Holdkey

Don’t type it. Say it.

Hold Right Ctrl, talk, let go. The text lands wherever your cursor is: Claude Code, Cursor, ChatGPT, Gmail, Slack, your terminal. Three times faster than typing, a five-minute thought is never lost, and it all runs on your PC. Fully offline.

Get the launch emailFree. No account. Nothing leaves your PC.
You, holding Right Ctrl

…

Claude Code, a moment later
>
HoldkeyThe overlay you actually see. Latency shown is a real measurement on a desktop CPU.

How many hours would you get back?

You type at about 45 words a minute. You talk at 150. Move the sliders.

Assumes you speak at 150 words a minute, about 0.5 s of overhead per dictation, and 252 working days. Talking is not free; this is the difference, not the whole.

3.6 days
back in 12 months
21 minutes a day. 340,200 words you would say instead of type.
typing, 5.3 days total talking, 39 h total

Full calculator and how it is worked out →

~70 ms
from letting go of the key to text on screen, in Raw mode
0 bytes
of audio or text sent anywhere. Recognition and cleanup run on your CPU and GPU
1 key
to learn. No wake word, no app to focus, no menu
How these are measured →

Hold. Talk. Release. That is the whole product.

  1. 1

    Hold

    Press and hold Right Ctrl in any app. Holdkey listens only while the key is down; nothing else on your PC changes.

  2. 2

    Talk

    Say it the way you would say it out loud. While you pause, the words are already being recognized and cleaned in the background.

  3. 3

    Release

    Let go and the text is in the box under your cursor, usually before your hand is back on the keyboard. Tap again for the verbatim version.

Says what you meant

Clean mode keeps your words and drops the noise: ums, restarts, the sentence you took back. It never changes your meaning, and a tap of the key swaps in the verbatim version.

Knows your vocabulary

Add product names, libraries and coworkers once. Holdkey biases recognition itself, so Supabase stops coming out as “super base”.

Works wherever your cursor is

Including the places other tools drop the ball: the terminal, Claude Code, a password-manager field, the tiny reply box in Linear or Discord. If you can type there, you can hold a key there.

Questions people ask

Doesn't ChatGPT already have a mic button?

It does, and the ChatGPT mic only works inside the ChatGPT window. Holdkey works in every text input on your computer: Claude Code, the terminal, Gmail, Slack, Linear, the form you are filling out right now. One key, everywhere.

What happens on long thoughts?

If you have ever talked to ChatGPT for two minutes and watched the message vanish, you know. There is no undo for that. Holdkey transcribes as you pause, keeps every dictation in a local history, and a five-minute ramble lands as fast as a five-word one.

Do I have to click something and wait?

No clicking a button, no focusing a box, no spinner after you stop. Hold, talk, release. It works from a webcam mic on the other side of the desk.

Is anything sent to the cloud?

No. Recognition, cleanup and your vocabulary run on your PC with open models. It works with the Wi-Fi off. There is no account, and your history is a file on your disk you can open or delete.

Does Clean mode change what I said?

It removes ums, restarts and the sentence you took back, and fixes punctuation. Numbers, negations and who you are talking to are checked before anything is inserted; if the cleanup would change them, you get your exact words instead. A tap of the key swaps in the verbatim version.

Why not just build this myself with Whisper?

You can, and a weekend version works in one app. The months go into the parts you notice later: text landing correctly in terminals and password fields, cleanup that never flips a meaning, vocabulary biasing inside the recognizer, and release-to-text under a second on a plain CPU.

What does it cost?

Raw mode is free forever, no account. Clean mode is $8 a month or $69 a year. Founders get a lifetime licence for $99 while the first 150 seats last.

Mac?

Windows first. Leave your email below and you get one message when the Mac build lands.

Runs on your PC. Fully offline.

Speech recognition, cleanup and vocabulary run locally with open models. No account, no cloud, no screenshots, no usage tracking. Your dictation history lives in a file on your disk that you can open or delete.

  • Fully offline after the one-time model download: on a plane, on a locked-down network, with the Wi-Fi off
  • No per-word caps, no “free minutes”, no outages
  • Under 100 MB of memory while idle
  • Optional: bring your own API key for a cloud rewrite model

Pricing

Raw mode is free forever. Pay only for the cleanup.

Free
$0
  • Raw mode, unlimited
  • Clean mode, 30 dictations a day
  • Personal vocabulary, 20 terms
Pro
$8/month
or $69 a year
  • Everything unlimited
  • Per-app modes
  • Bring your own cloud model
  • Priority fixes
Founder lifetime
$99
first 150 seats, then gone
  • Pro, forever
  • Your name in the credits
  • Direct line to the founder

Get the launch email

The Windows build is in private testing. One email when the installer ships, and one when Mac lands. Nothing else.

Questions

Do I need a GPU?
No. Recognition runs on the CPU in real time. Clean mode uses a small local model that is faster on a GPU but works on CPU too.
Which languages?
English first. Twenty-five European languages are available with the multilingual recognizer; more are planned.
Does it work in the terminal?
Yes. Windows Terminal, PowerShell and Claude Code are first-class targets, not an afterthought.
Mac?
After Windows is solid. Leave your email above.
What happens to my audio?
It is processed in memory on your PC and discarded. You can optionally keep local recordings to tune recognition to your voice.
Why not just use the built-in Windows dictation?
Try both on a technical sentence in a terminal. Holdkey is faster, understands your vocabulary, and does not need the cloud.