previous step?

Screenshots and recordings as verification artifacts are useful for UI changes, but for backend logic or API behavior the visual output doesn't tell you much about whether the code actually works correctly. What does TryCase return as verification evidence for non-visual changes, like a database write, a webhook handler, or a background job, and how does the agent know the difference between "it ran without error" and "it did the right thing"?

the disposable Linux environment idea is such a clean solve for the "I cant actually verify this" problem with coding agents. nice execution.

finally something that lets my coding agent actually run the app instead of just staring at it. the screenshot capture came back clean on the first try

The 'an agent handles verification' step is where I've watched this quietly break. When the same model family does the work and the check, the verifier tends to trust the doer's framing of what success looks like, so it happily confirms a screenshot of the wrong screen. What helped me was feeding the verification agent only the original task spec plus the artifact, never the doer's transcript, so it can't inherit the optimistic story of what happened. Does TryCase hand the checker the full run log, or just the recording and a fresh prompt?

'Return verified code instead of asking you to test manually' is the exact gap. My coding agent writes the fix, then I'm the one clicking through the app like it's 2015. The agent proving its own work with screenshots and recordings flips the trust equation completely. How isolated are the environments - can it safely test against a copy of production data? Congrats on the launch.

congrats on the launch ben. the port collision thing you describe hit me the first time i let two agents run dev servers at once, so the throwaway box approach makes a lot of sense. one thing i'm curious about: booting an app end to end usually means real env vars, api keys, db urls. where do those secrets live while a sandbox runs, and does teardown wipe them for good?

Hey Product Hunt! 👋

I'm Marcos, maker of MentionDrop.

MentionDrop watches the places where buyers actually talk: Reddit, Google News, search results, and selected public web pages. AI triages every mention: summary, sentiment, relevance, and whether to reply, share, monitor, or ignore.

Today's launch gives your AI agent that same view. Connect Claude, Cursor, or Windsurf with an API key and ask:

• "What should I pay attention to today?"

• "Find competitor complaints from the last 7 days"

• "Draft a reply for the most urgent mention"

11 tools, read-first design. Your agent drafts, you decide what gets posted.

Honest boundaries: we do not monitor X, LinkedIn, or "the whole internet". Bounded sources, useful mentions.

Launch offer: 14-day free trial plus free MCP setup help. I will personally help you create your first monitors and connect your agent.

Tell me what your agent workflow needs. Ask me anything!

love that you name the limits out loud (no X, no 'whole internet') instead of overclaiming. does the agent learn which sources matter per brand over time?

@MentionDrop The read-first design with account-scoped API keys is a brilliant security boundary for B2B engineering. If an organization has multiple distinct autonomous agents or team members concurrently querying the MentionDrop MCP server for the same client brand keyword, how do you handle rate-limiting and query caching on your end to prevent overlapping API consumption spikes?

Setting it up caught way more niche forum mentions than I expected, and the AI summaries actually pull out the useful bit instead of just repeating the sentence. The sentiment scoring felt a little hit or miss on sarcastic posts but the suggested actions save me from digging through everything myself.

How are you handling the massive scale of 8 billion pages scanned daily? Is your AI summary and sentiment analysis based on a specific NLP library or a custom implementation?

How does it handle false positives or really low-quality mentions, like random forum spam or machine-generated content? Curious how strict the filtering is before something actually lands in your dashboard.

how does the sentiment scoring hold up across languages that use sarcasm or idioms a lot, like portuguese or japanese?

the "nothing auto-posted, drafts for review" part is what separates this from the stuff that gets brands in trouble. i've been doing something similar by hand across a few communities and the actual bottleneck is never finding mentions, it's triaging which ones are worth a human reply versus noise. curious how the triage decides what's worth surfacing versus what gets buried, is that tunable per source or is it one global scoring model right now

Could this be used to monitor product launches like Product Hunt and automatically surface high-priority conversations?

How does the AI handle sarcasm or context-heavy mentions where sentiment isn't obvious, and does that affect the action suggestions it makes?

