PH热榜 | 2026-08-28

Hey Product Hunt 👋

I'm Mingtian, co-founder of PageIndex.

PageIndex lets you ask across your entire document set and verify every answer down to the source line. If you work with long and professional documents where accuracy and traceability matter, financial reports, legal contracts, textbooks, research papers, etc. and you can't afford to trust an answer you haven't checked, this is for you.

✍️ Here's how it works:

1. Drop in your entire folder. Folder structure is preserved, and everything you add stays in your knowledge base, so you build it once instead of re-uploading the same files into every new chat.

2. Ask your hardest question across all of it. Ask the one you actually need: a number buried in an appendix, a clause that only makes sense with the definition twelve pages back, a figure that has to be pulled from a table and compared across four files. PageIndex goes and gets all of it.

3. Verify answers in one click. Every answer comes with micro-citations. Click one and the source document opens beside your chat at the right page, highlighted at the exact line the number came from.

Why you want PageIndex:

• Checking a number takes one click instead of an afternoon of opening PDFs

• One question runs across hundreds of documents, not one file per chat

• Tables and charts are read in context

• Your library compounds each quarter instead of starting from an empty chat

Leading accuracy on FinanceBench. 30K+ people use it, and the retrieval engine underneath has 35K+ GitHub stars and hit #1 on GitHub Trending.

🎉 Launch offer: code PRODUCTHUNT gets you one month of Pro free.

👉 app.pageindex.ai

Thanks for checking us out. I'll be here all day.

Hi PH, Ray here, co-founder and CTO of PageIndex.

I did my PhD in databases at Oxford. Years spent on indexing, and I was firmly on the side of vector databases.

I've changed my mind about them being the right infrastructure for retrieval in AI systems. I want to be precise, because the claim is not "vectors are dead".

An index is defined by the one question it can answer. A vector index answers: which chunks look most like this query?

• Right question for broad recall over messy, conversational text. Vectors will keep doing that job well.

• Wrong question for a long, professional document.

Two reasons it breaks down there:

• The passage you need may share almost no wording with how you asked for it.

• The answer often sits behind a cross-reference like "see Appendix G". No amount of similarity gets you through that pointer.

So we index the structure instead of the surface.

• The document's tree sits inside the model's reasoning context.

• It decides where to look next, not what looks similar.

• Retrieval becomes navigation. Navigation leaves a path, which is why every answer can point back at the source line.

If you want to stress-test it: upload the longest and most complex document you own, then tell me what happens. Bug reports are worth more to me today than upvotes.

Can I share a single answer with its citations as a link, so a colleague can see the sources without an account?

For contacts and reports I'd rather spend a few seconds checking a source than blindly trust a summary. Does it preserve the orignal document formatting when I kump to the citation?

Hiii PH, I'm Cathy, GTM at PageIndex.

I'm the non-engineer on this team, which makes me test subject number one. If I can't get a verified answer out of a folder in thirty seconds, it goes back to Ray.

Before this I worked in finance. I studied it at LSE and ground through CFA Level I. None of that helps you at 11pm when you're pulling the inputs for an EBITDA build, then tracing every single number back to the page it came from.

👇 That's the bit that changed for me:

• My filings sit in one folder that stays indexed.

• I ask across all of them at once.

• Every number in the answer carries a reference. One click lands me on the exact highlighted line, right next to the chat.

• The verification pass that used to take an afternoon is now one click.

One favour if you're testing it: please don't ask what Apple earned last year. Any chatbot answers that. Ask the thing only you would know where to look for:

• a covenant threshold buried in an appendix

• a segment number that moved between restatements

• a figure that only appears in a footnote

If anything confuses you in the first two minutes, tell me bluntly. That's the feedback I act on fastest. I'm in the comments all day ☕

can you suggest how pageindex is better than chatgpt or claude code. what makes pageindex better?

Looks pretty scientific :) Do you collaborate with some universities and research centres? :)

Asked three follow-up questions in a row and it kept the context each time without me re-explaining anything.

Congrats! Avoiding embeddings is an ambitious architectural choice. I'm especially curious whether this makes the results more interpretable, since each search decision could potentially be traced through the document tree.

Feels like what Perplexity does for the web, but pointed at my own document library instead.Great Launch!

Really good product from popular opensource lib. Congrats on this launch!

