PH热榜 | 2026-08-07

Hi Product Hunt, Ohans here from @Coldtea.ai.

Coldtea brings your terminal, end-to-end testing and production monitoring into one agentic development environment.

• Terminal: Run the agents you already use, in parallel, with shared context between them. No lock-in. Your workflows stay yours.

• Visual QA agents: They drive your real app and catch regressions before your users do. And let you automate every release testing (supports iOS, Android and web)

• AI production monitoring: Watches production after the deploy and tells you what broke, in plain language. Connects your logging, observability tools, and agent traces in one place, so you wake up to investigated fixes, not blind alerts

• Tasks: Brings your engineering tasks closer to where you work. One place your team and your agents both work from, whether locally or in the cloud

Try it now: coldtea.ai

Would love your feedback; I'm in the comments all day.

Thanks!

Shipping faster was never our hardest problem.

When I was a Staff Engineer at the world's largest meal kit brand, HelloFresh, what I personally obsessed over was the part after merging a PR: how do you know a regression didn't slip through, how do you know what production is doing right now, have we architected the product right, and where does the whole team go to see it all of this in one place? That question followed me home.

Then agents arrived and made it louder. I tried every tool out there, and my day turned into context-switching between them or stringing together a custom workflow. Orchestration in one tab, QA in another, monitoring somewhere else.

It took me a while to see why none of them fixed it. Every other agentic development environment out there is pointed at the same thing: build. Run more agents in parallel, ship faster, orchestrate better.

All of that is genuinely good, and all of it is the first act. Software delivery was never just building, so the tools that only do build hand the rest back to you to figure out.

So we spent the last five months rethinking what a holistic agentic development environment for engineering teams could look like, alongside some fast-moving teams.

This is how Coldtea was born.

Coldtea is an agentic development environment for the entire software lifecycle. Not bolted on, not an afterthought. Day zero. Agents can already work across the entire stack. The next era of development environments has to do the same. That's our thought process.

Would love your feedback, especially if you're the one who gets paged when the thing your agent shipped breaks.

Love this direction. When most dev environments focus on building faster, @Coldtea.ai takes a different, calmer approach, focusing on building better. Agentic QA and monitoring, terminal and tasks, all in one place for less context-switching, and beautifully crafted by @ohansemmanuel and team.

@naviradjou could name @Coldtea.ai "the frugal ADE."

Try it now: coldtea.ai

Excited to finally launch Coldtea on Product Hunt!

Proud to be part of the team behind Coldtea and grateful for everyone who helped us get here. We'd love to hear your feedback, answer your questions, and learn how you're using AI in your development workflow.

Thanks for your support—it means a lot!

Try it now : https://www.coldtea.ai/

A PR can be green, and the app can still be broken. A deployment can go through, and users can still find the bug. Your logs can tell you something went wrong, and you can still spend an hour figuring out what.

We spent a lot of time on that gap but also on what comes before it. What should an agent work on? What context does it need? And when it says it’s done, did it actually work?

That’s what pushed us to build Coldtea around the whole loop. The context carries through, from deciding what to build to an agent using the real app to check it to figuring out what happens when something breaks in production.

How much of that should an agent actually own? Where would you draw the line?

We build something internally like this, but this is next level greatness for any dev-team to use. Solving absolutely a problem shipping faster with AI. Nice integration with LangWatch ;)

Congrats on the launch @ohansemmanuel and team! Lets goo!

Congrats on the launch.. My big build issues lie upstream of the QA agent.

Where does the test intent come from? If the coding agent and the QA agent both work off the same idea of what the feature should do, the check is a mirror.

And can Coldtea catch a test that always passes? Mine sat green for a week while a bunch of stuff was breaking, because it was only checking that a metadata column was present. That column could never be null, so every run came back legit.

A check that has never failed once is the thing I want flagged.

The promise sound great , but "agent speed without breaking anything" is a bold claim. I'd be interested in seeing how it perform on large , real-world codebases.

We love coldtea! We use it to catch regressions and monitor our agents so our small and mighty team can move fast. It wouldn't be possible for us to move as fast as we do at @Brew without coldtea. 10/10!

Congrats on the launch @ohansemmanuel @favour_ohans 🎉 How does the visual QA agent handle apps with a ton of dynamic content, like feeds that change every load? Does it know what's a "real" regression vs just normal content shuffling?

Does everything work as one integrated system, or can the different parts be used independently? We're currently developing with Claude Code and Claude Design, but we don't have any really good tools for QA. Could we use your service specifically for QA, independently of the other parts?

What about downtine? How is that taken care of?

Congrats @fmerian for hunting another gem and Congrats #makers for the successful launch.

What happens when two agents make conflicting changes to the same part of the code?

Huge congratulations, Ohans and team! 🙌 For iOS/Android testing, do I need to already have an existing test suite, or can Coldtea generate the visual QA flows from scratch?

Question: What are the 3 hardest things in software delivery that you managed to solve with this platform? thanks 🙏

This is a very interesting design. How do the visual QA agents hold context of the actual intent?

