Private by design
Your voice goes to one place: the computer you already trust with your code. The relay never stores it, the agent holds it only in memory until the text is ready, and neither ever logs the audio or the text.
New Voice input
Dictate to your coding agents.
Your own computer does the listening. Say your prompt instead of thumb-typing it: Mobaide records while you talk, your Windows, macOS or Linux machine turns the audio into text, and the text waits in the command field for you to check. Nothing is sent until you send it.
Free and opt-in. In the Android and iOS apps and in the browser, for Claude Code, Codex CLI, Antigravity CLI, a raw terminal or your own agent types.
Reading src/upload/client.py …
✓ 48 tests passed
What should I do next?Now add retries with backoffto the upload client andcover it with a test.
✓ Transcribed on macbook-pro
- 25 languages, detected automatically
- Runs on your CPU — no GPU
- Audio and text are never stored
- Free, and off until you turn it on
How it works
Your voice takes the same path as your keystrokes: from your phone or browser, through the Mobaide relay, to the agent on your own computer. The difference is that your computer writes it down for you first.
-
01
Tap the mic and talk
In the command field on Android and iOS, or in the session toolbar in the browser. A level meter shows it hears you; stop whenever you are done.
-
02
The relay passes it on
The audio streams in short chunks over TLS through the Mobaide relay, which routes it in memory to your agent and keeps nothing.
-
03
Your computer transcribes
The agent runs NVIDIA's open Parakeet speech model on its own CPU and turns the audio into text. No cloud speech service is involved.
-
04
You review, then send
The text appears in the command field — in a review box in the browser. Fix a word if you need to; nothing reaches the agent until you send it.
See it in the app
The same recording controls on every platform: the elapsed time, a level meter that moves with your voice, Stop to transcribe what you said and ✕ to throw it away.
Why on your own computer
Nothing sent behind your back
A misheard word should never become a command. The transcript always waits in the command field, so you read it, edit it, and send it yourself.
No GPU, no fuss
Transcription runs on the CPU at below-normal priority, so it gives way to your other work. It loads only when you dictate and frees its memory after three quiet minutes.
Speak your language
Twenty-five languages, and the model works out which one you are speaking — there is nothing to switch when you change language.
Every session kind
The transcript is plain text in the command field, so it works for Claude Code, Codex CLI, Antigravity CLI, a raw terminal and your own agent types alike.
Off until you want it
Voice input is opt-in per computer. Nothing downloads until you enable it, the app shows the cost first, and you can remove it again at any time.
Set it up
Turn it on once for each computer. Your agent already supports voice input if it has updated itself to version 0.3.17 or later — agents update on their own.
On Android
- Open a session and tap the mic in the command field — or go to Settings → Voice input.
- Check the download and memory cost, then tap Enable voice input.
- Allow the microphone the first time you dictate.
On iPhone and iPad
- Open a session and tap the mic in the message field — or go to Settings → Voice input.
- Check the download and memory cost, then tap Enable voice input.
- Allow the microphone the first time you dictate.
In the browser
- At app.mobaide.com, open Agents and choose Voice input on your computer's card (it must be online).
- Check the cost, then choose Enable voice input.
- In a session, press the mic in the toolbar and allow the microphone when the browser asks.
The download runs on your computer in the background, about 0.7 GB once. You can close the app meanwhile — the mic switches on by itself when voice input is ready.
What it needs
- Computer
- Windows, macOS or Linux on 64-bit x86 or ARM — the computer that already runs your Mobaide agent.
- Download
- About 0.7 GB, once: the speech model plus a small runtime. Agent updates never download the model again.
- Disk
- Up to about 730 MB. Removing voice input frees it.
- Memory
- None while idle. About 700 MB while a dictation is transcribed, released after three idle minutes.
- Recommended
- 12 GB of memory, 4 CPU cores and 2 GB of free disk. On a smaller computer the app warns you first, and you can still turn it on.
- Per dictation
- Up to two minutes. Dictate again to add more.
25 languages
- Bulgarian
- Croatian
- Czech
- Danish
- Dutch
- English
- Estonian
- Finnish
- French
- German
- Greek
- Hungarian
- Italian
- Latvian
- Lithuanian
- Maltese
- Polish
- Portuguese
- Romanian
- Russian
- Slovak
- Slovenian
- Spanish
- Swedish
- Ukrainian
The model detects the language as you speak. Japanese, Chinese, Korean, Arabic, Hindi and Turkish are not supported yet.
Where your voice goes
- The microphone records only while you dictate, and only with Mobaide on screen.
- Going through the relay, the audio is TLS-encrypted on every hop and goes only to your own agent.
- The relay routes it in memory and keeps neither the audio nor the text.
- Your agent holds the audio in memory until the text is ready, then drops it. It never logs or saves the audio or the text.
- No third-party speech service is involved. The only extra downloads are the model and its runtime, once, from GitHub (Hugging Face as a fallback) and PyPI.
The privacy policy describes voice input in full. The model is NVIDIA Parakeet TDT 0.6B v3, released under CC BY 4.0; the apps show its license in the voice input settings.
Questions
Is voice input free?
Yes. It is part of Mobaide, with no subscription and no per-minute fee — the transcription runs on hardware you already own.
Will it send my prompt by itself?
No. The transcript always waits in the command field — or in the browser's review box — until you send it. In the browser you can tick Insert dictated text directly to skip the review box; the text still lands in the session's input without pressing Enter.
How long can I talk?
Up to two minutes per dictation. The last 15 seconds count down, then it stops and transcribes what you said. Dictate again to add more.
What if my connection drops or I switch apps?
A shaky connection does not stop the recording: the audio is kept in memory and re-sent when the connection comes back, and if a dictation still fails, Retry reuses the recording — you never have to say it twice. Leaving the app, hiding the browser tab or taking a call stops the recording and transcribes what it captured so far.
Why don't I see the mic?
The agent must be online and at version 0.3.17 or later — it updates itself. A computer the voice model cannot run on (for example a 32-bit Python) shows no mic. If you chose Hide the mic for this agent, turn it back on in the agent's voice input settings.
Can I still use my keyboard's dictation?
Yes. The command field is an ordinary text field, so your phone's keyboard dictation keeps working. Mobaide's voice input is the option that keeps your audio between your device and your own computer, and it works the same way in a desktop browser.
How do I turn it off or remove it?
Open Settings → Voice input in the app, or the agent's Voice input settings in the browser, and choose Remove. That deletes the model and frees the disk space. Removing the agent from the app with Also uninstall Mobaide from this computer deletes it too.
Can I set it up on the computer itself?
Yes, which helps on a headless machine. Run this as the user the agent runs as — from an elevated prompt on Windows:
# on the agent's computer
mobaide-agent voice install
The apps pick up the result on their own. mobaide-agent voice
status shows the state and the cost.
Try it on your next prompt
New to Mobaide? Install the agent on your computer, pair your phone with one QR scan, then turn on voice input.
