Hi Hacker News!

I found myself asking my AI assistant to help me with some non-trivial tasks (describing visual changes to an app, filling in forms, setting up a server instance in a new provider, changing some dns settings in another one, mostly things I know how to do but that would take me extra time to figure out on an unknown UI). I was solving this by taking screenshots all the time and pasting them, when I got tired of that I just copy&pasted the whole text.

With DesktopVisionMCP, I just ask my AI to "look" and it takes a screenshot on-demand of all my monitors, passes it directly to the AI (no file), so it gets the context instead of me figuring out the best way to describe it and my intent.

I'd like to hear your thoughts and feedback.

It is live on Product Hunt today.

PS, this whole thing was written by me, by hand but proofread using DesktopVisionMCP ;)

A few things I expect to come up:

- Privacy: nothing leaves your device.

- Other options (like Otter or Granola): AudioTap is not cloud-based, recordings are yours as local files (instead of a limited access to the transcripts), no extra non-identifiable data analysis to train anything else, offline first (although use case is online calls) and one-time payment vs subscription.

- It uses stereo channels to identify you (as [Me]) and analyzes voice embeddings to cluster other speakers in the system-audio channel (letting it know how many speakers were in the call significantly improves diarization).

Happy to go deep on any of these.

(edit formatting)

Thank you for letting me know, I’ll update both links asap.

nsa.tech is a domain name I acquired some years ago for fun and I’ve been meaning to use for some time now. I understand that most people know of an agency that uses the acronym but hey! We’ve been thinking “network secure access” around the office ;)

I want to build an open protocol for building p2p apps that don't rely on any central servers. For this I'm looking into BitTorrent's DHT for discovery, libp2p for connection and encryption. The idea is that you own your identity, you pick who you trust, and nothing gets stored on servers you don't control. It will work for things like messaging, photo sharing, and private networks where you want real decentralization/federation, not just a blockchain slapped on top of the same old client server model.

What does ActivityPub protocol lacks in your opinion?

The infra intended here exposes TCP and/or UDP connections, it is protocol agnostic; so end-user apps can use smtp, http, activity pub, at protocol, etc. I am going for an infra that can scale by leveraging the DHT as DNS and p2p for end user apps/apis.

(Edit typo)

I've been enjoying working with LLMs for some weeks now, and realized that as faulty as these can get, asking them to put material in place to learn some resource may be worth my time.

I am looking to learn more about Clojure, I read through it all and was happy that it ends up with examples for an actual api app. I may ask for more in depth lessons as I find more suitable apps to be built.

Interesting... the way I workaround'd non-macos-native-sshfs is by using docker and mounting a local folder, do you plan on publishing your tool?

Absolutely: https://github.com/jabedude/sshfs

This looks fun! But the fact that the self hosted version lacks support for some core apps is sad, I'd love to be able to build puter apps! Any plans for an app store-like ecosystem?

Yes! definitely. Please stay tuned :)

Some time ago I was building a mitm proxy myself, then I found out about: https://www.mitmproxy.org/ Maybe you already had it in your radar

That's for HTTP/S and related, as parent said his is for BGP which is a completely different protocol

If you are open to suggestions, libp2p has a good amount of sdks for several languages and you can integrate kademlia for peer discovery

We are open to suggestions :) And the very first generation of P2P Matrix was indeed built on libp2p (and Protocol Labs led Element's Series B). However, the thought experiment here is whether we can get away without a full global P2P overlay at all in the interests of keeping it simple & stupid. We might well end up back at libp2p tho!

Edit: another option on our todo list to look into more is Iroh (https://github.com/n0-computer/iroh)

Some years ago, I co-founded a startup that would run workflows when email messages arrived, any email from any source was parsed and it would trigger "actions" that could include notifications - this sounds like a good use case for it! You don't need the service to expose an api to listen in, since most services end up sending email as last fallback.

Thank you for you feedback!

Do you still run that startup?

With Cozy Watch, I use the GitHub API, never thought about using emails as triggers.

I’ve actually got GitHub emails disabled, they can get pretty spammy.

Sadly no, the feedback loop with users lacked consistency and we never got around finding use cases. One of the visions for it was to help people clean up inboxes, because as you put it, some sources can get pretty spammy. I'll keep an eye out for Cozy Watch, I hope you are successful!

Thank you!

Please let me know if it can be helpful for you or your team.

reach me here: tiago@cozywatch.com