Hey Product Hunt,

I’m Wenhao Yu, founder of Soloop.

Before Soloop, I built products at a large tech company. Our team could solve hard product problems, but we spent too much time waiting for decisions to move through meetings, quarterly reviews, reports, and handoffs.

AI coding tools later made it possible for one person to ship a working product. I could build the first version alone, then faced the work code could not cover: choosing an audience, finding users, reading their feedback, and deciding what to build next.

I built Soloop to give a solo founder a coordinated AI team.

You give the AI CEO a company goal. It makes the plan and coordinates the other agents. The AI CTO builds and iterates from user signals. The AI CMO finds potential users, runs distribution, and brings their responses into the next decision.

While building Soloop, we stopped treating each agent as a standalone assistant. Separate assistants left the founder carrying context between chats and reconciling their plans. We organized the team around one company goal and a daily review. Soloop brings decisions that need judgment back to the founder for approval.

Soloop is early, and we want solo founders to shape it with us.

• Where does the path between a working demo and a real business break down for you?

• Which part would you trust an AI team to run?

Hey Product Hunt,

I’m Qingfeng Meng, CFO of Soloop.

I first knew this company as an investor. At Shuimu Tsinghua Alumni Seed Fund, I worked on investments in more than 30 Tsinghua-founded startups. I later worked on global GTM at Zhipu AI. After the fund invested, I crossed the table and joined the team.

Early-stage investing taught me to treat every important decision as an allocation of capital. That capital includes cash, a founder’s time, and the paths given up by choosing one direction over another. Building, selling, and waiting all have a cost.

I measure an AI team by how much uncertainty it removes each week. Soloop helps a founder turn a company thesis into a focused bet, test it in the market, and bring back evidence before making the next commitment. Most AI tools make an individual task cheaper. Soloop aims to reduce the cost of deciding what the company should do next.

That is the vision I joined: a founder with conviction should be able to build a real company without first assembling a large organization.

My job is to help make the economics of that model work. I’d be glad to hear your stories and explore collaborations with founders, operators, and investors building or backing one-person companies.

Hey Product Hunt,

I’m Xinyu Shi, CTO of Soloop.

Building Soloop taught me that creating a useful AI team is very different from putting several agents into the same product.

The hard part is not getting an agent to generate code, write a marketing post, or produce a plan. The hard part is making sure every agent understands the same company goal, works from the same context, and turns its output into something the next agent can actually use.

As CTO, I’ve focused on building the system behind that coordination.

When the AI CMO discovers a user need, the insight should not disappear inside a chat history. It should become structured context for the AI CEO to evaluate and, once approved, a clear task for the AI CTO to implement.

That requires more than prompts. It requires shared memory, clear ownership, reliable handoffs, execution environments, approval checkpoints, and feedback loops that help the system improve without taking control away from the founder.

We are still early, and many of the most interesting technical problems are ahead of us:

• how much autonomy agents should have

• how they should recover from failure

• and how founders can understand why a decision was made.

I’d love to hear from other builders:

• What would make you trust an AI agent with a real part of your company?

• And where do today’s coding agents or multi-agent systems fail you most often?

I designed almost exactly this for my own company — departments instead of loose agents, each with a manager, shared memory, daily reviews. The one place I landed differently is approval. I started approval-first and moved to "execute and report, stop only at what's irreversible", because the approval queue became the bottleneck it was meant to remove. As a solo founder my scarcest resource is attention, and a system that asks me twelve times a day is a slower version of doing it myself. Curious how you're drawing that line — is "requires approval" a fixed category, or does it widen as the founder builds trust?

Love the focus on keeping the founders in control. Ai should amplify decision making , not replace the vision and judgment that make a company unique.

Congrats on the launch team clearly having real conviction here.
For someone with zero technical background, how much do they actually need to understand to give useful approvals to the AI CTO?

I love the concept, but what do you mean by "approval-first"?

Hi Product Hunt,

I’m David, co-founder of Soloop, leading market, brand, and commercialization.

Before Soloop, I worked across media, platform marketing, and creator operations, then co-founded a Gen Z social startup in MiraclePlus’s 2020S program. That experience led me to a conclusion: at the start, attention is all you need. Once you make the right people care, you can figure out the rest with them.

Vibe coding gave one person a software team. Soloop is building the rest of the company.

Soloop brings an AI CEO, CTO, and CMO into one company-building system for a founding team of one. They share the same company memory, coordinate product and GTM execution, bring key choices to the founder, and carry each decision into action. You keep the vision and final say. Soloop carries more of the operating load as the company moves toward users and revenue.

My job is to keep that ambition close to what users actually do. We’d love to hear your founder stories, compare notes, and explore collaborations with people building the next generation of one-person companies.

Congratulations

Ambitious idea! Best luck on "really make things work"!

