For a few months now I've been working with Claude Code out loud. I talk, it talks back, and honestly it's the most natural way I've ever worked with an agent.
Then I noticed nobody around me works like this. My colleagues dictate their prompts, then go straight back to reading everything Claude writes.
Building a voice mode for an agent sounds easy. Speech-to-text in, a voice model out, done in a weekend. That's what I thought too.
My first version was a mess. It kept hearing its own voice through the speakers and taking it as me talking. The interruption handling was either too sensitive or too slow. And once the agent started working, it stopped listening. To stop it or change direction, I had to go back to the terminal, which defeated the whole point.
Each of those needed real engineering to fix: proper echo cancellation, telling an actual interruption apart from a nod, catching what I say while it's working, and getting it to answer in a couple of spoken lines instead of reading whole paragraphs out to me.
Now I just stay in the doc or the code and talk it through, like with someone sitting next to me.
So I cleaned up my setup, packaged it and open-sourced it. It's called Larmor. It works with any AI agent and runs entirely on local models on your Mac. For now it's built just for Apple Silicon.
If you code with an agent, give it one session. I'd love to hear how it feels to you.
curl -fsSL https://t.co/fmX6wij5L4 | bash
https://t.co/rWJ59pHTlV
https://t.co/qbUOelMj2E
Is it really a good idea to give your AI access to your passwords? I agree we're moving towards more autonomous AI that doesn't need babysitting. But giving it access to your most sensitive stuff is literally playing with kerosene and matches. What if the AI is misaligned? Or worse what if someone from outside attack it?
Cool demo but I'd still keep my passwords out of my agent's reach.
I've been working with Claude Code by voice for a while now, and my favourite part is the back and forth before it starts working. Quick question, quick answer, like talking it through with someone sitting next to you.
It keeps listening while it works too, so if I change my mind halfway, I just say it and it picks it up at the next tool call.
I packaged it up and open-sourced it. It's called Larmor, and it runs fully on your Mac.
https://t.co/rWJ59pHTlV
(The video's a mockup with the pauses trimmed, but the agent's voice is the real one.)
@iamdelulusami @TVKVijayHQ@arivalayam First day la Gajanava vazhichu eduthutu poi irukanga nu oath ceremony laye dmk va adichitu. Innaiku nerla poi paathuta healthy politics ayiduma. Naaliki edhume pannama adhukum dmk mela blame podama irundha sari.
@buildwithshyam For me it's A) . atp where we can build almost anything , finding the moat and where we can actually add value to the user , becomes very difficult to find.
I desperately want an idea which I can confidently go all in.
In an ideal world, I wouldn't be proud of who I am, you wouldn't be proud of who you are, and together we can talk to each other without hurting anyone.