Dimmy: voice dictation, command mode, and AI meeting recaps for macOS and Windows

v0.6.65 · now on mac & windows · 4 stars on github
Detto, fatto. said, done.

A voice power tool for macOS and Windows. Dictate clean text anywhere, speak commands that edit in place, and turn meetings into structured recaps. Offline by default.

HoldSpaceor tap the pill below. Try it right here.
~/work/draft.txtready
Hold Space and say something.e.g. "draft a reply saying we're aligned and want to ship by Friday."
offline·whisper-large-v3-turbo·78 mb·~0.4s round-trip on m1
Get notified when Linux shipslinux build · join the waitlist
free forever·open source·opt-out telemetry·no account·~12 mb installer
3
voice superpowers
~0.4s
dictation round-trip
13
rewrite styles
100+
apps it works in
scroll to explore
works everywhere

If it has a text input, Dimmy works in it.

Slack
Notion
VS Code
Figma
Discord
Linear
Gmail
Cursor
Terminal
Obsidian
ChatGPT
Claude
Microsoft Word
Google Docs
Telegram
Whatsapp
Things
Bear
Xcode
Safari
Chrome
Arc
Mail
iMessage
Zed
Sublime
JetBrains
Slack
Notion
VS Code
Figma
Discord
Linear
Gmail
Cursor
Terminal
Obsidian
ChatGPT
Claude
Microsoft Word
Google Docs
Telegram
Whatsapp
Things
Bear
Xcode
Safari
Chrome
Arc
Mail
iMessage
Zed
Sublime
JetBrains
one voice · three ways to use it

Your voice, made powerful.

Most apps stop at turning speech into text. Dimmy treats your voice as an input you can do real work with: dictate it, command with it, or capture a whole meeting and keep the recap.

PILLAR 01
Dictation
Speak, and clean text lands where your cursor is.
Two passes, both local by default: one catches every word, the other polishes punctuation, grammar, and tone.
"hey can you draft a quick reply"Hey, can you draft a quick reply.
see how↓
PILLAR 02
Command Mode
Say what to do. It happens in place.
Select text and speak an instruction. Dimmy reads it, applies your request, and replaces the selection. Nothing selected? Your words become a writing request at the cursor.
"make it more formal"Would it be possible to move the deadline?
see how↓
PILLAR 03
Meetings
Record the call. Keep the recap.
Captures both sides, transcribes live, and writes a structured recap: decisions, action items, open questions. Audio and history stay searchable, and an MCP bridge hands it all to your AI tools.
47-minute planning callTL;DR · decisions · action items
see how↓
02live demo

Press, speak, release.
That's the whole UX.

One icon, one status dot, a right-click menu. That's all the UI you ever need to see. Watch it cycle live.

DimmyFileEditViewWindowHelpFri 1 May 20:21
ready
Tasks
Toggle pill
Open Settings…
Quit Dimmy
Style
Off
Correct
Summarize
Elaborate
Comprehensible
Professional
Prompt
Gen Z
Boomer
Emoji
Acronyms
Imbruttito
Custom
Translate to
No translation
🇬🇧English
🇮🇹Italiano
🇫🇷Français
🇩🇪Deutsch
Dimmy (Dev)
Hide menubar icon
real ui · pure react · no video·auto-cycling: idle → rec → transcribe → done·click the icon to toggle the menu
PILLAR 01Dictationspeak → clean text
01how it works

From breath to pasted text
in less than half a second.

Four stages. Three of them happen on your machine. One is optional. All of them are visible in the pill.

10ms
Capture
your microphone
16kHz mono, VAD trims silence in real-time. Audio buffer never touches disk.
2~120ms
Transcribe
whisper-large-v3-turbo
Runs locally on CPU/GPU via ggml. 78 MB quantized model, ~5× realtime on M1.
draftaquickreply
3~250ms
Rewrite
Gemma 4 · on your machine
Google's open-weights LLM. Apache 2.0, ~3 GB, no API key, no network. Pick a style: correct, professional, summarize. Skip for raw.
professional→✦ ✦ ✦
4~10ms
Paste
into the focused window
Clipboard simulation via system APIs. Slack, Notion, VS Code, anywhere.
Hey, can you draft▌
⌃⌥ pressed → text in your editor. ~380ms median on m1, fully offline.
two passes · both local

Transcribe.
Then enhance.