As a solo founder who's already shipped, the build was never my wall — it's the distribution grind (posting daily, finding the right communities, actually reading feedback) that eats the week and doesn't get cheaper the more I code. That's the part I'd hand off first. Saw in the thread you're bringing existing products into the CMO flow soon — that's the version I'd genuinely try.

Hi team, was trying to take the product out for a spin - just have a look when you get some time. I am on chrome. Love the concept btw, and hence was trying to check this out

Approval-first is the hill I’ve been dying on all year, so congrats, you built my argument.

I’m a solo founder running 25 AI marketing agents, and the two rules that keep that sane are exactly your premise: agents produce drafts, not actions, and when they’re missing information they ask instead of inventing. (Mine once invented an employee named Teri and assigned her support tickets. Approval-first is why Teri lasted one document review instead of reaching a customer.)

So here’s the question I’d ask before subscribing, because it’s where I’ve watched this model strain: how do you fight approval fatigue? At some volume, “approve” stops being a decision and becomes a reflex, and a human rubber-stamping fifty prompts a day is arguably worse than no gate at all, because now there’s an audit trail saying someone reviewed it. Do you do anything structural to keep approvals meaningful: batching, risk-tiering, surfacing only diffs, anything?

Genuinely asking. This is the design problem I think decides whether approval-first scales or becomes approval theater.

Seems like there is a problem...

Hey, everyone from Soloop, do you use agents to seed your posts here with your tool, or do you actually posted it yourself (I don't assume you put time to write it :))

Congrats on the launch.. Is there a level of autonomy planned to baked in? Also is the product only meant for solo folks - or in future can also accomodate co founders. Last connected issue is that our calendars, emails and usually slack - are filled with discussions and clutter - could those be connected and baked into the plan. Last but not the least -- can real product signals from say feedback widgets, posthog, newrelic/sentry, langfuse traces etc automatically be parsed to create and suggest product stories (sorry having been a PM for so many years - this is something I keep thinking about, product discovery).

Have signed up will try to take this out for a spin soon :-)

Soloop is a bold step toward empowering solo founders. Being a solo entrepreneur means juggling countless tasks, and having agentic AI roles — CEO, CTO, CMO — that handle planning, building, and selling can significantly improve efficiency. The balance of keeping judgment and ownership while delegating unfamiliar work is exactly what makes this system exciting. Look forward to seeing it envolve.

I tried Soloop today and I really like the idea of having an AI team (CEO/CTO/CMO) helping solo founders move faster.

One thing I’m still trying to understand is the workflow after choosing an AI role. I selected a CEO agent and started discussing my idea, but I wasn’t completely sure what happens next.

Does the agent mainly help founders refine ideas, validate opportunities, and make decisions? Or does it also work together with other agents (like CTO/CMO) to actually execute and build the product?

I think clarifying this journey would help new users understand the value faster. Excited to see how this evolves!

Hi Product Hunt!
I'm Diana, product manager for Nitro. I've been in this role for eight years, so I've watched Nitro grow up.
When we started, Nitro was a simple platform for one thing: getting short texts translated by professional human translators, fast. Over the years we built it into something much more powerful — adding context so translators get it right (screenshots, tone of voice, character limits) and expanding what you can send.
Each of our past Product Hunt launches marked one of those leaps:

• First launch: the core service, API, and export to Google Sheets

• Second launch: file support — Google Docs translation with formatting preserved

• Third launch: developer and content formats like JSON, HTML, iOS Strings, XML, and Google Sheets

Despite being a niche product, Nitro got noticed here. A #4 Product of the Day and two #5s for a product this niche still surprises us, and it's thanks to this community. Grateful and a little humbled every time — that's why we're back today to show you what's next.
What Alconost has always strived for is the same: pair genuinely good human translation with the most powerful tech we can build around it. This launch is the latest step. Nitro now supports MPP — the machine payments protocol — so an AI agent can order a human translation and pay for it on its own, with no account and no API keys. You hand your agent the task, it sends the text and pays per request, and a professional native speaker delivers the copy.
We think it'll be interesting for folks building with agents — Nitro is the first translation service to offer this — and we hope the localization crowd finds it just as interesting. We'd love your thoughts and questions!

Publication-ready in hours is a bold claim. I'd love to see actual turnaround data from real orders before I trust it for anything client-facing.

As someone who's dealt with clunky localization vendors, skipping the account and API key setup entirely is honestly refreshing. I'd try this for a small app update just to see the turnaround time.

If there's no account and the agent placed the order — where do the finished translations actually come back to?

80+ languages with no minimums is the part that stands out to me. I've been priced out of smaller translation jobs before because agencies wanted a monthly contract.

Do you see this being used mostly for app localization, or are email campaigns and marketing copy becoming the bigger use case?

It's impressive to see how Nitro has evolved over multiple launches while staying focused on quality. Wishing the team another successful launch! 🎉

I wasn't expecting machine-to-machine ordering with human delivery to be the next step in localization. It'll be interesting to see what kinds of AI agents people end up connecting to Nitro first.

great idea, i’m seeing many successful products that make interactions more convenient for ai agents useful. congrats on the launch!

