On July 30, we launched SKI (Voice coding for your agents) on Product Hunt, and by the end of the day it became the #1 Product of the Day. It was a moment we had dreamed about for months, and seeing all the hard work finally pay off is something we'll never forget. To everyone who upvoted, left a comment, reported a bug, shared feedback, or simply gave SKI a try, thank you. Every bit of support meant more than you know.
The launch introduced SKI to thousands of people, brought invaluable feedback, improved our visibility across search engines and AI search, and opened doors to opportunities we wouldn't have had otherwise. Beyond the numbers, it reminded us that great products aren't built in isolation. They're shaped by the people who use them.
As we continue building at Pattern AI Labs, we want this community to grow alongside us. That's why we've started a Discord community where we'll share early previews, product updates, behind-the-scenes progress, and new launches. More than that, it's a place for founders, developers, and anyone curious about the future of AI to connect, collaborate, share what they're building, learn from one another, and help shape what we build next.
does the local ASR model stay resident and eating RAM/CPU the whole session, or only spin up on push-to-talk? that's the same always-on-local-audio battery tradeoff we keep running into
SKI
@sabber_ahamed Stays in RAM, but doesn't use CPU. If it doesn't stay in memory, the cold start takes up time and makes the latency higher. Starts processing once the audio channel sends data.
@johnkg003 got it, so it's memory-resident but idle until the audio channel fires, not actually chewing CPU at rest. that's the distinction we needed, good to know
SKI
@sabber_ahamed Yes, correct :)
SKI
@sabber_ahamed Thanks again for checking SKI
the answer-out-loud part is the interesting bit, most voice tools stop at input. if I start talking again while it's mid-response, does it interrupt itself or do I have to wait it out?
SKI
@sabber_ahamed The current version doesn't have interrupt. It has a silent mode where you can stop the agent from speaking and talk back while it is mid-response. You don't have to wait for the response to end before talking to it again.
@johnkg003 makes sense as a v1 scope - true barge-in with echo cancellation while the agent's own voice is playing is a genuinely hard audio problem, silent mode is a reasonable stopgap
SKI
@sabber_ahamed Yes, the major concern when bringing in interruption is the noisy environment where TTS might pickup something and the agent will get interrupted accidentally. So, we haven't included that in the v1 scope for now.
SKI
@sabber_ahamed Thanks again for checking SKI
Rindler
This is definitely a valuable addition to the coding community. I currently use whisper flow for coding but I am wondering how this differs and maybe how it will be better given that it's tailored toward programmers.
SKI
@arthurdls Whisperflow is a STT application which converts your speech to text, which you can copy and paste to work.
SKI is a skill used by coding agents (in your case, or AI agents in general) to understand what you are speaking to them. Your speech is transcribed and sent to the agent as instructions (with review, which is optional - can set in preferences). Agent then does the task. Once it is complete, the agent updates the status or asks queries to you via voice.
SKI is a complete hands-free coding loop for agents. Please try it out and let us know your feedback.
SKI
@arthurdls Thanks again for checking SKI
Cleanlist AI
The "it answers back" part is the bit that actually matters. Dictation just moves typing to your mouth, you're still sitting there reading output. Having the agent talk back changes the whole loop.
The ambient bit is smart too. Half the reason I don't use voice tools is the ceremony of opening one first.
Going to give this a proper run this week. Congrats on shipping.
SKI
@victor_paraschiv Thanks a lot for taking a look at SKI.
With the agent updating the status back, the loop is complete. Please try it out and let us know your feedback. Thanks a lot for your support!
SKI
@victor_paraschiv Thanks a lot, Victor! You're absolutely right—we believe an agent becomes much more capable when it can talk back and communicate what it's doing.
If you're using Claude Code, the screenshot feature combined with Claude Computer Use is incredibly handy. Give it a try, and we'd love to hear your feedback and learn about your experience.
SKI
@victor_paraschiv Thanks again for checking SKI
@anand_balakrishnan5 since both the speech-in and voice-out models run locally, how much does it lean on CPU or battery while just sitting in the notch waiting for the hotkey, versus when it's actively transcribing? Curious if it's noticeable running all day alongside the coding agent itself!
SKI
@clement_avq It doesn't process anything when sitting idle. When the audio comes in, it gets to work. It keeps everything in-memory. And doesn't start processing audio until data is available. It would not be a noticeable difference on battery backup.
SKI
@clement_avq Thanks again for checking SKI
Congrats on the launch! How do you prevent voice interactions from becoming more distracting than typing during long coding sessions?
SKI
@tarqiya_forgah Thanks a lot for checking out SKI.
The idea behind connecting the skill to agents was that a human could in fact speak much faster than he could type. And once he does that, it gets transcribed and handed over to the agent, who then goes on to complete the assigned task. Once done, it just updates the summary or the status completion by speaking out. If you want it to just show what it ways to say, without voice, we have a silent mode option where the notch in Mac just displays what the agent has to say (no audio). This is also available in the pill mode in both Mac and Windows.
SKI
@tarqiya_forgah Thanks again for checking SKI
if you bring it into a live meeting, how's it telling your commands apart from everyone else talking in the room? seems like the hard part once it's ambient and always-on
SKI
@sabber_ahamed Thank you for checking out SKI.
It has speaker diarization and knows the person by name. But it is always advised to take the agent to the meeting in a space where people are authorized to have information. The agent is grounded in the project it is connected to. The advanced agents, such as Opus / Fable will deny taking destructive actions unless you have authorized it to. Also, you can set a list of allowed commands for the meeting so that you have a better control over the tool use while the agent is in meeting.
@johnkg003 diarization for the who plus a hard allowed-commands list for the what is a good split. the permission scoping is the part that actually makes ambient-in-a-meeting feel safe, not just the voice ID
SKI
@sabber_ahamed Yes, the agents usually know who their owner is and will most likely refuse to take any destructive actions on call.
SKI
@sabber_ahamed We've been working in the meeting intelligence space for quite some time. The experience we've gained from building www.agentcall.dev and www.getsun.io has been invaluable in helping us build this.
Please try SKI and let us know your feedback
SKI
@sabber_ahamed Thanks again for checking SKI