The AI summaries actually capture nuance better than I expected, especially the suggested actions which feel useful rather than generic. Setup took a couple minutes and it picked up mentions from forums I forgot even existed.

The MCP angle here is smart, pulling this into the agent's workflow instead of another dashboard to check. Question on the foreign-language sites part of the pitch - sentiment scoring is already tricky in English with sarcasm and industry slang, how does accuracy hold up once you're scoring sentiment on a mention that's been through translation first? That seems like the place this could quietly go wrong without anyone noticing until a genuinely bad mention gets triaged as neutral.

The sentiment scoring paired with suggested actions is such a thoughtful touch, turns a noisy firehose into something you can actually act on instead of just staring at.

Finally a tool that catches the foreign language forums I always missed, and the AI summary actually saves me from scrolling through each thread.

How does it decide which sources count as relevant mentions, and is there any way to scope it down so it doesn't drown me in noise from tangentially related sites?

The sentiment score on mentions is useful but the suggested action is the harder part to get right, since the right action for a negative Reddit thread about your product depends heavily on context that's hard to capture automatically, like whether it's a power user venting or a prospect researching. How does MentionDrop avoid the suggested action defaulting to "respond and engage" for everything, which would just be noise?

The nothing-auto-posted, drafts-for-review boundary is the right default for an MCP tool — an agent that has live web signal AND can publish is how brand incidents happen. With account-scoped API keys, is a key's rate/quota shared across all 11 tools or metered per tool, and can I scope one key to read-only mention pulls without exposing the reply-drafting tools? I'd want to hand a teammate's agent the monitoring half without the outreach half.

The 'suggested action: reply / share / monitor / ignore' field is the interesting bit once this is behind MCP. When Claude or Cursor pulls a mention in, does that action come back as a plain label the agent re-reasons over, or is it structured enough to wire straight into an autonomous loop? The failure mode I keep hitting with signal-feed MCPs is the agent slurping 40 mentions into context and burning the window before it ever triages, so I'm curious whether you pre-rank or paginate server-side rather than handing back the raw firehose.

Skipping X and LinkedIn is an honest call, but that's often where brand mentions actually happen for consumer products. How do teams work around that gap, run a separate tool alongside MentionDrop for those platforms?

That feeling when AI agents have better interaction than humans :D

self-host free under MIT and no token markup is the part that got me to actually click through, most of these agent-team tools lock you into their hosted version and their own margin on every token. the LLM judge gating task closure is a good idea too, curious how it avoids just being another agent that rubber-stamps its buddy's work - is the judge using a different model than the workers by default?

The LLM judge before a task closes is a smart guardrail — kanban plus channels is closer to how I actually want agent teams to report back than another generic group chat.

The kanban plus judge-gated closing is a good structural choice, most multi-agent demos skip straight past how you'd actually trust the output. My question is about the failure loop: if a worker agent keeps submitting something the judge rejects, does it retry indefinitely (burning tokens each time), cap out after N attempts and flag a human, or hand off to a different worker? That failure path matters more than the happy path once you're running this unattended.

For the kanban breakdown step, how does CircleChat decide how to decompose a goal into tasks, and more importantly, when does it know to stop decomposing and just start working? Over-decomposition is a real failure mode in multi-agent systems where agents spend more time planning and reporting than actually producing anything useful.

Love how the objective input sits front and center before anything else, it makes the whole multi-agent idea feel way less intimidating. Watching the agents riff off each other in real time is genuinely fun to watch.

How does CircleChat actually pick which agents join the conversation, and do I have any control over which models or personas show up?

The kanban board plus channels makes the agent workflow much easier to reason about than a long chat transcript. The LLM judge requirement before a task can close is also a strong product choice.

How do you handle cases where the judge is confidently wrong? Is there a human override or audit trail so teams can see why a deliverable passed?

Hi, I'm Ishaan, the one person behind Pennen.

I built it because every journal app I tried eventually made me feel like I was feeding it. Infinite documents I never finished. Streaks that shamed me after one missed day. And lately: AI "insights" reading my most private sentences back to me. I didn't want insights. I wanted a page.

