Launch HN: Hyprnote (YC S25) – An open-source AI meeting notetaker

5 days ago

Hi HN! We're Yujong, John, Duck, and Sung from Hyprnote (https://hyprnote.com). We're building an open-source, privacy-first AI note-taking app that runs fully on-device. Think of it as an open-source Granola. No Zoom bots, no cloud APIs, no data ever leaves your machine.

Source code: https://github.com/fastrepl/hyprnote Demo video: https://hyprnote.com/demo

We built Hyprnote because some of our friends told us that their companies banned certain meeting notetakers due to data concerns, or they simply felt uncomfortable sending data to unknown servers. So they went back to manual note-taking - losing focus during meetings and wasting time afterward.

We asked: could we build something just as useful, but completely local?

Hyprnote is a desktop app that transcribes and summarizes meetings on-device. It captures both your mic input and system audio, so you don't need to invite bots. It generates a summary based on the notes you take. Everything runs on local AI models by default, using Whisper and HyprLLM. HyprLLM is our proof-of-concept model fine-tuned from Qwen3 1.7B. We learned that summarizing meetings is a very nuanced task and that a model's raw intelligence (or weight) doesn't matter THAT much. We'll release more details on evaluation and training once we finish the 2nd iteration of the model (still not that good we can make it a lot better).

Whisper inference: https://github.com/fastrepl/hyprnote/blob/main/crates/whispe...

AEC inference: https://github.com/fastrepl/hyprnote/blob/main/crates/aec/sr...

LLM inference: https://github.com/fastrepl/hyprnote/blob/main/crates/llama/...

We also learned that for some folks, having full data controllability was as important as privacy. So we support custom endpoints, allowing users to bring in their company's internal LLM. For teams that need integrations, collaboration, or admin controls, we're working on an optional server component that can be self-hosted. Lastly, we're exploring ways to make Hyprnote work like VSCode, so you can install extensions and build your own workflows around your meetings.

We believe privacy-first tools, powered by local models, are going to unlock the next wave of real-world AI apps.

We're here and looking forward to your comments!

183 comments

yujonglee

Reply
itsalotoffun  4 days ago

I'm always amazed at these relatively tiny projects that "launch" with a "customers" list that reads like they've spent 10 years doing hard outbound enterprise sales: Google, Intel, Apple, Amazon, Deloitte, IBM, Ford, Meta, Uber, Tencent, etc.

  • Aurornis  4 days ago

    This one is especially bad because I doubt all of those companies allow employees to install unapproved software that records meetings and uses so many 3rd party APIs

    The social proof logo list is an old scheme on the growth hacking checklist. There was a time when it was supposed to mean the company had purchased the software. Now it just means they knew someone who worked at those companies who said they’d check it out.

    At this point, when I visit a small product’s landing page and see the logo list the first thing I think of is that they’re small and desperate to convince me they’re not.

    • snthpy  3 days ago

      Wow, I had no idea that the bar was that low. That's ridiculous. I think I'll follow your approach from now on.

  • herval  4 days ago

    We’re firmly in a world where “cheat on everything” is an acceptable business, startups that were hacked together in a week at YC claim they have SOC2 and vibecoded GPT wrappers claim they “trained a model”. Shameless lying took over tech, and if anyone catches you lying, you double down, make a scene and a bunch of podcasts will talk about you. Free advertising!

    Of course, dishonesty is as old as time, but these last couple of years have been hard to watch…

  • paulhart  4 days ago

    Yeah, IBM employee here, not speaking on behalf of the company, own opinions etc. The odds this is approved for employee use are essentially zero.

  • johntopia  4 days ago

    have to admit that we did some logo plays. but our users are really all over the place and just wanted to show it off! i am not sure how it looked but that's why we didn't use terms like "teams" or "customers" to be honest while showing some validation.

