How Rubber Duck Transcribes With Zero External API Calls (100% On-Device)
Quick answer: Rubber Duck transcribes your speech entirely on your Mac using on-device Whisper, so no audio is ever uploaded. The only optional network call is note cleanup, and it runs through your own Claude Code or Codex subscription rather than a server we run. Turn cleanup off and the app makes no external calls at all.
Most dictation and meeting tools follow the same path: your microphone audio is streamed to a company's servers, transcribed in their cloud, and stored in their account system. It works, but it means the most unfiltered thing you produce all day, the sound of you thinking out loud, lives on someone else's computer.
Rubber Duck is built the other way around. Here is exactly what happens to your voice, and what does not.
What "on-device" actually means
When you hold the dictation key and talk, the audio is captured and run through a Whisper speech-to-text model that lives inside the app and executes on your Mac's own chip. The Apple Silicon Neural Engine does the heavy lifting, which is why it is fast and easy on the battery. There is no round trip to a data center, because the model that turns sound into words is already on the machine in front of you.
That is the whole difference. "On-device" is not a privacy slogan bolted onto a cloud service. It is the actual architecture: the transcription happens locally because the model is local.
Where your audio goes: nowhere
The recording exists only long enough to be turned into text, and it stays in memory on your Mac the entire time. It is not written to a cloud bucket, not queued for a server, and not attached to an account. When the transcription is done, you have text; the audio was never anywhere but your own hardware.
This matters most for the things people actually dictate: half-formed product ideas, candid notes about a coworker, a client's confidential numbers, a doctor's-appointment summary. None of it needs to leave your laptop for the app to do its job.
The one optional network call, and how to turn it off
There is exactly one point where Rubber Duck can touch the network, and it is opt-in: cleanup. After transcription, you can have an AI tidy the raw text, fix punctuation, add a title, and tag the note. That step sends the already-transcribed text (never the audio) to an AI model.
Two things make this different from a normal cloud app:
- - It is optional. Turn cleanup off and the transcript is pasted and saved exactly as spoken, with zero external calls. The app works completely offline in this mode.
- - When it is on, the call goes through your AI subscription, not ours.
Bring your own AI: why you are not paying an API bill
Rubber Duck does not run an AI service and does not resell tokens. Instead, it uses the Claude Code or Codex subscription you already pay for. Your credentials, your account, your usage.
This has a few honest consequences. You are not funding a markup on every word. There is no separate metered API charge from us that balloons if you dictate all day. And because the cleanup runs against a developer AI subscription you control, you can see and reason about where that text goes, rather than trusting an opaque pipeline.
If you would rather keep everything local, don't connect any AI at all. You still get accurate on-device transcripts and organized Markdown notes.
Why this matters for meetings and sensitive work
Meeting capture is where cloud transcription gets uncomfortable. Recording a call usually means uploading everyone's voice to a vendor. With on-device transcription, you can capture a meeting, get a full searchable transcript, and never forget what was said, without that audio ever leaving your Mac. For anyone handling client data, health information, legal discussion, or unreleased plans, that is not a nice-to-have. It is the difference between "allowed" and "not allowed."
The tradeoffs, honestly
On-device is not magic, and it is worth being straight about the costs.
- - The best models are large. To keep the app download small, Rubber Duck ships a compact model and lets you download a higher-accuracy one once, on demand.
- - Your Mac does the work, so a very long meeting uses your CPU and battery rather than a server farm's.
- - Cleanup quality depends on the AI subscription you bring.
For most people, those tradeoffs are easy. You get speed, privacy, and no surprise bill, in exchange for a one-time model download and using your own hardware. That is the deal Rubber Duck is built on: your voice, your notes, your machine.
Frequently asked questions
Does Rubber Duck upload my audio anywhere?
No. Audio is captured and transcribed on your Mac and is never sent to any server. The raw recording stays in memory long enough to produce text and is not uploaded.
Do I need to pay for an API to use Rubber Duck?
No. Transcription is free and on-device. The optional cleanup step uses the Claude Code or Codex subscription you already have, so there is no separate per-word API bill from us.
Can I use Rubber Duck fully offline?
Yes. With cleanup turned off, dictation runs with no internet connection at all. A transcription model ships inside the app, and better models download once when you choose them.
Is my transcript stored online?
No. Notes are saved as plain Markdown files in a folder you choose on your own Mac. Nothing is stored on a server we control.
Think out loud. Rubber Duck writes it down.
On-device transcription that files your ideas and meetings as searchable notes.
Get early access