So Pennen is built on three stubborn principles:
A page has a bottom. One dated page per day. You write it, you close it, you live your life.
Written is written. Yesterday seals and becomes read-only, a place you can visit, not edit.
An audience of one. No accounts, no Pennen servers, no AI. Your handwriting is never OCR'd into machine-readable text, your words stay ink. Pages live only on your iPad and in your own iCloud.

Craft bits for the curious: it's all native PencilKit with a custom stroke-merge that survives two iPads writing the same day; the emoji stickers peel off the sheet with a real GPU paper-fold (SceneKit shader) and press down with a haptic; and the tiny counter in the status bar shows how many strangers are writing right now, never who, never what. And one honest study, since "handwriting is better" gets thrown around loosely: a 2024 EEG study (Frontiers in Psychology) found handwriting produces far more widespread brain connectivity than typing. Modest, real, cited on our site.

If you've ever abandoned notebook number four in a drawer, I built this for you. Tell me about it and I'll tell you which of Pennen's decisions came from mine. I'm here all day.

this is a nice contrast to basically everything else on here today. "the streak forgives, a one-line night still counts" is such a small detail but it's the difference between a habit tool and a guilt machine. every other journaling app i've tried eventually adds some AI summary feature nobody asked for, so keeping that out on purpose is the actual selling point, not a missing feature. only question is what happens if you lose the iPad, is there any backup path beyond iCloud or is that a deliberate no as well

the past-pages-seal-as-read-only detail is a quietly brilliant UX choice, makes it feel like a real notebook rather than an endless editable surface

Love that past pages actually seal read-only, such a quiet way to make the journal feel real and uncheatable without any gamification pressure.

How does the page sealing work exactly — does it happen automatically at midnight based on your time zone, or do you have to manually close it out?

Love this. I've journaled on paper and I've journaled on the notes app. This sounds like a perfect medium.

This is a genuinely nice idea, no accounts and no AI reading my journal is a real selling point these days, not just marketing copy. One thing that'd worry me a little as a single point of failure: if my iPad dies or I switch to a new one and iCloud sync hiccups for whatever reason, is there any way to export or back up the pages outside of iCloud, or is iCloud the only copy that exists? Years of daily pages feels like something I'd want a belt-and-suspenders backup for.

The pen-on-paper feel with the Apple Pencil is genuinely convincing, and I love that past days just lock themselves away. The one-line forgiveness on streaks is a small but thoughtful touch.

the pencil feel is genuinely nice, ink weight feels just right under the apple pencil and the dated daily page keeps me from overthinking what to write

The "no feed, no AI" positioning is clear and honestly refreshing for a journaling app. I also like the constraint of one dated page per day; it makes the product feel closer to a real notebook than another notes database.

The sealing choice is brave. Since you deliberately avoided edit exceptions, do you think of follow-up thoughts as today's page referring back to yesterday, rather than corrections on the old page?

I do like the idea of a non-AI app that is clean and simple. Your aesthetic is calming and nice too.

You do claim that the one-time lifetime costs less than a Moleskine but that's not true. A new moleskin is $26: https://www.moleskine.com/en-us/shop/notebooks/the-original-notebook/classic-notebook-sapphire-blue-8051272893601.html and lifetime access for your app is $40 per the bottom of your app store page: https://apps.apple.com/us/app/pennen-handwritten-journal/id6781577517

the "past pages seal and become read-only" detail is what sells this for me. it makes the journal feel like real ink instead of an editable text box you'll fidget with forever. and "the streak forgives, a one-line night still counts" is the opposite of every guilt-trip habit app i've quit. feels made by someone who actually journals. congrats on the launch.

Any plans to bring this to iPhone?

the "tap any word on a youtube video or podcast" part is the bit that matters. most immersion apps make you leave the content to look something up, which kills the flow and the motivation right when you had it. keeping the lookup in place on real native material (novels, podcasts, video) is how people actually stick with a language instead of grinding flashcards. congrats on the launch.