    • 48terry  4 days ago

      > i am not sure how it looked

      Well, it looks a lot like you're playing word games to get clout-by-association that you don't necessarily deserve. That doesn't seem like something an authentic person (or people) would try to do. Are the other claims about your team and software equally unserious?

yonl  5 days ago

Congrats on the launch. I never understood why an AI meeting notetaker needed sota LLMs and subscriptions (talking about literally all the other notetakers) - thanks for making it local first. I use a locally patched up whisperx + qwen3:1.7 + nomic embed (ofcourse with a swift script that picks up the audio buffer from microphone) and it works just fine. Rarely i create next steps / sop from the transcript - i use gemini 2.5 and export it as pdf. I’ll give Hyprnote a try soon.

I hope, since it’s opensource, you are thinking about exposing api / hooks for downstream tasks.

  • Aurornis  4 days ago

    > I never understood why an AI meeting notetaker needed sota LLMs and subscriptions

    I’m the opposite: If something is expected to accurately summarize business content, I want to use the best possible model for it.

    The difference between a quantized local model that can run on the average laptop and the latest models from Anthropic, Google, or OpenAI is still very significant.

    • pollyrocket  3 days ago

      For summarizing context, it's not that far off. I've summarized notes using Claude Sonnet 3.7 and Qwen3 8b, and there a difference, but not huge.

  • satvikpendem  5 days ago

    Can you share the Swift script? I was thinking of doing something similar but was banging my head against the audio side of macOS.

  • yujonglee  5 days ago

    What kind of API/Hooks you expect us to expose? We are down to do that.

    • headcanon  4 days ago

      registering an MCP server and calling an MCP tool upon transcript completion (and/or summary completion) would help (check out actionsperminute.io for the vision there).

      Calendar integration would be nice to link transcripts to discrete meetings.

      1 reply →

teiferer  5 days ago

Nice!

Would be great if you could include in your launch message how you plan to monetize this. Everybody likes open source software and local-first is excellent too, but if you mention YC too then everybody also knows that there is no free lunch, so what's coming down the line would be good to know before deciding whether to give it a shot or just move on.

  • yujonglee  5 days ago

    For individuals:

    We have a Pro license implemented in our app. Some non-essential features like custom templates or multi-turn chat are gated behind a paid license. (A custom STT model will also be included soon.) There's still no sign-up required. We use keygen.sh to generate offline-verifiable license keys. Currently, it's priced at $179/year.

    For business:

    If they want to self-host some kind of admin server with integrations, access control, and SSO, we plan to sell a business license.

polyaniline  4 days ago

I just tried to build on Linux and it keeps panicking because it requires dozen(s) of API keys. I was not expecting that from local first software.

crashabr  4 days ago

Do you intend to reach feature parity with something like MacWhisper? I'd love to switch to something open source, but automated meeting detection, push to transcribe (with custom rewrite actions) are two features I've learned to love, beside basic transcript. I also enjoy the automatic transcription from an audio, video or a even a YouTube link.

But because MacWhisper does not store transcripts or do much with them (other than giving you export options), there are some missed opportunities: I'd love to be able to add project tags to transcripts, so that any new transcript is summarized with the context of all previous transcript summaries that share the same tag. Thinking about it maybe I should build a Logseq extension to do that myself as I store all my meeting summaries there anyway.

Speaker detection is not great in MacWhisper (at least in my context where I work mostly with non native English speakers), so that would be a good differentiation too.

  • johntopia  4 days ago

    definitely planning to catch up to other tools - we ship FAST!

    automated meeting detection - working on this. push to transcribe - want to understand more about this. (could we talk more over at our discord? https://hyprnote.com/discord)

    if you're using logseq, we'd love to build an integration for you.

    finally, speaker identification is a big challenge for us too.

    so many things to do - so exciting!

    • dexterdog  4 days ago

      > finally, speaker identification is a big challenge for us too.

      But your home page makes it looks like you already have it. I just tried it in a 30-minute meeting with 20 people and it put the entire conversation under a single speaker, in a single paragraph.