Love how easy it is to verify an answer without leaving the chat.

I like the exact-line citations. Congrats!

Love that I don't have to keep scrolling through a 500-page PDF anymore！

what is the latency for this kind of search?

Super useful tool. Congrats Mingtian and team. Does PageIndex tell me which version it pulled from? Data rooms I read usually have the same number living in three versions of the same file - original, amended, restated.

Great use case at the core of the project! I often check contracts with ChatGPT the same way when I need to quickly find payment terms or deadlines.

Congrats on the launch!
Being able to actually verify an answer instead of just trusting it is the real unlock here, most doc AI tools skip that part entirely.
Upvoted 🚀

Congrats on launching PageIndex! Verifying answers with clickable citations solves a real pain point. How do you handle documents with scanned or image based pages?

This seems especially valuable for investors and researchers.

The document should always be the source of truth. I like that AI is supporting the reading process instead of replacing it.

Hey Product Hunt 👋

I'm Jason, one of the makers of Caddi. Caddi is an agent that builds agents.

We build for law firms, RIAs and accounting firms, and the same thing kept coming up: the work isn't hard, it's just endless. A contract comes back signed, someone downloads it, renames it to the firm convention, files it to the matter, logs it in the CRM. Forty times a month. Every month.

Automating that has always meant a project. Scope it, spec it, configure it, get a developer. So it never happens, and people keep doing it by hand.

Caddi takes a different path, and it starts before you build anything: it reads your stack and tells you which work you repeat most, ranked by impact.

You pick one, and then you teach it like a new hire, except this new hire already knows the job. It follows your guidance or guides you with best practices, and probes for edge-cases: not just "download the contract", but "say how you tell an executed copy from a draft".

When it hits something ambiguous it stops and asks, with the options laid out and a confirm button. What if only one party has signed? Your answer doesn't disappear into a transcript. It becomes a rule in the agent.

What comes out reasons with AI where judgment is needed and runs deterministic code where it isn't, so the messy cases get handled without the exact parts going off-script. Then it runs in the cloud, executing thousands of actions at a time.

And you can see all of it. Every run is a log of what it decided, step by step, with the permission it used to do it: read, create, send.

We'd genuinely love your feedback, especially on where you'd trust an agent and where you wouldn't. That line is the whole product for us.

The runby run visibility is a really strong touch. Being able to see what the agent decided and which permissions it used Makes the automation feel much more accountable.

I like how Caddi focuses on the repetitive work people actually do instead of asking them to become workflow designers.

How does it handle small change in the process? That's usually where automations break for me.

I like the screenshare approach. Feel more natural than setting up a workflow step by step.

Talking out loud while doing the task feels way more natural than manually building a giant diagram.

The screen share to agent workflow feels very thoughtfully designed. It removes a lot of the friction that normally comes with setting up automation from scratch.

The interesting change here is not just AI automation, but also turning traditional workflows into something reusable. This could lead to the automation of many documentless business processes in the future. This is a great idea.

Love it!

Wow, that's super cool. In most firms two people doing the same task do it differently so whose version becomes the rule? Does it even matter?

I can see this being useful for all those boring admin tasks we keep doing every day.

Can it handle workflows that hop across 3 or 4 different apps in one go?

How well does Caddi handle exceptions when the real process differs slightly from the example?

I literally saw this and I said out loud: "No Way!". Early I was talking to someone on my team, this is literally my new assignmnet but with Copilot. I want to mention this tool to my team.

Congrats to the team. Good call having it stop and ask on the ambiguous ones instead of guessing.

Would this work for healthcare use cases?

what happenes if there is no api to use for a process? how can it handle that?

Here's a pure <60 sec demo of how Caddi discovers and automates repetitive work:

https://youtu.be/GdCP1bJp_3o?si=lcJFF1mE6S0l3DYK

Congratulations on the launch! Probably the most common use case is working with websites. Can I ask it to go to Amazon, find a specific product, and place an order for me?

How much flexibility do users have when they need to adjust only part of an exisiting workflow?

ui looks very interesting.

what happens if a website changes its layout slightly? Does it adjust or break?

Hi everyone!

Reachy Mini has a new little buddy 🦆

Quite a week for @Hugging Face. Right after the reported big deal, they’re launching something very on-brand with Pollen Robotics: a $399 open-source biped you can train yourself.