Congrats on the launch! qq When an agent orders with no account, how does the translation get back to it?

Congrats on a new launch, guys! I am wondering something: how do you ensure quality if an agent is placing the order?

Hi everyone!

I'm Alex — I run Alconost, and Nitro has been my baby for years. I came up with it out of a simple frustration: for short texts — an app update, an ad, a few lines of email — professional translation was far too heavy. You shouldn't need a project manager and a quote to translate three sentences. Nitro was my answer to that.

I've since moved on to other projects, but I still keep an eye on it and bring ideas to the team — and seeing it come back to Product Hunt a fourth time, now with AI agents able to hire human translators directly, feels kind of surreal. Huge thanks to the Product Hunt crowd for all the support over the years.

Hey Product Hunt people! 👋

This launch makes me a little nostalgic - I ran the first and second Nitro launches here, and spent a long while in customer support and customer development before that. 

What I love is that after all these years Nitro keeps catching up with what's new; it doesn't sit still. This launch is a great example: the moment AI agents became a real thing, we found a way for them to bring human translators in rather than replace them. It's been a joy to watch it evolve. We'll be happy to answer any questions about Nitro!

Congrats on the launch team!!

Aren’t you worried that AI could eventually make your service obsolete? Translation quality is already very good with AI, at least for the most widely spoken languages.

How do I connect my agent to Nitro? how does it order a translation?

Rindler is a map of a website that an agent calls instead of reading the page. The map knows what screens the site has and what you can do on each one, and it knows what comes back when you do it. So your agent gets the same fields back every run, in the same shape. We've mapped more than 1000 sites so far.

How a run works

We build the map ourselves, once, before anyone runs against it.

1. Our ingest picks up the site and works out what it's looking at.

2. The mapper drives the site click by click and drafts the map.

3. A fleet of agents goes through and fixes what the draft got wrong, and the map has to pass a check against the live site before it serves anyone.

Every run after that just draws on the map. There's no model sitting there working the page out again, so the run is fast and you get the same fields back in the same shape every time. Most browser agents redo all of that work on every single visit. That's why the field names come back different each run, and it's why every run takes as long and costs as much as the first one.

Inside a run we pick the transport step by step. For one action we drive a real browser, and for the next we call the site's own endpoint. If you bet everything on those endpoints you end up thin at the hard flows. Drive a browser for every click and it's slow.

What happens when a site changes

Every map runs against the real site once a day, checking a list of things that site still has to be able to do. Most of the time that means we catch a break before it shows up in your run. Fixing it is on us, so it doesn't end up as a ticket in your backlog. And if a step does fail mid run, a model works that one step out again while the run is going, and the request finishes.

What it costs

Starter is $100 a month for 100 sessions, with a 7 day free trial and one site you pick that we map for you. Teams is $1,000 for 1,000 sessions and scheduled automations. A session costs a dollar at every tier, and you don't pay for a run that fails.

https://chat.rindler.ai/ is a playground with a few sites open so you can poke at one without signing up. If the site your team is stuck on isn't in there, name it and tell me what you need done. I'll be here all day!

Michael

A clean approach to browser automation always gets my attention. Looking forward to seeing how developers use this in real workflows.

Automation that feels reliable is hard to build. Excited to see what the team has created here. Best wishes for the launch!

This made me stop scrolling. Instead of building another AI app, you're improving the way AI interacts with the web itself. That's a much bigger opportunity than it might seem at first glance. Looking forward to seeing real-world use cases.

Congrats on the launch! Love the idea of turning websites into reusable maps instead of making agents rediscover everything on every run. Curious how Rindler handles frequent UI changes or A/B tests—does the map update automatically? Wishing you an amazing launch day!

The login and bot protection part caught my attention. How do you handle sites that frequently change their security flow?

Congrats @michael_serrano & team!

Hi Product Hunt! 👋 Nikhil here, co-founder of BrowserOS. Fun fact: I’m building this product with my twin brother Nithin (@nithin_sonti)

Today we're launching BrowserOS neo: a browser built NOT for you, but for YOUR AI.

Here's the problem we ourselves kept hitting: Claude Code and Codex got scary good at reasoning about work, but the moment a task touched the web, like checking a dashboard or filling out a form or booking your dentist appointment, they hit a wall.

We tried several solutions out there today and none of them were a good fit. Cloud browsers (like browser-use) run in a datacenter, so logging into your accounts is a pain, and sites like Twitter and LinkedIn block you because you are on a datacenter IP. And headless browsers (like agent-browser) don't let you see what your agent is doing. 

We realized what’s needed is a browser that is run locally, has your logins but can be easily driven by YOUR agents.

That’s what we built with BrowserOS neo! You import your logins from Chrome in one click and connect your agent (Claude Code, Cowork, Codex, OpenClaw, etc). And then you can start giving your agent any complex task and see it get magically done using neo!