      1 reply →

btown  4 days ago

How are you balancing accuracy vs. time-to-word-on-live-transcript? Is this something you're actively balancing, or can allow an end user to tune?

I find myself often using otter.ai - because while it's inferior to Whisper in many ways, and anything but on-device, it's able to show words on the live transcript with minimal delay, rather than waiting for a moment of silence or for a multi-second buffer to fill. That's vital if I'm using my live transcription both to drive async summarization/notes and for my operational use in the same call, to let me speed-read to catch up to a question that was just posed to me while I was multitasking (or doing research for a prior question!)

It sometimes boggles me that we consider the latency of keypress-to-character-on-screen to be sacrosanct, but are fine with waiting for a phrase or paragraph or even an entire conversation to be complete before visualizing its transcription. Being able to control this would be incredible.

  • yujonglee  4 days ago

    It is more like ai model problem(then app logic. doing it more frequently will require more computation. Things like speculative decoding can help though).

    Doing it locally is hard, but we expect to ship it very soon. Please join our Discord(https://hyprnote.com/discord) if you are interested to hear from us.

mentalgear  4 days ago

Looks great & kudos for making it local-first & open-source, much appreciated!

From a business perspective, and as someone looking also into the open-source model to launch tools, I'd be interested though how you expect revenue to be generated?

Is it solely relying on the audience segment that doesn't know how to hook up the API manually to use the open-source version? How do you calculate this, since pushing it via open-source/github you would think that most people exposed to it are technical enough to just run it from source.

  • yujonglee  4 days ago

    I mentioned about the monetization plan in other threads! (search with 'license').

    Hope that make sense

jihomie  3 days ago

Hello, I am a manager at a tech firm. Came across your product, and feeling this could be a thing I was looking for these days - recently got banned of some notetakers. Impressive demo video btw!

First of all I wanna try this on my personal device. How can I implement your product in my devices?

I also want to discuss some points (on-device, self-hosting, open-source, etc.) to persuade our compliance team, as well

  • johntopia  3 days ago

    would love to support you! you can just download it from our website if you're using a mac.

    • jihomie  3 days ago

      Thanks for the prompt response.

      How about the Window(more of native Window OS than VM or WSL) case? What's your plan for Window version launch and support?

      3 replies →

jpencausse  2 days ago

Question :

In a nutshell you ask an LLM to summarize the transcript and input some notes to help frame the results that's all ? It could be done with ChatGPT + Notepad ?

Yes it could be fully local, but to do that you need a serious computer and a way to distribute Whisper (on Windows)

FYI I tried something close, but comparing Plaud and Whisper (on Windows) Plaud wins far ahead.

p2hari  5 days ago

I just downloaded on mac M4 pro mini. I installed the apple silicon version and try to launch it and it fails. No error message or anything. Just the icon keep bouncing on the dock. I assumed it needs some privacy and screen recording and audio permissions and explicitly gave them, however still just jumps on the dock and the app does not open. (OS, mac sequoia 15.5)

theodorewiles  5 days ago

Looks really cool - I noticed Enterprise has smart consent management?

The thing I think some enterprise customers are worried about in this space is that in many jurisdictions you legally need to disclose recording - having a bot join the call can do that disclosure - but users hate the bot and it takes up too much visibility on many of these calls.

Would love to learn more about your approach there

  • johntopia  5 days ago

    yes, we’re rolling out flexible consent options based on legal needs - like chat messages, silent bots, blinking backgrounds, or consent links before/during meetings. but still figuring out if there's a more elegant way to do this. would love to hear your take as well.

nashashmi  5 days ago

I was talking about this a week ago. One person wanted to make a pdf tutorial of how to use a software. I asked him to record himself in teams and share his screen and have AI take notes. It will create a fabulous summary with snapshots of everything he is going over.

ljosa  4 days ago

I just tried this for a standup meeting, and the inability to tell who said what is a show stopper. You'll need to add speaker diarization for this to be useful for more than 1:1s.