Two models work back-to-back on your laptop. One catches every word, the other makes it read like you wrote it. No API key, no login, no audio leaving your machine.

PASS 1
Transcribe
Catch every word, in any of 99 languages.
OfflineSub-second99 languages
raw transcript
"hey can you draft a quick reply saying we are aligned and we want to ship by friday"
powered by Whisper Large v3 Turbo
PASS 2
Enhance
Polish the punctuation, fix the grammar, switch the tone.
13 stylesNo API keyYours to keep
enhanced output
"Hey, can you draft a quick reply saying we're aligned and want to ship by Friday."
powered by Google Gemma 4 · runs offline
on your machine0 cloud calls0 API keys requiredworks on a plane
03speed

Faster than
typing you.

You speak at ~150 words per minute. You type at ~40. Dimmy closes that gap with sub-second latency, so the bottleneck is your thinking, not your fingers.

~380msmedian round-trip on M1, fully offline
~120mstranscription with whisper-large-v3-turbo
~60 fpspill animation on a 2018 ThinkPad
~12 MBinstaller size · ~78 MB whisper model
live race
Ktyping~2 wpm0.07s
H
Ddimmy~150 wpm equivalent0.00s
Hey team, I think we should ship the new pricing page on Friday.
⚡Dimmy finished in 0.38s. Typing is still going at 0.07s.
0413 styles

One sentence. Thirteen tones.

Pick a style before you press the hotkey. Dimmy will rewrite your speech in the tone you chose, including imbruttito, our love letter to Milanese grumpy.

you said
"ok so thursday afternoon design review with the team sound good"
Professionalpolished, business-appropriate tone

I'd like to propose a design review on Thursday afternoon with t

auto-cycling. Click any style to lock.
PILLAR 02Command Modespeak → action
02command mode

Say what to do. It does it in place.

Flip the pill into command mode and your voice stops being text and starts being an instruction. Select something, speak a change, and Dimmy rewrites it right where it sits. No copy-paste into a chatbot and back.

any app · text selectedready
hey can we push the deadline a couple days
“make it more formal”
Nothing selected? Just ask.
With no selection, your words become a writing request inserted at the cursor: “draft a Slack message asking for the Q3 numbers.”
One toggle, or one hotkey
Make it sticky from the pill's right-click menu (amber dot), or fire a single command with a dedicated one-shot shortcut.
Same model, no extra key
Runs on the same local model as the AI styles, or your Claude / ChatGPT subscription. Windows and macOS.
PILLAR 03Meetingsrecord → recap
03meetings

Record the call. Keep the recap.

A dedicated Dimmy Meeting window records both sides of a call, transcribes it live, and the moment you stop, writes a structured recap. The audio and the transcript stay yours, on your machine, searchable forever.

Planning sync recap
47:12 · auto-generated on stop
planning
TL;DR
›Agreed to ship the billing rewrite by Apr 30; mobile slips to Q3.
Key decisions
›Drop the legacy importer, nobody's used it since Jan. (12:04)
›Stripe stays the only processor for launch. (28:41)
Action items
›Marco: wire the new webhook + tests. (31:10)
›Sara: draft the migration note for users. (39:55)
Open questions
›Do annual plans get grandfathered pricing? (44:20)
Both sides of the call
Microphone + system audio, so you capture everyone, not just yourself.
Live, in ~15s chunks
Transcribes as you talk, with timestamped notes and pause/resume.
Tuned to the meeting
A type classifier (1:1, standup, planning, interview…) shapes which sections matter.
Consent, up front
A recording-consent gate every session, with all-party-consent regions flagged.
Audio + history kept
Every meeting is saved and full-text searchable. Re-open the recap or the recording anytime.
send it anywhere
Notion
Recaps as real Notion pages, optionally auto-sent.
A folder
Point it at Obsidian, Drive, or Dropbox for free sync.
Claude Desktop
Send a recap straight into a Claude conversation.
MCP bridge
dimmy-mcp lets Claude list, read, search, and save recaps.
05one click away

Right-click the tray.
Switch styles.