As a Chinese speaker who also ships an iOS app, the "no accounts, no streaks, runs on-device" stance is the part I respect most — most language apps throw up a login wall and a streak counter before you can read a single sentence. Curious from the build side: bundling full JP + CN engines and offline dictionaries usually means a chunky download and some battery cost. Roughly how big is the app, and does the parsing stay snappy on older phones?

How does the on-device engine handle really obscure kanji or rare compound words that might not be in the bundled offline dictionaries — does it just leave you stuck on those?

the synced word-tappable transcript for podcasts and youtube is the part i'd want to stress test before trusting it. japanese in particular has a ton of homophones and casual speech drops particles constantly, so an on-device engine transcribing real conversational audio (not clean narration) seems like the hard part. does it show any confidence signal when it's guessing on a mumbled or fast line, or does it just silently give you its best guess as if it were certain

The on-device offline dictionary part is what stands out to me, most language tools want you online for lookups. On the Chinese side specifically, word segmentation is the hard part since there are no spaces to tell you where one word ends and the next begins, and it's easy to tap-split a compound wrong. Also curious whether it handles both simplified and traditional, since a lot of Chinese content someone might paste in (Taiwan sites, older text) is traditional even if the learner studied simplified.

how does it handle the furigana lookup for kanji compounds that have multiple readings depending on context, like 当て字 or rare names?

Hello makers & creators.

Pete from @findable. here. Claude recently made Sonnet 5 their default model. And as a marketer you want to know what it says about your business. So today we are launching Claude Sonnet 5 Brand Report.

Why we built this:

This week Anthropic made Sonnet 5 the default model in Claude: For every free and pro user.
New training data, knowledge cutoff Jan 2026. That means what Claude "knows" about your brand just got a big update. And most companies have zero visibility into it.

So we built a way to check => in seconds, for free.

How it works:

1. Type in your domain (try nintendo.com if you want to see a demo)

2. Click "Get brand report"

3. See exactly what Sonnet 5 says about you:

→ Knowledge score (0-100): does Claude actually know what you do?

→ Recommendation score: would Claude recommend you?

→ Confusion risk: is Claude mixing you up with someone else?

→ Plus your products, competitors, and buyer personas — as the model sees them

Important to know: this isn't "the truth" about your brand.
It's what's in the training material of the model millions of people now use by default. If the knowledge score is low, you're invisible when consumers or buyers ask Claude for solutions in your category. That's exactly why you want to see it.

It's completely free and takes a couple of seconds.

Would love to hear:

→ What did your report say? Any surprises?

→ What other models should we cover next?

We'll be in the comments all day. Thanks for checking it out!

Pete & Tosh at findable

ran it on my own brand and the gap between how i describe myself vs how the model describes me was honestly eye opening, the competitor breakdown was a nice touch too

Ran it on my side project and the competitor breakdown was surprisingly candid, picking up on positioning gaps I hadn't noticed. Wish the report went a bit deeper on sentiment though.

Ran my own brand through it and the competitor breakdown was surprisingly accurate, even picking up on positioning angles I hadnt considered before. Wish the report went a bit deeper on sentiment analysis but solid starting point.

Ran it on my side project and the competitor breakdown was actually sharper than what I get from expensive SEO tools, though it missed a newer launch I'd expect it to know about.

Hey Product Hunt 👋

VoicePad AI turns your voice into text, instantly, on any device — and it does it 100% offline.

What it does: You talk, it types. Real-time dictation that drops clean text wherever you need it — documents, emails, chat, notes, code comments, forms. The speech recognition (Whisper) runs locally on your own hardware, so there's no lag waiting on a server and nothing ever leaves your machine.

Where you use it:

Windows & Mac — dictate into any window. Write emails, reports, messages by voice instead of typing.

Android & iOS — same engine in your pocket.

VoicePad Direct (Android) — a full voice keyboard. Tap the mic, speak, and your words land straight into any app — WhatsApp, Gmail, notes, search bars — no copy-paste, no switching apps. Live on the Play Store.