To be clear, BrowserOS neo is NOT a Chrome replacement. It's a secondary browser that sits next to Chrome, and we’ve made it agent friendly. 

The features you get today are:

• 🔓 Free and open source. It’s available on macOS and Windows.

• 👀 Watch your agent live. There's a cockpit view on your new-tab page that shows every agent, every site, every step.

• 🪙 Less tokens. BrowserOS neo hands your agent simplified snapshots of the web page instead of screenshots, so you burn way less tokens than using Claude’s Chrome extension and other solutions.

• ⏪ Rewind anything. Every agent session gets saved as a scrubbable video, so you can find the mistake and coach the agent.

• 🗂️ Your tabs are sacred. Agents work in their own tabs and can't touch the doc you're writing :)

• 🤖 Parallel agents. You can run as many agents as you want at once, each in its own tabs.

• 🔒 Privacy first. Everything stays on your machine, and every action is logged to a local file on your disk.

We'd genuinely love your feedback more than anything else today: what's the first task you'd hand an agent? Reply and I'll tell you honestly whether BrowserOS neo can do it today.

We also open source and you can find our full source code here: github.com/browseros-ai/BrowserOS 

We're in the comments all day.

PS: We are big fans of OpenClaw, we initially named the product as BrowserClaw but had to rename it to BrowserOS neo because of SEO issues.

Love using Neo, this really has been a passion project from the whole team!

Browser as a tool for your AI is one of those things where you can’t believe you lived without it before. This is literally a game changer my coding workflows where I need to drive a browser reliably.

Any plans for Linux app?

I am more than satisfied with the functionality of BrowserOS neo (I have been using it since it was BrowserClaw). The primary pain point it has resolved for my workflow has been the ability to log in to platforms when providing browser access to desktop/CLI agentic tools. It has been one of the best tools for AI that I have used interms of giving me that feeling of actually colaborating. There are only a handful of AI tools that I use on a daily basis as a non-developer/non-technical person, and BrowserOS and BrowserOS neo are definitely among them. Wonderful job!

Hey Product Hunt 👋

I’m Lyubo, the Product Manager behind Progress AI Observability.

We know a lot of teams are building AI agents now. The first demo comes together quickly but then you try to run it in production and things get more complicated.

Traditional monitoring can tell you that an app is running, but it usually can't explain why an agent chose a particular tool, ignored useful context, entered an expensive loop or produced an answer that looked convincing but was wrong.

We built Progress AI Observability to give engineering teams that missing visibility.

We want to move teams from “something went wrong” to understanding why it happened, what needs to change and whether the next version is actually better.

You can start tracing in minutes with support for .NET, Python, and JavaScript/TypeScript. There’s a free plan, and no credit card is required.

We'd love to hear your feedback on:

• What’s hardest to debug once an agent reaches production?

• Which signals are most useful to you: traces, evaluations, latency, token usage, or cost?

• What would you need to see before using this with a production agent?

Thanks for checking it out 🙏
- Lyubo and the Progress AI Observability Team

@yochev Congratulations. And happy product launch.

I assume if I want to monitor how Claude Code works, then I have to use Fiddler? (I love Fiddler)

the Allow/Deny buttons on the creature are the sneaky-useful bit, I lose so much time hunting for which window wants something from me. How does it behave when it needs approval: does it freeze my window and wait or just sit there until I notice?

@ravindra_sisodia1 congrats, this is lovely!

hehhe loved it! All the best here Ravindra

Headless browsers, built for agents. Indeed, we are moving to the post-human internet.

The stateless-per-call design is a real bet. Curious how you handle workflows that need persistent login state across steps, does the agent have to re-authenticate on every call, or is there a session token it can pass in? That tradeoff seems like the main friction point for multi-step agent tasks.

Detection is based on deterministic rules or some local small ML-model? Should I setup a rules for my projects before sharing or DataBlur will just do the work?

Hey ProductHunt 👋!

We’re 5 founders who’ve collectively done over 250 interviews - everywhere from startups to FAANG+ to quant shops. Most of the interviews we’ve done were Leetcode based or tested skills that weren’t used on the job.

Meanwhile, at each of our companies, though, PR counts have nearly tripled. Most of us haven’t manually edited a line of code in a year. Our teams are putting more and more emphasis on code and architecture reviews, yet hiring processes haven’t changed whatsoever.

Even the AI-assisted ones we’ve done still assess code output as the primary evaluation metric. As agents develop, we truly believe this will not be the most challenging ability for an engineer to have.

We’ve seen that the real difficulty with using AI is not just reviewing code your AI generates; it's reviewing code that another engineer’s AI has generated, having little context yourself.

That’s why we’re launching Merge today: to help hiring teams assess engineering judgement. Here’s how it works:

1. Candidates are shown a small codebase to understand and a PR to review and comment on.

2. An AI agent addresses each PR comment via a code change or reply, simulating a real engineer.

3. Candidates can repeat until 5 revisions are used up or time runs out.

Throughout this process, we assess the following:

1. Coverage - How many bugs or vulnerabilities did the candidate identify and address?

2. Communication - Was the candidate efficient and constructive with their feedback?

3. Efficiency - How many revisions and tokens did the review take?

We’re the first platform that can show you exactly how efficient a candidate is with token use, LLM costs, and PR revisions — all of which are becoming exceedingly important in industry positions.