Drafting an email? Email. Notes for yourself? Off. Voice memo to flatten into bullets? Bullets. The tray shows your favorites. 13 styles total, all configurable.

🇬🇧🇮🇹🇫🇷🇩🇪🇪🇸🇮🇹+ napoletano
Translate while you dictate.
Speak Italian. Output English. Or napoletano, perché no.
Tasks
Toggle pill
Open Settings…
Quit Dimmy
Style
Off
Correct
Summarize
Elaborate
Comprehensible
Professional
Prompt
Gen Z
Boomer
Emoji
Acronyms
Imbruttito
Custom
Translate to
No translation
🇬🇧English
🇮🇹Italiano
🇫🇷Français
🇩🇪Deutsch
Dimmy (Dev)
Unpin from taskbar
Close window
← five styles
  one click
06no pill, no problem

Hide the pill.
The tray icon takes over.

You're presenting on a Zoom call. You're recording a Loom. Or you just don't want a colored pill floating at the edge of your screen all day. Toggle it off. The tray icon does everything the pill did, quietly.

Windows · taskbar
macOS · menu bar
Fri 1 May 14:32
auto-cycling · ready · nothing on screen
Screenshare-safe
Nothing floats over your slides. Recording is invisible to viewers, but the tray tells you it's on.
Fullscreen-friendly
Games, video editors, presentations. Apps that hate floating overlays still get dictation.
Distraction-free
Some users just don't want a pill in their peripheral vision. The tray icon is enough.
07it's a real app

Settings that
respect your attention.

No 7-step onboarding. No account. Search, click, change. Done. Click around. It's the real thing.

live prototype · click anywhere · theme follows the page
Live preview
See pill states before applying
9-grid position
Snap anywhere, drag to fine-tune
Border styles
Rainbow, solid, breathing, off
Per-app rules
Different shortcut, different style
08two minutes to setup

Pick your engine.
Cloud-fast or fully offline.

Groq's free tier is genuinely the fastest cloud STT we've benchmarked. Whisper local is fully private and works on a plane. Both are first-class.

1
Sign up at console.groq.com
No credit card. Free tier covers ~100k tokens/day, way more than you'll use.
🔒 console.groq.com
Welcome to Groq
The fastest LLM inference on the planet.
Sign in with Google
2
Create an API key
🔒 console.groq.com/keys
API Keys
namedimmy+ Create key
gsk_•••••••••••••••••••••••••••••••• 7Hk2
3
Paste it into Dimmy
Settings → Transcription → Cloud (optional) → Groq API key. Encrypted with the OS keychain. Done.
⚡
That's it. Round-trip latency drops to ~180ms. Your machine does zero compute.
09privacy

Your voice doesn't
leave your machine.

Default mode is fully offline. The cloud is opt-in, explicit, and per-provider. We don't know who you are. We'd like to keep it that way.

0
telemetry, if you opt out
anonymous counts only. flip one toggle and we get nothing, ever.
0
accounts created
no signup, no email, no waitlist
0
audio bytes saved to disk
ring buffer in RAM, then gone
AES-256-GCM
key encryption
if you connect a cloud STT, your keys live in the OS keychain
$ sudo lsof -i -P | grep -i dimmy
# nothing. as expected.
Read the full privacy policy# see what we collect
10pricing

Honest pricing.
No dark patterns.

The code is public. If you can compile, you don't owe us anything. Paying gets you the signed binary, one-click updates, early access, and a human on email. Same software, different distribution.

Free
€0forever

Clone the repo, compile it yourself. Full app, no paywall.

Get the source
✓Same code as Pro. Zero feature gates.
✓You compile, sign, and notarize yourself
✓Local Whisper transcription, all 13 styles
✓Bring your own API key (Groq, OpenAI, Deepgram, Gemini)
✓macOS, Windows
✓Community support on GitHub Issues
most popular
Pro
€39/ yearsave 35%

We compile it for you. You get updates, early access, and a human on email.