It’s desk-sized — 25cm tall, under 800g — and already walks, picks things up, gets back up after a fall, and even roller-skates.

You can mess with how it moves in the browser first, then put the new behavior on the real robot.

The whole software stack is open under Apache 2.0.

Reachy is the one you talk to. This one you teach to walk!

This is truly fun. Big fan of fun. :)

im ordering for my son right now, love it.

So sweet!
I like it 💞

Congrats on launch number twenty. Apache 2.0 on the software stack of a $399 robot is a lot to give away.

Love that you made a real bipedal robot hackable under Apache 2.0, turning sim-to-real RL from a lab luxury into something anyone can retrain on their desk.

the sim-to-real part is what i'd actually want to know more about. training in the browser sim and just dropping it onto the real robot sounds great in a demo but usually there's a gap - does it walk fine on the first try or does it need retraining once it's on actual hardware

I’m happy to share this! Gemini Omni 1.1 Flash brings some seriously powerful new controls for AI video generation and editing.

🎬 Extend scenes while preserving characters, lighting & context
🎯 Control the first & last frames for smooth transitions
⚡ Draft ideas quickly in 360p, then upscale your favorites
✨ Upscale outputs up to 4K for production-ready results
📽️ Add up to 3 seconds of video references for movement & consistency

It’s rolling out now to Google AI Studio, Flow, and the Gemini Enterprise Agent Platform. Scene extension is also available in the Gemini app for Plus, Pro & Ultra subscribers.

The level of creative control here is getting crazy! 🚀

Draft in 360p then upscale the favourites is the part that actually changes what a video costs, not the 4K. Where that flow usually breaks is the cheap draft and the expensive final not being the same generation, so someone picks a draft they like and the upscale quietly reinterprets a hand or a face. If the 360p pass is a true preview of the same seed and motion, that's the headline and it's buried. First and last frame control is the other one I'd have led with, it's the difference between a clip and a shot you can cut into something.

Our Claude Tag bill is way to high for it to burn tokens doing nothing

Really smart idea!

As Claude Tag takes off, more and more people will want cheaper agents to do the tasks vs Claude alone. This is the future!

Congrats on the launch!! Been wanting to try a Slack coworker. How does your credit system work in terms of billing?

Congrats.

What makes you different compared to Claude Tag/Victor/Scarlett and other solutions on the market?

Pretty wild amount of developer context. Curious how much this cuts down on token usage.

Congrats on launch number twelve. Good to see a recall number published against 1,179 real developer queries.

This is crazy, this is a problem every company I've talked recently is having and ourself are experiencing. This is an incredible product

How can I control which tools Almanac has access to? What if I only want read access to my emails?

This is awesome. What happens if there's conflicting context on the same topic? like an old slack thread vs a newer doc? How does the brain decide what's the latest? I mean it's not even just dates. Sometimes half the info is in slack and half in a doc.

This is great! The personal vs. company wiki split is very important. How do you handle situations where private context should inform a company task without accidentally exposing that private information?

saw the reply above about currently asking for all permissions upfront rather than scoped access. that's the one thing that'd stop me from connecting real accounts to something that also keeps running after i close my laptop. is scoped/read-only access per tool on the roadmap or is all-or-nothing the plan for a while

Hey Product Hunt 👋

We built aramb because shipping one agent meant signing up for nine different products.

A model provider. A voice API. A browser infrastructure vendor. A sandbox provider. A vector DB. An integrations platform. A metering service. Then Stripe. Then a Redis to hold it together.

Nine SDKs. Nine sets of keys. Nine rate limits to reason about. Nine invoices at the end of the month — and none of them agree on what a "session" is. You spend two weeks on integration plumbing before you write
a single line of the thing that's actually yours.

Every agent product converges on the same seven services. So we stopped rebuilding them and made them primitives:

- Agent — any model, any provider. Route on price, latency, or intent. Swap without a rewrite.
- Voice — sub-400ms realtime. Whisper, ElevenLabs, Cartesia, Deepgram behind one interface.
- Browser — headless Chrome that survives captchas, logins, and 6-hour sessions.
- Sandbox — ephemeral Linux boxes, fenced per end-user, metered by the second.
- Memory — vector + episodic + semantic on one query surface. No RAG plumbing.
- Tools — 3,000+ integrations, pre-authed. OAuth, refresh, retries, rate limits: solved once.
- Billing — every primitive emits a usage event. Pipe it to Stripe, Paddle, or your own ledger.