Why it's different:

Fully offline. No internet, no account, no telemetry, nothing uploaded. Your voice stays on your device — the whole point for anyone handling private or client data.

One-time payment. Buy once, own it. No subscription.

All four platforms, built by one developer from scratch.

English + German, language always forced for accuracy (no auto-detect guessing).

→ Claim your spot at www.voicepad.tech
→ And tell me what's broken, what's missing, what would make this a daily tool for you. I read every reply.

This is my first Product Hunt launch. I'll be here all day.

Alex
voicepad.tech

VoicePad — We're alive, we grow, we create.

How does it handle background noise when you're actually out riding or in a busy shop environment?

The floating Orb is such a smart idea — flicked it on in VS Code and it just dumped clean text right where my cursor sat, no fuss. Love that it's fully offline too.

Offline Whisper on-device is exactly the tradeoff I want — cloud dictation means my IDE comments and DMs leave the machine, which is a non-starter for me. Which Whisper size ships by default, and is it quantized enough to stay usable on an older laptop with no GPU? And when the Orb drops text at the cursor, is that going through the OS accessibility API so it works in sandboxed apps, or synthetic keystrokes?

How does the AI actually come up with the names and locations for those popups, are they pulled from real customer data or generated on the fly?

the location-aware copywriting is genuinely clever, feels like it would actually drive conversions rather than just collect dust in a corner of the site.

How does the AI actually verify those purchases happened though, or is it just generating plausible-looking notifications?

The location swap actually fooled my coworker, he thought real customers were rolling in. Setup took maybe a minute.Tested with a VPN bouncing between regions, the names swapped cities instantly, kind of eerie in a cool way. Setup was genuinely under a minute.The location-based name swaps caught me off guard, my colleague in another country saw a totally different popup with the right city. Painless to set up.

Congrats on the launch @sanjux97

Does it integrate with payment providers to show real user names of the person who just signed up or is it randomly generated?

The generic "Peter from New York" problem is one of those things that becomes obvious the second someone names it, and stays invisible until they do. I've watched dozens of DTC brand websites and none of the social proof felt regionally believable, Mumbai visitors were getting Ohio names, London visitors getting Texas cities. Once you notice it, you can't unsee it.

The location-adaptation is the fix, but I'd push further on relevance. Beyond just location, there's a "person like me" signal that goes deeper than country match. A 24-year-old scrolling a hair serum ad feels closer to "Aisha, 26, just purchased" than "James, 52, just purchased", even if both are geographically local. Is there a path to demographic hints from the visitor context (session behavior, referrer, page), or is that a can-of-worms you're intentionally staying out of?

Also respect for naming the Stripe/Paddle rejection in a launch post. That's the kind of honest founder detail that makes a launch feel like a launch rather than a marketing exercise.

Hey Product Hunt 👋

I’m Kevin, founder of Codeswtch.

I spent seven years building infrastructure at Google and X, and over the past year I’ve watched something incredible happen.

AI has made it possible for founders to build products faster than ever.

With tools like Claude, Cursor, Lovable, and Bolt, a single founder can accomplish what used to require an entire engineering team.

That's a huge step forward.

But after the first customers arrive, the questions start to change.

Not...

"Can we build this feature?"

But...

• "Why is every new feature getting harder to ship?"

• "Why are AI agents making the codebase messier instead of better?"

• "Can a new engineer actually understand this repository?"

• "Will this architecture survive our next stage of growth?"

• "Are we building momentum—or accumulating technical debt?"

Those aren't coding problems.

They're engineering judgment problems.

That's why I built Codeswtch.

Our goal isn't to slow founders down with enterprise process or endless refactors.

It's the opposite.

We help AI-native companies preserve the speed that got them their first customers while building a foundation they can confidently grow on.

That's what Engineering as a Service means to us.

Not more engineers.

Senior engineering judgment, available when it matters most.

My belief is simple:

AI is making software creation cheaper than ever.

Engineering judgment is becoming more valuable than ever.

I'd love to hear from other founders and builders.

If you've built software with AI, what changed after your product started getting real users?

