Yesterday morning I was paying a monthly subscription for an AI dictation app. By dinner I had cancelled it. Not because I stopped dictating — I dictate constantly, including most of this article — but because the thing that replaced it is faster, completely private, costs zero dollars a month, and I own every line of it.
It's called Vox. It's free and open source. Hold a key, talk, release — clean text appears wherever your cursor is, in about a second and a half. And here's the part that matters if you run a business: I didn't build it alone. My AI agents did most of the work, in one day, while I gave feedback by voice.
The numbers first
- ~1.5 seconds from releasing the key to pasted text (my old cloud app took noticeably longer)
- $0/month — the speech model runs on the Mac's own chip. No API, no tokens, no account
- 100% private — audio never leaves the machine. For anyone dictating client work, contracts, or anything sensitive, this alone is worth the switch
- 1 day from "I'm tired of paying for this" to cancelled subscription
- 4 AI agents shipping code to the same repo — with me as the only human in the loop
Why I stopped renting
Dictation apps like the one I cancelled are good products. But structurally, you're renting a workflow: your voice goes to their cloud, gets transcribed on their servers, and comes back — and you pay monthly for the round trip. The platform updates when they decide, changes when they decide, and bills you forever.
Meanwhile, the actual intelligence — OpenAI's Whisper speech model — is open source and runs beautifully on Apple Silicon. The only thing the subscription was really selling me was glue code.
Glue code is exactly what AI agents are spectacular at writing.
The build, hour by hour
Morning: I told my main agent (Claude, running on my Mac Mini) what I wanted: press a key, record, transcribe locally, use a local LLM for cleanup, paste at the cursor. It picked the stack — whisper.cpp for transcription, Hammerspoon for the hotkey and paste, Ollama for optional AI cleanup — installed everything, wrote the app, and tested the transcription pipeline with synthesized speech before I ever said a word into it.
Midday: I used it and complained, by voice, into the tool itself. "It's slow." The agent profiled the pipeline and found the model was reloading from disk on every single dictation — so it switched to a persistent server that keeps the model in RAM. Then it discovered language auto-detection was silently running a second full pass and pinned my language. Latency fell from ~5 seconds to ~1.5.
Afternoon: Accuracy pass. Here's my favorite engineering decision of the day: we turned off the LLM cleanup by default. Testing showed the small local LLM occasionally paraphrased what I said — unacceptable for dictation — while a simple deterministic find-and-replace dictionary fixed the actual errors ("super base" → "Supabase") at zero latency. Sometimes the boring solution beats the AI solution, and it takes honest testing to know which is which.
Evening: Polish and distribution. A little animated alien in a pill at the bottom of the screen shows when it's listening. Music fades down automatically while you talk and comes back when you're done. Then the fleet took over: one agent built the landing page, two others — running on my other Macs — found and fixed installation bugs on their own hardware and pushed the fixes to the shared repo. By the time I checked GitHub, four different agents had shipped to the same codebase without stepping on each other.
What installing it looks like now
One line in Terminal:
curl -fsSL https://automatescale.com/vox/install | bash
It installs everything, verifies the model download against a checksum, walks you through the two macOS permission prompts, and finishes with a self-diagnostic. Every install keeps itself updated automatically. There's a one-line uninstall too — owning your tools should include the freedom to delete them.
The lesson that actually matters
This isn't really a story about dictation. It's about the new build-vs-buy math.
A year ago, "just build it yourself" was terrible advice for a busy founder — the glue code would eat a week you didn't have. Today, an agent fleet writes, tests, debugs, and ships the glue code while you review outcomes and make taste decisions. The build column of the spreadsheet just got dramatically cheaper, and every SaaS line item on your P&L that's mostly "glue around an open model" deserves a fresh look.
That's the exact thesis AutomateScale is built on — owned infrastructure over rented workflows. Vox happens to be a dictation tool. The same playbook applies to your CRM automations, your content pipeline, your reporting stack.
Get it
Download Vox free → Works on any Apple Silicon Mac. Free, open source (MIT), no account, no telemetry. Source code on GitHub, full command reference at /vox-docs.
And if you're staring at your own stack of subscriptions wondering which ones an agent fleet could replace — that's literally what we do.