One API. One key. One bill. You bring the knowledge — a system prompt, a playbook, your domain. We handle the runtime.

const session = await aramb.run({
system_prompt: "./triage-playbook.md",
user: "u_812",
agent: "gpt-5-thinking",
voice: { provider: "cartesia" },
browser: { persist: true },
memory: "episodic",
tools: ["gmail", "stripe", "linear"],
meter: "per_user",
});

Six lines to your first agent. Usage streams straight to your ledger.

The part we're most proud of: per-tenant metering. Every session is tagged to an end-user id, so you can rebill your own customers on day one instead of month six. Plus ATK (Aramb Token Kompressor, patent
pending) — prompt and context compression that cuts token spend up to 50% for the same output.

Pricing: free tier with upto 5,000 credits/mo, then $19 for solo builders and $49 for teams. One shared credit pool, no per-seat fees — unlimited teammates, agents and projects. Idle agents cost nothing.

We're in private beta and genuinely want to be argued with. Which of the seven did we get wrong? Which one would you rip out and replace with your own?

First 500 hunters get a free month of Studio 🎁

— the aramb team
hello@aramb.ai

What made me stop scrolling was the start making money part. Part of me wants to try using this. But your website did not load, so it sounds too good to be true.

Cool app! Tried it out! Love that everything is kept in one place and people can start earning right away. What has been the most surprising thing someone has built with it so far?

using Screenpipe for the last 4 months. I have never seen a team ship so fast and still be so responsive.

Screenpipe and the team is our inspiration for a lot of what we've built - super excited for Louis and the team for making this widely available and accessible in a fundamentally open source way!

Screenpipe has been so helpful for my mental wellness. As a founder, I context switch a lot and it can get really overwhelming, but screenpipe helps me stay on track. I don't have to worry about forgetting and following up on important things that come up everyday. I also like how responsive and supportive the founding team is whenever I have any feedback. Congrats on the launch.

I LOVE using screenpipe to chat with what I did during the day.

Lately, I’ve been using Screenpipe to remind me about unfinished work and create skills for tasks I repeat.

I built screenpipe because capable AI still loses the context of work that already happened on your computer.

The goal is simple: let the agents you already use work from that history instead of another manual recap.

I want the honest version: where does missing context hurt your AI workflow most?

You shouldn’t need to give your AI a long prompt, 10 screenshots, and 2 meeting transcripts before it can help with work you already did on your computer.

We built screenpipe to stop the loop of pasting screenshots, recapping meetings, and explaining what changed.

screenpipe records your screen, audio, and activity on your own computer, then makes that history available to the AI agents you already use through MCP.

So you can ask about a customer call, a bug you chased, or a document you edited without explaining the work all over again.

screenpipe is local-first, source-available, and available on Mac, Windows, and Linux.

We’ll be here throughout launch day. What’s the first workflow you’d stop re-explaining?

Launch-day offer: use code BUSINESS20 for a discount on annual plans. Ends August 28 at 11:59pm PT.

Hey Hunters 👋

My wife and I built Glisio together.

We wanted a simple way to create high-quality product demos and launch videos, but the tools we tried were either too expensive or didn't give us the quality and experience we were looking for.

So we started building Glisio.

Glisio is a native Mac app for product demos, launch videos, and tutorials. Use Clip to record with auto-zoom, system audio, mic, and optional webcam, then edit and export local MP4. Use Snap for styled screenshots: background, crop, annotate, then copy or export. No cloud uploads. Files stay on your Mac.

What started as a tool we wanted for ourselves turned into a product we've been continuously building and improving.

You can record, edit, copy, and export for free. Free exports have a small watermark, while Pro removes it. Pro Lifetime is $79 one-time, and we're offering 40% off for this Product Hunt launch. If you prefer monthly, Pro is $9.99/mo.

We'd genuinely love your feedback. If you make SaaS demos, courses, tutorials, or launch videos, what would you want us to build next?

Download Glisio: https://glisio.com/download

Hey Hunters! 👋

So excited to finally see Glisio on Product Hunt!

Building Glisio together has been such a fun (and sometimes challenging 😅) journey. What started as a tool we simply wanted to use ourselves slowly turned into something we’re really proud to share with others.