How does the subscription actually work in practice, do I get a dedicated engineer or a rotating pool, and what happens if my request sits in the queue longer than expected?

How does the subscription handle requests that need deep context over weeks, like onboarding to a messy existing codebase, or does it reset each time you submit something new?

Interesting idea, I have seen codebases messy or setup in a bad way if AI is fully in charge. And an experience software engineer reviewing does add real value. Though the value is diminishing as AI improves.

My honest opinion is the pricing is too high. Maybe I'm just not your target client but just sharing my thoughts.

Anyways, wish you the best of luck!

Subscribed last week and got a Slack response in under an hour about a stubborn auth bug. Loved not having to scope a whole project just to ask one question.

Quick question: how are the Google Ads and Meta tokens scoped and stored on your side? Proby moving bids and budgets means those tokens can spend real money across every account you hold. Congrats on the launch btw, the ROAS-vs-bank-account framing is painfully accurate.

how does the AI Agent Proby actually decide when to scale or pull back budgets, and can I set guardrails around it?

Finally gave this a spin on my Shopify store and the COGS + VAT tracking actually held up against my spreadsheets, which was a nice surprise. The Meta product label sync saved me a solid hour of cleanup.

How does it pull margin data from Shopify or WooCommerce, do I need to set up custom feeds or does it read product costs automatically?

How does Profit Bid handle the lag between ad spend and actual margin data coming back from platforms like Shopify — is the bid optimization working on real-time margin or are you using delayed conversion values?

Finally found something that looks at profit instead of just revenue, which has been a headache on our Shopify store. Setting up the COGS sync was straightforward and the Google Ads margin reports actually made sense.

The ROAS-vs-bank-account gap is one of those DTC realities that hits people hardest 6-12 months into their scaling. Everyone starts with ROAS as the north star because the platforms show it prominently, then eventually notices that their "successful" campaigns are burning margin they can't see until quarterly close.

The refund retraction detail is what makes this feel actually built by someone who's operated an ecommerce brand rather than just analyzed one. Most POAS tools I've seen treat conversions as final but in DTC beauty (my space), a "$47 order" that gets refunded to $0 two weeks later completely inverts the campaign math retroactively. Handling that at the platform layer instead of relying on the merchant to manually true-up is the honest architecture.

Genuine question, the A/C/X label sync to Google Ads. Do you handle cases where a product's true profit changes over time (new supplier prices COGS up, seasonal discount campaign runs, VAT bracket changes)? Because if the labels are set at first calculation and don't re-evaluate, the winners-scale/losers-exclude logic slowly drifts from reality.

Also: creative-side note. Bid optimization is one lever, creative rotation is another. Most brands optimize the first and neglect the second, running the same 3 ads until performance collapses. The teams that pair POAS-aware bidding with fresh creative cadence get compounding gains rather than diminishing ones.

Hey everyone!

I'm Val, designer and developer behind tinypad.

This simple product started with a frustration I'd had for years: I realized that I browse the web a bit differently from most people, and none of the major browsers gave me the simplicity and clarity I look for in every digital tool I use.

I've always considered myself as a digital minimalist (and minimalist in general). I close tabs when I'm done, keep a relatively small collection of bookmarks I actually use, and organize them just enough. Modern browsers, however, invest heavily in tabs, history, and search, while bookmarks often feel like second-class citizens.

So I designed and built tinypad: a fast and minimal bookmark manager designed to keep all the links on hand. In short, it is a combination of a bookmark system and web launchpad. On a single homepage, with a clean and snappy interface. Always in sync between browsers and devices.

After sharing early versions with friends and later with the Reddit community, I received a lot of encouraging feedback and surprisingly realized I wasn't the only one who thinks this way.

What's started as a simple side project, turned into my first Product Hunt launch after more than 10 years of building software, both in teams and on my own.

Thanks you for checking it out!

I'd love to answer questions, hear your thoughts, and collect any feedback or feature ideas.