  • johntopia  4 days ago

    agreed - we understand that speaker identification is crucial when it comes to extracting action items and we're actively working on this!

whitefang  3 days ago

I started building something like this for myself and though it would be useful. I have a decent mac only working working but what you have built is great.

This one is completely open source and local, no api calls required at all. Quality can better with newer models which has to be worked on.

https://github.com/acorn-globus/waveflow

rushingcreek  5 days ago

Congrats on the launch! I'm very bullish on how powerful <10B-param models are becoming, so the on-device angle is cool (and great for your bottom line too, as it's cheaper for you to run).

Something that I think is interesting about AI note taking products is focus. How does it choose what's important vs what isn't? The better it is at distinguishing the signal from the noise, the more powerful it is. I wonder if there is an in-context learning angle here where you can update the model weights (either directly or via LoRA) as you get to know the user better. And, of course, everything stays private and on-device.

  • yujonglee  5 days ago

    > How does it choose what's important vs what isn't?

    The idea of Hyprnote is that you write chicken-scratch-raw note during the meeting(what you think is important), and AI enhance based on it.

    On-device learning is interesting too. For example, Gboard: https://arxiv.org/abs/2305.18465

    And yes - we are open to this too

b0dhimind  5 days ago

Look forward to testing the Windows version. Hope it has ability to also upload recordings, etc. Meetily is nice but setup feels too convoluted, with a backend and frontend being required to separately install...

woadwarrior01  5 days ago

Congrats on the launch. Is there a reason why the app isn't sandboxed?

tito  2 days ago

Hey, from the demo video, I thought it was entering the answers to your pre-entered questions based on the transcript, but it looks like you're just typing in the answers. Is that right?

marcofiset  5 days ago

The only issue I have with those tools, and I have not seen a single one even acknowledge this, is that it becomes completely useless when holding meetings in a hybrid fashion where some people are remote and others are in the office with a shared mic.

Almost all of our meetings are hybrid in this way, and it's a real pain having almost half of the meeting be identified as a single individual talking because the mic is hooked up to their machine.

It's a total dealbreaker for us, and we won't use such tools until that problem is solved.

  • yujonglee  5 days ago

    It can be solved with speaker segmentation/embedding models, although it is not perfect. One thing we do with Hyprnote is that we have a Descript-like transcript editor that allows you to easily edit/assign speakers. Once we integrate a speaker diarization model with that, I think we'll be in good shape.

    If you are interested, you can join our Discord and follow updates. :) https://hyprnote.com/discord

    • mijoharas  4 days ago

      Oh awesome, I was reading through to see about whether it had speaker diarization (why I got rid of my whisper script I use).

      I'll look forward to the Linux version.

      Is there any chance of a headless mode? (I.e. start, and write transcript to stdout with some light speaker diarization markup. e.g. "Speaker1: text")

      1 reply →

    • apwell23  5 days ago

      our conference rooms even have some sort of rotating camera contraption that automatically focus on the person speaking

  • abtinf  5 days ago

    I forbid this kind of meeting on my teams.

    Either everyone is in the same physical room, or everyone is remote.

    The quality of communication plummets in the hybrid case:

    * The physical participants have much higher bandwidth communication than those who are remote — they share private expressions and gestures to the detriment of remote.

    * The physical participants have massively lower latency communications. In all-online meetings, everyone an adjust and accommodate the small delays; in hybrid meetings it often locks out remote participants who are always just a little behind or have less time to respond.

    * The audio quality of remote is significantly worse, which I have seen result in their comments being treated as leas credible.

    * Remote participants usually get horrible audio quality from those sharing a mic in the room. No one ever acknowledges this, but it dramatically impacts ability to communicate.

  • nashashmi  5 days ago

    you might need an AI for in-person meeting first. Such tools are available to doctors who see patients. The note taking is great but I think it is skewed towards one-person summary where the name of the patient remains unknown. I wonder if the same tool can take notes if two patients are in the room and distinguish between each one.