If you’re interested in the next-generation of engineering hiring, book a demo with us! Feel free to ask any questions below as well.

how do you stop a candidate from just running the take home assessment through an AI code review tool of their own before submitting it back to you? feels like an arms race where both sides eventually just have AI grading AI

Hey PH!

I’m Antoine, and over the past year, as I tried to scale our agentic coding workflows and software factories at my company, I kept hitting the same set of problems. So I built HAR to solve them, and I'm open sourcing it today.

Getting a single coding agent to work in a repo is easy. Scaling to a real multi-agent workflow, where several runs at once and you still trust the output, is where it breaks down. A few things go wrong at the same time:

1. No standard way to run or verify a repo. That knowledge is scattered across a README, a CLAUDE.md, editor rules, and CI config, all drifting out of sync with each other and the actual code.

2. Agents on one repo collide. Shared dev server, shared database, shared ports, conflicting git state.

3. Trusting a change means re-verifying it yourself. Which defeats the point of running a fleet.

4. Vendor sandboxes lock you in. If the setup lives in someone's hosted dashboard, switching agents later means rebuilding the whole thing.

What HAR does

HAR is a CLI and an MCP server. It works with Claude Code, Cursor, Codex, or any MCP agent, and it closes each of those gaps:

1. Isolation. Each agent gets its own git worktree, branch, ports, and database. Nothing is shared with the main checkout or another agent's slot, so a fleet runs in parallel without colliding on a dev server, DB, or ports.

2. Deterministic validation gates. HAR runs your project's real checks through a fixed pipeline, same result every time. The result is bound to the exact code that passed and enforced at commit time, so an unverified tree cannot land.

3. Verifiable proof. Every run leaves logs, artifacts, and a validated tree hash tied to the exact code checked. A reviewer inspects the evidence instead of trusting the agent's self-report.

4. Full observability. Mission Control is a local dashboard showing every repo, worktree, run, and validation in one place, so you can watch a whole fleet as it works.

All of this lives in one contract committed to your repo, which every agent reads the same way. It replaces the usual scatter of a README, a CLAUDE.md, editor rules, and CI config that drift apart. You start from a profile that matches your stack, your agent adapts it to the real repo, and you extend verification with plugins (like Playwright) or with any command you already run.

Give it a try and let me know what you think :)

I'd genuinely love feedback, especially from anyone already trying to run multiple agents in parallel.

the single in-repo contract replacing scattered CLAUDE.md / README / editor rules / CI config is a genuinely sharp design call. config drift across those files is what makes multi-agent setups brittle, every agent reading a slightly stale version of the truth. curious how opinionated the contract schema is on init: does `har init` scaffold a full contract template, or does it leave most of the structure to the team to define?

Super interesting concept, not an obvious solution to that problem. Will need to give it a bash!

Hey Product Hunt, Rajesh here, founder of Troopr.

I've spent years watching engineering leads act as human middleware: chasing updates in Slack, reconciling Jira against GitHub, running a standup that exists mostly to find out what's going on.

And the problem is getting worse, not better. AI agents are joining teams as contributors and multiplying output, but every contributor adds coordination cost. The bottleneck is shifting from producing work to knowing what's happening.

We've been building standup tools longer than we'd like to admit, and for years we watched the same loop: the bot pings at 9am, everyone alt-tabs away, and at 4pm someone types "same as yesterday." Eventually we admitted the uncomfortable part. The bot was never the problem. The form was. A standup bot is just a meeting that follows you into Slack.

So we built the AI Scrum Master to know the real state of the team.

It doesn't ask. It reads what your team actually did: the PR that merged last night, the ticket that hasn't moved in four days even though its owner said it was done, the thread where someone mentioned they're blocked on staging access. From that, it writes each person's standup. You read it, fix a line if you want, done.

If your team kept the live standup, Troopr joins your Google Meet, listens, and delivers a report cross-checked against your live Jira and GitHub state. Every notetaker we tried gave us a context-free summary of one call. Troopr's report knows what the team said last week and what the tools show right now, so it catches what doesn't add up.

The part I'm proudest of is the memory. Troopr learns how your specific team works, who owns what, what "done" means for you, and every standup gets more accurate because of it. Everything it retains is inspectable: you can see, correct, and delete exactly what it knows about your team. No training on your data, no raw message storage.

We're not new at this. Troopr has served 600+ engineering teams including Netflix and Snowflake, and today it becomes AI-native.

It's free for 10 seats as a launch offer, everything included, so you can point it at tomorrow's standup and judge the report yourself. I'll be here all day answering everything.