We’ve put so much time and care into making Glisio simple, fast, and completely local on Mac — and we’d love to hear what you think!

If you try it, let us know what you love, what you’d change, or what you’d like to see next. Your feedback means a lot to us.

Thanks for supporting our little project!

Most of my Mac recordings end up as GIFs in docs, and hand-cropping the zoom is the slow part. What does the auto-zoom actually follow, clicks or the cursor?

Wait a sec, is this a 3D asset builder or a video game builder? or both?

Liked the idea of working together Gabe! Thanks for hunting tihs!

Love the fact that you're able to build collaboratively, wherever your teammates are

hey, classic question: how is it different than langsmith evals or weights and biases evals?

40 evaluators is the number I'd push on. Most eval suites I've used come down to another model grading the output, so the eval inherits the same failure mode as the thing it's grading. For each of those 40 I'd want to know upfront whether it's deterministic or a judge model, because I trust those two very differently. BYOK with no markup on API spend is the right call though.

notch utilities feel like a crowded lane right now, i counted at least 3-4 similar apps on this same page. what's pulling people to fide island specifically over the others - is it the currency/calc stuff, or something else

Hey Sergey, the product looks super cool. Unfortunately, I couldn't install it on my mac as it flagged for malicious files.

The notch is dead space on my MacBook. The apps I work on run long file transfers, and a glanceable progress bar up there would earn its spot. What does the island show when nothing is running?

Hi Product Hunt! I built CTRL Micro because I loved the idea of a physical control deck for AI work, but options like Codex Micro are out of stock, and I wanted something I could use immediately on the iPhone already in my pocket.

What began as tactile shortcuts became a fuller companion for Mac and agent workflows. CTRL Micro lets you:

• Monitor Codex, Claude, and Cursor and feel when a task needs attention
• Open tasks, read responses, answer prompts, and act on approvals
• See your Mac in a live Desktop View
• Move the pointer, scroll, click, and type with Trackpad and Keyboard surfaces
• Dictate into desktop apps with transcription processed locally on your Mac
• Connect nearby or through end-to-end encrypted Remote Anywhere

Privacy guided the architecture. Your controller pairs with one or multiple Macs, no CTRL Micro account is required. Screen, input, command, and dictation content are processed locally and encrypted.

The iPhone and iPad app is live on the App Store, and the Mac Host is available from ctrlmicro.com.
I would especially love feedback on the Game Boy-style Desktop View and which controls you keep reaching for.

Using the Taptic Engine instead of waiting on deck hardware is clever. I work on Mac apps all day, so a spare iPhone as a control surface fits neatly. Does the haptic pattern differ when an agent finishes vs fails?

nice idea, rule of thirds is one of those things that's easy to explain but hard to actually feel while framing a shot. since it needs camera access in the browser, is the scoring happening on-device or does it send frames to a server? would want to know before pointing it at anything private

Ok I am trying to understand it. Low key I do not understand it. Like I know its a good marketing tactic. It IS different. I feel bad having you explain but I really did not understand the examples.

Hey Product Hunt 👋

I’m excited to launch NotchDrop v2 today.

I originally built NotchDrop around a simple idea: the MacBook notch takes up space anyway, so why not make it useful?

With v2, NotchDrop has grown into a small workspace for everyday Mac tasks, with features like NotchDrop Studio, Live Wallpaper, Calendar, Calculator, System Stats, Keep Awake, Draw Signature, AirDrop, local file sharing, and a redesigned Notch Terminal.

A lot of this update came from feedback and feature requests, so thank you to everyone who has supported NotchDrop so far.

I’d love to hear what you think especially which feature you’d use the most, and what you’d like to see next.

Check it out!

the name-your-own-price ticket thing is the part that stands out to me. is that peer-to-peer resale between users, or are venues/artists actually setting a floor on it? asking because those usually turn into pretty different products depending on who controls the price

Hey I’m Max, one of the co-founders and CEO of CrowdVolt! I’ve built this alongside my co-founders @aria__mohseni and @joshkarol7, and the rest of the dedicated CrowdVolt team.

We believe every night out should be effortless and open to all. There are a million different ways to go out right now (you may go look at Instagram, check a few different apps, etc.), but everything is dispersed. So we’re building the infrastructure to unify it - connecting its fragmented, gate-kept systems into one open layer, with the fan at the center of it all.