    The second tool is likely hardware limitation. A multi-cam-mic with beam forming capability to deconstruct overlapping sounds.

    • johntopia  5 days ago

      hyprnote can be used for in-person meetings as well! we have doctors like ophthalmologists or psychiatrists using it right now. and yes - definitely going to be working on speaker identification as it crucial.

  • dkdcio  5 days ago

    I think if you put N-1 mics in the room (where N is the number of people) you could easily identify all individuals...

AlexErrant  4 days ago

I'm building a voice-controlled app, so it requires transcription. The problem is that the app talks back, so some amount of AEC is needed. It looks like your "generic" AEC is ML powered, but your planned platform specific AEC may use non-ML tech (i.e. looks like you're considering Windows/Mac specific APIs). Have you had a bad experience with AI powered AEC?

yunohn  5 days ago

Super cool, congrats on the launch - will be trying this soon! I noticed it’s using Tauri - what are your main takeaways from building a local inference desktop app with it?

  • yujonglee  5 days ago

    Thanks. I learned that running a server on the Rust side and calling it from a TypeScript frontend is a good approach. For example, we run an OpenAI-compatible server using a Tauri plugin (https://github.com/fastrepl/hyprnote/tree/main/plugins/local...) and call it using the Vercel AI SDK.

    • ashishact  5 days ago

      I wanted to build this for myself. Could never figure out how to get audio output from Mac. Tried almost all audio loopback driver (Blackhole, Soundflower ...). There was problem everywhere wrt security.

      Even tried making a teams meeting bot. But Teams doesn't give live audio to developer unless you are a special partner.

      Glad you made this. Will play around

      2 replies →

kpen11  5 days ago

This is really cool! I've been using Obsidian more and more as a second brain and getting data in has consistently been the point of failure, so I've been wanting something just like this. Specifically something that runs locally and offline.

Is the future goal of Hyprnote specifically meeting notes and leaning into features around meeting notes, or more general note taking and recall features?

Myrmornis  4 days ago

This looks exactly like something that I've been saying to myself that I want! Installed, and already used to transcribe two meetings. At the moment I think that I'm going to carry on using this. Keep going, it looks great!

Adrig  5 days ago

Since this isn't available yet on Windows, what would be the glue & duct tape alternative? Record audio and dump it in chatGPT? Or do you need to create some kind of automation with n8n / Zapier? I don't have that many meetings but it could be nice to have

anyg  4 days ago

This is perfect timing! I just cancelled my fireflies.ai subscription yesterday because it just felt unnecessary. I prefer using less platforms and more tools, especially those that can work under the surface.

joos3  4 days ago

Let's go! Trying it out, appreciate you building this as the hosted/online alternatives (Granola, Notion recorder, ChatGPT recorder etc) don't really make sense if you can record locally.

turnsout  5 days ago

Congrats! I'm currently a Granola user, and wanted to build this myself a while back. But I probably wouldn't have gone as far as fine-tuning a small model for meeting summarization. Can't wait to try it out!

  • yujonglee  5 days ago

    Super! let us know how it goes! (we have discord channel too)

    • turnsout  5 days ago

      Installation and onboarding was nice & smooth!

      A few random bits of realtime feedback:

      You have an icon with the Finder face labeled "Open Finder view." I would expect this to open the app's data folder in the macOS Finder. Instead, it opens an accessory window with some helpful views such as calendar view. I'd encourage you to find another name for that window, because it's too confusing to call it "Finder" (especially with the icon).

      I'd also add a menu item for Settings (and Command-comma shortcut) in the Application menu.

      You also need a dark mode at some point.

      Finally, I'm not sure where note files end up. Seeing that there's an Obsidian integration, I would love an option to save notes in Markdown format into a folder of my choice. I'm an iA Writer user, and would love to have meeting notes go directly into my existing notes folder.

      I'll let you know how the actual functionality is working for me after my next few meetings!