Congratulations on the launch! Having an AI that actually understands team patterns and flags discrepancies in updates is a huge step up from standard standup bots.

Siva here 👋 I'm on the engineering team at Troopr, and the best part of building this has been using it with my own team every day.

We point it at our daily standup, and every morning it writes each person's update from what they actually did — the PR opened late last night, the ticket that moved, the discussion where someone got unblocked. Most days I read my update and think, "Yep, that's exactly what I did." Nobody has to type a status update anymore.

My favorite part is the validation. If a ticket is marked Done but the PR is still open, it catches it. If someone reports the same work for days without any activity in Jira or GitHub, it flags that too. And when there's a real blocker, it doesn't just say "Blocked" — it includes the context and possible next steps. It has no problem calling me out either 😅

For teams that still have live standups, it handles both sides:

• A pre-read arrives before the meeting, so nobody spends 10 minutes reading updates.

• During the call, it listens, captures decisions, action items, and owners, then publishes a recap that reflects everything that changed during the discussion.

One thing that surprised me is how much it improves over time. It has learned our projects, who owns which services, and even what "blocked" means for our team—without us configuring any of that. The reports today are noticeably better than they were in week one.

We've completely stopped preparing for standups. We just show up, skim the report, and spend the meeting discussing the two or three things that actually need attention.

If you run an engineering team, try it with just one team for a week. That's the fastest way to understand the value.

Happy to answer any questions about how it works under the hood! 🚀

Really like the shift from collecting stand ups to actually understanding what's happening across the team. Excited to see where this goes.

Been following Troopr for a while, mostly for the Slack and Jira side of it. Good to see the AI layer land on top of something that already worked rather than showing up as a brand new tool with no track record behind it. Congrats to the team.

Mani here from the Troopr GTM team.

Every conversation I've had with an eng lead this year lands in the same place: output is up, visibility is down. Agents ship more, and nobody has a reliable read on what's actually happening.

That's the gap we built for, and it's why the memory matters more than the meeting.

Curious whether that matches what people here are seeing, or whether it's just the teams that end up talking to us.

I like that that the retained memory is inspectable instead of being a black box. Congrats team!

Soumya here from the Troopr Product team 👋

The hardest part of building this honestly wasn't the standups. It was getting the AI to remember the right things without drowning everyone in noise.

A team says a hundred things a day, and most of it is clutter. What took the longest was teaching it to hold onto the few that actually matter - the blocker from Tuesday, the call you made last sprint, that risk someone flagged once and everyone forgot about and resurface them right when they matter again.

That's what we wanted Troopr to do : remember what the team forgot, and give everyone a clear, honest picture of where things really stand so the whole team stays in sync as the work moves

If you run a distributed or async team, I'd really love to know if this resonates.

Hi Product Hunt! 👋 I'm Garvit Bandil, Product Manager at Troopr.

Here's the problem we kept running into: the true status of an engineering team is scattered across three places that rarely tell the same story. Jira says one thing, GitHub says another, and the real context is buried in Slack. Someone ends up piecing it all together manually, usually during a meeting that's mostly about surfacing work that already happened.

That's why we built Troopr.

Troopr reads your team's real activity across Jira, GitHub, and Slack, then handles the coordination work for you. It drafts each person's standup including relevant code diffs, so meaningful work doesn't get lost behind vague commit messages. It flags stalled reviews, unresolved blockers, and low-signal updates, and even checks in with quieter teammates before the report is posted.

Most importantly, every insight links back to its source: a Jira ticket, pull request, commit, or Slack thread-so you can verify everything yourself.

Few more things

• Humans stay in control. Troopr asks for approval in a DM before making any changes to Jira.

• Free forever for up to 3 seats. No credit card required.

I'll be here in the comments all day, and I'd genuinely love your feedback.

Where do you think Troopr would break on your team's workflow? What would you want it to handle better?

Hey Product Hunt 👋

I kept writing the same kind of document: click here, then here, screenshot, crop, arrow, number, repeat. An hour of work to describe four minutes of clicking. Every tool that promised to fix it was a browser extension — useless the moment the workflow touched System Settings, a native app, or a terminal.

So StepShot is a native Mac app. It watches clicks and keystrokes through the accessibility layer, grabs the screen before each click lands (the frame you want is the one showing the button you're about to press, not the menu that opened), and assembles a document while you work. Two capture modes: action-driven for tight step lists, screen-recording when you want the full video kept as evidence. Three targets: active window, full screen, or a fixed region.

Four decisions I'd defend:

The preview is the export. The editor canvas, the preview pane, and all three exporters run the same renderer. A tool whose entire output is images cannot let you approve one image and ship another.

Local-first, and specific about it. Projects and screenshots live in a folder on your Mac. There is no account and no telemetry. Passwords are never read at all — secure-input fields are excluded at the engine level, so the characters never reach memory. The AI is deterministic and local by default; if you want a model to polish the wording, you bring your own OpenAI-compatible endpoint and key, and it only refines a draft that's already complete. Sending screenshots to a model is a separate, off-by-default consent, and samples are redacted locally before they leave.

