pi-voice  ·  github  ·  npm

i wanted to talk to my coding agent.

so i made it happen. this is pi-voice: you talk, pi does the work, and a voice tells you what happened. it's all in your terminal, on your mac.

π — ~/dev/parser

^ that's the bar above pi's input. it shows who has the floor: you talking, the voice talking, or pi working.

how this happened

when gpt-live showed up in the chatgpt app on my phone i was honestly a little too excited. it felt like the first voice thing that actually talks with you. you can cut it off mid-sentence, it keeps up, it doesn't feel like a walkie-talkie.

and my first thought was: i want this hooked into my coding agent. now.

but there was no api. so i waited. and kept checking. and waited some more.

the day the api came out i opened codex and built my own /voice command. it worked, and it was really fun.

why pi

because this shitty harness is what a simple man needs. (said with love.)

pi is small, it gets out of the way, and you can bend it however you want with an extension. that's all this is: one pi extension. gpt-live is the voice, and your pi session is the brain. same model, same tools, same permission prompts, nothing hidden.

π — ~/dev/parser

stuff i cared about

try it

pi install npm:@lprsrts/pi-voice

then set OPENAI_API_KEY (or put it in your keychain as openai-api-key), type /talk in pi or hit alt+v, and say something. same again to stop.

you need macos 14+, node 24+, the xcode command line tools, and an openai api key with billing. a chatgpt plan doesn't cover it. the voice is about $0.05 a minute, and pi's model costs whatever it costs. the free package is MIT. more in the readme.

heads up

this is a side project. it works for me every day, but it's young and there are rough edges. if something breaks, open an issue and i'll take a look.

— alper (@lprsrts)