Start 14-day trial
✓Pre-built binary, signed and notarized
✓Auto-update inside the app, one click
✓Early access to beta builds before public release
✓Email support · a human replies in <24h
✓Vote on the public roadmap
✓Cancel anytime, 30-day refund
buy once · own forever
Lifetime
€99once

Pay once. Every release for 3 years from purchase. No subscription, no upsell.

Buy lifetime
✓Everything in Pro, forever · no renewals
✓3 years of updates from purchase
✓Prebuilt for macOS · Windows
✓Priority email support · founder replies
✓Your name in the credits if you want
Three years of updates, paid once. No renewals to forget, no subscription anxiety.
all plans · 30-day refund · cancel anytime · students 40% off
11manifesto

We built Dimmy
because we were tired.

Tired of voice apps that take 8 seconds to wake up, weigh 200 MB, and want a subscription to do something a hotkey should do. Tired of giving our voice to a server we'll never see.

So we wrote it native. SwiftUI on macOS, WinUI 3 on Windows, a shared Rust core for the parts that matter. The whole thing is 12 MB, boots in 80ms, and the source is public on GitHub.

Audio never touches disk. Audio never leaves your machine, unless you turn on a cloud provider. In that case your key sleeps in the OS keychain, and you can revoke it whenever.

Made with espresso. /ˈdɪmmi/: "tell me", in Italian.

A team of nerds in Italy
native desktop apps, pair-built with Claude Code
♥ Sponsor on GitHub
12roadmap

Mobile when you
want it badly enough.

We can't ship five platforms in parallel. Vote and we'll prioritize. The first platform to hit 1,000 ★ on GitHub gets built next.

iO
iOS
system-wide dictation replacement
An
Android
gboard alternative + system input
Ap
Apple Watch
tap, speak, send to phone
vi
visionOS
the obvious one
votes are stored in your browser, not on our server (because we don't have one)
13frequently asked

Yes, but really, how?

Yes, if you compile it yourself. The whole code is public on GitHub and builds on macOS and Windows. Pro and Lifetime sell you the convenience: a signed pre-built binary, one-click in-app updates, early access before public releases, and a human on email. Same software, different distribution.
We ship Whisper Large v3 Turbo, quantized to 78 MB, running on ggml. On an M1 it's ~5× realtime; on a 2018 ThinkPad it's still under 500ms for a typical sentence.
Default mode is fully offline. Your audio never leaves your machine and is never written to disk. If you connect a cloud provider (Groq, OpenAI, Deepgram, Gemini), it's opt-in per provider, and your API key is encrypted in the OS keychain with AES-256-GCM.
Because we wrote it native, one platform at a time, with a shared Rust core for the parts that matter. A 12 MB binary boots in 80ms and runs smoothly on a 2014 laptop. The harder path was the right path.
Mobile is the most-requested feature by a mile. We'll ship the platform that crosses 1,000 ★ on GitHub first. Vote up above.
All 99 that Whisper supports. It auto-detects, and you can mix mid-sentence: English and Italian, English and Mandarin, whatever you actually speak.
Yes. The 'Prompt' style structures speech as an LLM prompt; the 'Comprehensible' and 'Professional' styles work great for commit messages and doc strings. People mostly use it for Slack, email, and ChatGPT though.
Instead of typing what you say, your speech becomes an instruction Dimmy runs in place. Select some text, say "make it more formal" or "translate to English", and it rewrites the selection right where it sits. With nothing selected, your words become a writing request inserted at the cursor. Same local model as the AI styles (or your Claude / ChatGPT subscription), on Windows and macOS.
Yes. The Dimmy Meeting window records both sides of a call (microphone + system audio), transcribes live, and the moment you stop writes a structured recap: TL;DR, key decisions, action items, open questions, with timestamps. The audio and transcript stay on your machine and are full-text searchable. Export a recap to Notion, a synced folder (Obsidian, Drive, Dropbox), or Claude Desktop, and the dimmy-mcp bridge lets Claude list, read, search, and save recaps.
Short for 'dimmi', Italian for 'tell me'. We're Italian. We thought it was cute. We're aware that's a bias.

Dimmy.

Italian for "tell me."

Get notified when Linux shipslinux build · join the waitlist