Redaction is destructive. Sensitive regions are resampled down and back up, burned into the bitmap. Not a blur, not an overlay someone can strip. Local Vision OCR proposes them; you confirm each one before export unblocks.

One-time purchase. $129 Pro, $599 Team (5 seats), perpetual, all future versions included. The free tier isn't a trial: capture and editing are unlimited and never expire, exports cover the first 10 steps in Markdown and HTML.

macOS 14+, Universal (Apple Silicon + Intel), signed and notarized. UI in English, 简体中文, and 日本語.

If you'd rather try the paid tier than read about it: drop your email in a comment and I'll send you a 30-day key with everything Pro unlocked — unlimited steps, PDF, branding. No card, no account, one per address. When it ends the app goes back to the free tier and nothing you recorded is touched.

Honest about what's not there: no Notion export, no auto-translation, no cloud sync or team accounts, and Windows is a private test build — not a release. I'll be here all day for questions, including the annoying ones.

Hey, looks interesting. May I email to get the trial please?

Hi

My name is Rishi, I’m the one of the creators of Blueberry.

Our journey began when I very publicly got roasted for being a horrible texter.

In my defense, I didn’t do it on purpose. When I “accidentally ghost”, it usually goes something like this:

1. I see a text while I’m busy
2. I think, “I’ll respond to this later”
3. I completely forget about it

Once I’ve forgotten about it, it gets buried by the endless spam, group messages, and general noise that clutters my iMessage inbox, never to be found again.

This has had real consequences for me. Genuinely, I’ve lost work opportunities, fallen out of touch with previously close friends, and let potential dates fizzle out simply because I forgot to reply.

So we made Blueberry, it's a mac menu bar app that finds the messages that need a reply and auto-drafts a response in your voice.

You can view and send responses in one click, you can edit the response, or choose an alternative variation. And, it works with individual or group messages.

I want to be very clear, we don’t want to replace your iMessage app or automate your texting. Our goal is to remove the small bit of friction that turns “I’ll answer this later” into three months of accidental ghosting.

We also added some fun features like an unhinged mode, which is deranged and hilarious, and a flirt mode, for those of us who struggle to send that perfect text ;)

I want to also address privacy and security. Blueberry needs Full Disk Access to read your iMessages and Automation permission to send replies you approve.

Recent thread context is sent to the cloud to generate drafts. Our models are hosted on AWS Bedrock or Google Vertex servers. This means that your iMessages stay private and secure, and model providers cannot see your data or use it to train on.

Blueberry will never send messages automatically - you will always be the person to hit send, and we have a 5 second unsend just in case you change your mind!

It's great for people who:

- Live in Do Not Disturb mode,
- Read texts but don’t reply
- Get overwhelmed by group chats
- Constantly say, “Sorry, I just saw this.”

We'd love your feedback on Blueberry, what is it missing? What can we do to make this something you use every day?

This is our first launch on product hunt, and we're excited to share all of the other products we're building.

I think that this is somehow related to me. My Facebook messages are unanswered from 2021 :D

The "nothing goes out until you click send" choice is the part I like most. I'm building a voice AI at Callie Care that calls aging parents every day, so I spend a lot of time on the same problem from the other side: people love someone, but the activation energy to reach out is just high enough that weeks slip by. Curious how you capture "your voice" per recipient early on. Does the tone shift based on relationship closeness (a parent vs a coworker), and roughly how many corrections does it take before the drafts actually feel like you?

Hi,

Use this code for 30 days of free pro access, exclusively for folks on producthunt!

BB-6UWS-FZW2-BXZT-SVZC-JL5X-R7TL-PA

Hey all,

After a great first launch with 600+ Github stars and top 10 on PH, we’re launching a Desktop app for Rescript – the open source text-based video editor.

I built Rescript in a single weekend to prove that personal software can be as good as the commercial options. Rescript is an on-device, private, free and open source video/audio editor that allows you to edit your podcasts and recordings by simply editing the transcript.

No subscriptions, no accounts, no servers.

The desktop app pushes Rescript even further with improved performance and 10+ of the most requested features already shipped!

→ Try it https://getrescript.com/

→ Github https://github.com/wassgha/rescript

Text-based editing is what finally makes video approachable for solo creators. Ran into Descript's paywall hard after a few projects. The desktop app looks like the answer. Does it keep speaker labels consistent across a full session, or do you have to re-identify after switching between tracks?

Love that this runs fully in-browser — local-first tools are underrated. How do you handle long recordings memory-wise, and does export happen client-side too?

Built this after burning tokens and context for every new Claude thread I open. An embedding model uses a fraction of the memory a local LLM does, and gives me back what I (or Claude) are looking for instantly. It's local, offline, and now Claude can just ask the index directly. Would love to hear what you think!

Local + cited-to-the-exact-function is the right combo. Which embedding model runs locally, and how large can an indexed codebase get before search latency starts to hurt?