I got sick of typing, so I started talking to my terminal
Five days ago I typed a message into my terminal that I was too lazy to keep typing: "hey i am sick of typing. can i use voice directly to ghosty? or any terminal on my machine."
Yes, I spelled Ghostty wrong. That's the point. My hands were done.
The agent came back with a market survey. Top pick: a paid dictation app with a subscription. I told it, verbatim, "funny if you think ill pay to speak to my terminal. aqua can go to hell." Then it offered me three options with trade-offs, and I had to remind it I'm one person with one mouth. One answer. Go research like my money depends on it, because it does.
It came back with Spokenly. Free local models, types into any Mac app. I told it we were installing and testing together, no advisor mode, and if it turned out to be shit we would suffer together.
Install took a minute. Push-to-talk on Right Command, which is smarter than it looks: left hand keeps every terminal shortcut, right hand becomes the talk button. Then it asked me to pick a transcription model, and I planted the bomb I'd step on five days later: I chose the cloud model, because I wanted top accuracy. It was scary good. My review at the time: "the tool is not shit so far."
Then I just lived in it. Five days of talking to my terminal. Prompts, commit messages, half these notes. Somewhere in there I forgot one detail: every word was a network call.
The firewall remembers what you forget
Today dictation froze on "processing", which confused me, because in my head transcription ran on my machine.
It didn't. My firewall (LuLu, free, holds every new connection until you approve it) had a popup waiting: Spokenly connecting to a license server. Harmless call, but it forced me to look. The cloud model from day one was still selected. My voice had been commuting to someone else's computer all week, with my full but expired consent.
The fix was a tab I'd skipped during onboarding: Local. I downloaded NVIDIA Parakeet TDT 0.6B V3, 496 MB, and it is unreasonably fast. Sub-100ms, words land as they leave your mouth, and benchmarks have it beating Whisper and Apple's engine on English while being a fraction of the size. Nothing leaves the Mac anymore. Download, select, done.
The feature I deleted an hour later
Spokenly also has an MCP server, so Claude Code can pop a recorder and ask you questions by voice:
claude mcp add spokenly -- ~/Library/Application\ Support/Spokenly/mcp-bridge.shTested it. Worked. Removed it within the hour. By the time the agent has a question I'm deep in another window, so the recorder pops at nobody. And it inverts who holds the microphone: my hotkey is me deciding when to talk, the MCP is software deciding when to listen. No thanks.
claude mcp remove spokenlyWhat's left
A key I hold when I want to talk, a 496 MB model that types at conversation speed, everything on-device, free. This post was dictated with it. Worst error all day: "Spokenly is back up" arrived as "spoken list back up." I've typed worse.
If you read this far, check your pockets: you already have the setup. Spokenly, Right Command, Local tab, Parakeet V3. And a firewall that makes you think before your apps phone home.