      1 reply →

ls-a  5 days ago

Looks great. From my experience Tauri team has no clue what mobile is and they're not interested in fixing mobile issues. I can already tell the mobile version will be a disappointment.

  • yujonglee  5 days ago

    We tried Tauri mobile (few months ago) and had the same disappointment. We will use RN or Dioxus(to share Rust code) for the mobile version. So it will be cool :)

nkotov  5 days ago

Really cool - how does it compare to Granola outside of the OSS part?

  • yujonglee  5 days ago

    Local-first, controllability(custom endpoint part), and eventually extensibility(VSCode part of the post)

    We're putting a lot of effort into making it run smoothly on local machines. There are no signups, and the app works without any internet connection after downloading models.

    • ashishact  5 days ago

      One of the things I would want to do is - As the meeting is going on - I would like to ask a LLM what questions I could ask at that point in time. Especially if it's a subject I am not expert in.

      Would I be able to create an extension that could do this?

      3 replies →

pokechamp  5 days ago

another free tier (but not opensource) recording tool Ive been using is MacWhisper. Does this and more all locally too. Will try hyprnote out because its neat to do the transcription in real time and its note taking purposes

https://goodsnooze.gumroad.com/l/macwhisper

danneezhao  4 days ago

How to deal with the AI meeting recording function that comes with conference software？

  • johntopia  4 days ago

    hyprnote listens to both sounds coming out(system audio output) and going in(microphone input) so it will work perfectly fine with virtual meetings.

swyx  5 days ago

congrats, app looks gorgeous. def a good tauri codebase to study ( been using https://deepwiki.com/fastrepl/hyprnote)

any interest in the Cluely-style live conversation help/overlay?

  • johntopia  5 days ago

    thanks! we're focusing on local-first at the moment which makes real-time assist a bit challenging. but we definitely have something similar on our mind - for example, a live summary of what's been going on during the past 5 mins. planning to support this via extension in the future!

eawgewag  4 days ago

I honestly don't care that much if Granola uploads my notes to some external service. Work conversations are not private anyways.

HOWEVER, I was extremely disturbed to find out that Granola was automatically making each of my notes folders public to the entire organization. Including 1:1s, you guessed it.

Yeah, really shook my trust in this product by a lot.

  • johntopia  4 days ago

    hyprnote stores and processes data entirely on-device — but you can also bring your own model for better quality and have controllability.

kriro  4 days ago

Curious what made you decide to start with the macOS version. Some insight that most potential customers are on Macs or simply "we use Macs, dogfood time"? I'd always target Windows first for these tools.

Looks cool, I'll wait for the Linux version and try it.

Integration of automatic translations could be an interesting business plan value-add. Branching out into CRM things also makes sense to me.

Good luck, keep shipping.

  • yujonglee  4 days ago

    Agreed. We are working on Windows too.

    For macOS, it is dogfooding, and also lots of AI models are Mac optimized + overall Mac device more powerful than windows, so suited to on-device thing.

hidelooktropic  4 days ago

Unfortunately the GPL license makes it dead in the water for using something like this at work

:(

  • mentalgear  4 days ago

    Why? You can use it, only if you want to extend the tool as part of your (paid) work, you are meant to contribute the code changes back.

    But none of it should prevent someone from just using it (GPL does not mean any usage data is being made "public").

tomr75  5 days ago

why use whisper over parakeet? how will you monetise?

  • yujonglee  5 days ago

    Whisper is more lightweight and smaller. We will support Parakeet in the near future too.

    Monetization:

    For individuals: We have a Pro license implemented in our app. Some non-essential features like custom templates or multi-turn chat are gated behind a paid license. (A custom STT model will also be included soon.) There's still no sign-up required. We use keygen.sh to generate offline-verifiable license keys. Currently, it's priced at $179/year.

    For business: If they want to self-host some kind of admin server with integrations, access control, and SSO, we plan to sell a business license.