Hold the key to see what happens to a real ramble.
I think we should move the meeting to Friday at 3 pm, and also can you send me the report by end of day?
Word for word what the model wrote.
Hold a key, talk the way you talk, let go. Clean text lands wherever your cursor is. It runs on your Mac, free, with no word limit.
Apple silicon · macOS 14 or later
Same words. Scroll to clean them.
This is what you actually said.
It doesn’t just write down the sounds you made. Change your mind mid-sentence and it goes with the last thing you said. Count things out loud and you get a list. Say “best regards” and you get an email.
tell rahul the deploy is at nine no wait ten thirty and he should ping me on slack not email
Tell Rahul the deploy is at ten thirty and he should ping me on Slack, not email.
Real output from Ramblebee’s own cleanup model, trained for this one job, running on an M1 Pro MacBook. On its 8 standard test sentences it took about 0.3 seconds each.
You already know what you want to say. This is the same paragraph at both speeds, in real time.
Typical rates, and yours will differ. The app’s Home screen shows your own pace. On phones, a Stanford study measured speech at about 3× the speed of a keyboard.
No plugins and no copying and pasting. It writes where your cursor already is, and lays things out to suit: a greeting and sign-off in an email, a plain line in a chat.
Hi Sarah, Just following up on the invoice from last week, could you let me know when it will be paid? Thanks, Sam
Staging is up. Anyone looked at CI yet?
Not me, I’m in the design review.
I found the deliver function in AppModel.swift. What would you like to change?
My priorities this week: 1. Finish the onboarding flow. 2. Fix the crash on launch. 3. Write the release notes.
Are we still on for today?
Yes, after the deploy
When’s the deploy?
• Buy milk. • Call mom. • Pick up the dry cleaning.
Illustrations of each app. The words being typed are real output from our model. Names and logos belong to their owners.
Your words, your call
Private on your Mac by default, or a cloud model with your own key if that suits you better. Either way, nothing goes through us, because we don’t run a server.
Speech recognition and cleanup both run on your Mac. Want proof? Turn off your Wi-Fi and keep talking.
It goes online once to download its models, about 1.6 GB.
Your voice is understood right here. Nothing leaves your Mac.
Premium · Command Mode
Select some text and say what you want done to it. With nothing selected, it can change what you just dictated, answer the conversation in the window, or write something new at your cursor. Ask it a question and the answer comes up in a card, to read, copy or insert.
Hey team, so I just wanted to quickly let everyone know that the release is probably going to move to Friday, because QA found two bugs that we really need to fix first, and I will send out the notes tomorrow morning once we know more.
Word for word what the local model wrote, running on an M1 Pro MacBook in under a second.
Premium · Meeting notes Beta
When a call starts, it asks if you’d like notes. It hears your microphone and your Mac’s sound, so nobody joins the call, and it tells the people on the call apart by voice, so you can put names to them. A dot in the pill shows it’s listening. It stays light on your Mac during the call, and writes everything up once the call ends. It’s in beta: it works, and we’re still improving it.
They went through the launch dates. The final copy for the pricing page and a question from legal about the trial are still open, and the reminder email is to be ready by Thursday.
A sample meeting, laid out as the app shows it.
A summary, key points, decisions, and action points with who said they’d do them and when, plus a title. Stand-ups, one-on-ones and interviews get their own layout. It’s a draft from a model on your Mac, so check what matters against the transcript.
Type a question about the meeting and get the answer from the transcript, with the time to look at. When the transcript doesn’t say, neither does it.
It notices calls in Teams, Zoom, Webex, Slack, FaceTime and Discord, and Google Meet, Whereby or Jitsi in your browser. Any other app that keeps your microphone for 45 seconds counts too. Notes stop when the call does.
Meetings that mix English and Hindi come out in Roman letters, so the whole transcript stays readable and searchable.
Fix a name after dictating and it offers to remember it.
Say a short cue, get the long text.
Nowhere to type? They wait in a card, ready to copy.
Type or talk, in Markdown, with one shortcut.
Natural, concise, professional or casual, set per app.
Tap twice and keep talking, with no key to hold down.
Say a file or function name and it’s spelled the way it is in your project.
Say “scratch that” to drop a sentence. Cancelled by mistake? Undo brings it back.
Words, time saved against typing, your speaking pace and your streak. Share a card of it, made on your Mac, with none of your words on it.
Spaces, capitals and full stops fit the text around your cursor.
It pastes your words, then puts back whatever you’d copied.
It runs on your Mac, so your words cost us nothing. That’s why there’s no word limit. Premium is for people who want it to do more.
$0 for good
No card and no account.
$4 a month
Stop whenever you like.
7-day trial with a card. Charged when it ends unless you cancel first. See refunds.
Not by default. “Keep everything on this Mac” starts switched on, and it blocks every cloud request. If you add your own API key and choose a cloud engine, what you say goes straight from your Mac to that provider, under their terms, and the app shows you that it’s doing so. It never goes through us, because we don’t run a server.
Yes, if you want a cloud model. Add a key from OpenAI, Groq, Gemini, Anthropic or DeepSeek, or point it at any OpenAI-compatible server, such as Ollama or OpenRouter. Speech and cleanup can each use a different one, and you can choose per app. Keys stay in your Mac’s Keychain, and the provider bills you directly.
Only to download its models. Dictation needs about 1.6 GB, once. Command Mode and meeting notes use a larger model, about 2.5 GB more, if you want them. After that, turn off your Wi-Fi and it carries on working.
Open models, in a pipeline of our own. Express, Universal or Lite hear you, and Standard or Advanced clean up. Standard is our own model, trained for dictation. The others are published by NVIDIA, OpenAI and the Qwen team, and each one is credited with its licence in the app and on our Credits page. You can see each model’s size, and remove it, in Settings › Storage.
Macs with Apple silicon (M1 or later) running macOS 14 or newer. Meeting notes need macOS 14.2 or newer. Intel Macs, Windows and phones aren’t supported today.
Dictation is built for English. Meeting notes handle English, and also English mixed with Hindi, which is written in Roman letters.
Anywhere your cursor can type: mail, chat, documents, code editors, terminals and browsers. If a field can’t take the text, your words stay on screen in a card, ready to copy, so you never lose them. Password fields are never read or typed into.
macOS requires it for an app to type into other apps and to notice your shortcut. The microphone permission is for hearing you. Your audio is never saved.
Yes. Free dictation has no word limit, no time limit and no account, because it runs on your own Mac and costs us nothing per word.
You get 7 days of Premium with a card on file. It’s charged when the 7 days end, unless you cancel first. If you never upgrade, free dictation keeps working.
They work, and we’re still improving them with every release, so expect the odd rough edge. Keep your own notes for anything critical until you trust them.
No. They listen to your microphone and your Mac’s sound, so nobody joins the call. Summaries are drafts from a small model on your Mac, so check what matters against the transcript. Recording laws differ by place, so tell people you’re taking notes.
Apple silicon · macOS 14 or later. On your phone? We’ll email you the link for your Mac.