Cheers,

V

How does syncing actually work across browsers, especially if I switch between Chrome on my laptop and Safari on my phone — is there a Tinypad account I need to manage or does it just run in the background?

never thought it's such a disaster that all bookmarks are so disorganized. it seems the developers just skipped a part related lo bookmarks.... tinypad is really a good thing! is it possible to make tinypad page a browser start page on IPad? thanks!

How does it actually pick which model to weight more for a given question, or is it just running all of them in parallel and averaging the outputs?

Ran my pricing question through it and got back something I actually wanted to copy, not a generic chatbot shrug. Curious how it weights the different models when they disagree.

Asked the same pricing question to it and got a sharper answer than the one I'd been workshopping for an hour. The blend of perspectives actually caught an angle I wouldn't have hit solo.

how does the brand profile actually keep things consistent across different formats, like does it learn from past designs or do I have to manually set the rules every time?

How does the brand profile system actually work across different content types, like does it pull colors and fonts from my logo automatically or do I need to set everything up manually?

Curious how it handles brand consistency when I upload a logo and color palette - does it actually pull from those assets across every new creative, or do I need to re-enter the brand info each time?

How does the brand profile actually keep designs on-brand across different content types? Does it pull from existing assets or do I need to upload everything manually each time?

Congratulations, Team Magnut AI!

A very timely and practical solution for businesses, agencies and creators who are tired of jumping between multiple tools just to get one campaign live.

What stands out most is the brand profile approach. In real marketing workflows, consistency is usually where a lot of time gets lost, especially when teams are creating posts, ads, visuals, captions and product mockups across different platforms. Bringing all of that into one place, while keeping the brand identity intact, makes Magnut AI genuinely useful.

The vision of building an AI-powered marketing operating system is exciting. This can save teams time, reduce coordination gaps, improve creative output and help businesses move from idea to execution much faster.

Wishing you and the team great success with the launch.

Looking forward to seeing Magnut AI evolve further!

How does the deterministic scripting actually hold up when my project's toolchain drifts from the defaults you baked in, say if I'm not on OrbStack or I'm already wired into a different secrets manager?

Curious how this handles updates when Claude Code itself evolves, since the skills rely on specific toolchain versions. Do you ship patches for new model behavior or is that on me to retest?

How well does it handle sites that dont have a clearly defined palette yet, like early-stage startups still figuring out their visual identity?

the DESIGN.md export idea is genuinely clever, especially how it maps the same brand tokens to whatever tool you're working in. wish more startups thought about handoff instead of just pretty screens.

The "export DESIGN.md" handoff is such a smart move, skipping the usual screenshot-paste dance when passing brand context to coding assistants. Going from palette to a clean Cursor-ready doc in one step is genuinely useful.

Pasted my site URL and got a clean DESIGN.md in seconds that loaded right into Cursor without fiddling. Honestly surprised how well it pulled the actual palette instead of guessing.

The per-photo red and green flag breakdown is genuinely clever, way more useful than a single number. Nice that you didn't gate the lead-photo ranking behind a sign-up.

The lead photo recommendation was spot on, picked the one I'd actually had the best luck with. Glad the photos stay private too.

Ran my Hinge through it and the lead photo pick was actually the one I always get the most compliments on, weirdly accurate. The bio rewrite suggestions were pretty solid too, way better than what I came up with.

The hotkey setup was surprisingly intuitive, and latency was basically nonexistent when I tested it in a Discord call. Definitely bookmarking this for streams.

Finally tried AudioPad and the hotkey setup was way smoother than I expected, playback through my mic in Discord sounds clean with barely any lag.

Love it. Global hotkeys and dedicated volume sliders are fantastic 👏
Trying it right now on my Fedora 😊 Does it work on MacOS too?

This made me laugh 😭

Honestly, I respect the decision to keep this simple instead of trying to turn a soundboard into a "creator platform" with subscriptions, accounts, and cloud features nobody asked for. The open-source + no tracking approach is also a nice touch. For something that's literally just supposed to play sounds through your mic, that's exactly what I'd expect.

Curious... what was the hardest part technically? Getting low latency, virtual microphone compatibility, or making it work consistently across Discord, Zoom, Teams, and games?