Stop typing your prompts.Say them.
Tap a key and talk. A whole paragraph of prompt lands in your agent in the time typing would get you a sentence, and the agent can read your transcripts back over MCP. Your voice never leaves your Mac, because there is nowhere for it to go.
$29 once. macOS on Apple Silicon. Works offline.
Works with Claude Code · Cursor · Codex · Copilot
Claude Code, Cursor, and Codex connect over MCP and can search your transcripts, edit your dictionary, and send files to transcribe.
Full transcription on your Mac, in seconds, not minutes in the cloud.
Your audio never leaves the machine. No account, no server, no cloud.
153 words a minute spoken against 52 typed, in the study everyone quotes. On a full keyboard the gap is smaller and still worth having.
Everything built around your voice and your agent.
One key you can tap, hold, or rebind, a real on-device model, and a set of pieces that make voice the fastest way to drive a coding agent.
Your agent can search what you said
A local MCP server lets Claude Code or Cursor read your transcripts, search them, send in recordings to transcribe, and keep your dictionary current. Very few dictation apps ship an MCP server at all, and it is the piece that turns dictation into something your agent can operate.
Fast enough to forget it’s running
Dictation starts the instant you press the key, the model already warm. A 60-minute recording transcribes in about 24 seconds. Nothing ever waits on a server.
Real ASR on the Neural Engine
NVIDIA’s Parakeet v3 runs locally on Apple Silicon: accurate, punctuated, and quick enough that the text lands the moment you stop. It dictates in 25 European languages and works out which one you’re speaking on its own. Not a byte of audio ever leaves the machine.
Drop a file, get subtitles
Drag in audio or video and export SRT, VTT, or plain text in seconds. The recording stays on your Mac the whole time.
It spells your jargon your way
A plain CSV of fixes applied right before the text lands: useEffect, GitHub, your repo’s own names. No model second-guessing your code.
Every transcript, searchable
Everything you dictate stays on this Mac. Search it, copy it, export it. Your agent can read it over MCP. Now so can you.
What it does, in detail.
A handful of pieces, each doing one job. Nothing here is trying to be clever.
It types into whatever has focus
It uses a normal Cmd-V, so text lands wherever your cursor is, agent prompt, editor, any text field, and multi-line stays in one piece. If you’ve clicked away to another window since you started talking, Edict leaves the text on your clipboard instead of pasting it into the wrong place.
It never rewords what you say
You get your exact words, plus your own short list of spelling fixes. The agent you’re talking to already cleans things up, so a second model rewriting your text would only add lag and second-guess your identifiers. Even “scratch that” stays in, because the agent needs it.
It spells your jargon the way you do
Keep a short list of fixes: say “use effect” and it types useEffect, “git hub” becomes GitHub. It’s a plain CSV, applied right before the text lands. No model rewriting your code behind your back.
Your agent keeps the word list current
The fix list is a plain file your agent can read and edit. One button copies a prompt that tells Claude Code or Cursor to add to it as you go, so the list grows with your codebase.
use effect,useEffect fast api,FastAPI qbo client,QBOClient
last row added by your agent
Your agent can read what you said.
Edict runs a local MCP server. Your coding agent connects to it and can search your transcripts, pull the last one, and keep your dictionary current as your codebase grows. Your words, reachable by your own tools, on your own machine.
“What did I say about the auth refactor?” The agent queries your transcripts and answers from your own words.
One prompt and the agent keeps your spelling rules current as new names enter the codebase.
The agent can hand Edict a recording and get the transcript back: “transcribe this interview and pull out the action items.” All of it on your Mac.
One of the very few dictation apps that ships its own MCP server, and it runs locally: no account, no cloud.
Made for people who drive coding agents more than they type.
Built for builders.
You drive agents all day
You’ve got Claude Code going and you’d rather say what you want than type a paragraph into the prompt box.
Claude Code and Cursor users
Keyboard for the precise edits, voice for everything else. Describe the change, read the diff, fix it, repeat.
People running a few agents at once
Two or three going in parallel on Claude Code, Cursor, or Codex. Talk to whichever one has focus.
Anyone whose wrists are done
Years at a keyboard add up. Dictation that lands where you’re looking takes some of the load off.
Four steps. One key.
No ceremony. Press a key, talk, and it types. That’s the whole thing.
Tap Right Command
One tap starts it. The model is already loaded, so it’s listening right away. Rebind this to any key you like.
Talk
A small pill shows a live waveform while it listens. Pauses are fine, it keeps going until you stop.
Tap to stop
Tap again and it finishes the transcript. Or switch it to hold-to-talk and just let go of the key.
It types into your app
The text lands where your cursor is, through a normal paste. Whatever you had on the clipboard gets put back.
Drop in a recording. Get subtitles in seconds.
Drag any audio or video file onto Edict and it transcribes on your Mac, labeling who said what. A 60-minute video takes about 24 seconds to transcribe. Export it as SRT or VTT subtitles, or plain text. The file never leaves the machine, so a standup, a product call, or a confidential interview stays private.
About 24 seconds to transcribe an hour of video, around 150x real time on an M1 Pro. Newer chips go faster.
Speaker detection runs on your Mac too. The transcript comes back labeled by voice, ready for interviews, meetings, and podcasts.
NVIDIA’s Parakeet v3, from the model family that has topped Hugging Face’s open ASR leaderboard, with pyannote’s community model picking out the speakers.
SRT and VTT with sentence-aware cues and balanced two-line wrapping, ready to drop straight into an editor.
The same on-device engine as live dictation. No upload, no account, no per-minute cloud bill.
Your voice never leaves your Mac.
Speech-to-text happens on-device, on the Neural Engine, using speech models bundled inside the app. Your voice and your history stay on this Mac. Only the text you choose to paste goes anywhere.
The speech models ship inside the app, so dictation and file transcription run entirely on your Mac from the first launch, offline included. There’s no account, no cloud compute, and nothing you say is sent anywhere.
The app is signed and notarized by Apple, so macOS opens it as an identified developer.
lsof -i -c Edictnettop -p $(pgrep Edict)Most dictation apps type into anything. This one is built for coding agents.
Plenty of dictation tools are good. Here is how they line up on the things that matter once the text has to land inside, and feed back to, a coding agent.
| Edict | Wispr Flow | superwhisper | macOS dictation | Spokenly | |
|---|---|---|---|---|---|
| Your agent can search your transcript history over MCP | ✓yes | ✕no | ✕no | ✕no | ✕no |
| Your agent can edit the word list | ✓yes | ✕no | ✕no | ✕no | ✕no |
| Your agent can send it files to transcribe | ✓yes | ✕no | ✕no | ✕no | ✕no |
| Won't paste into the wrong window | ✓yes | ~partial | ~partial | ✕no | ~partial |
| Audio stays on the device | ✓yes | ✕no | ~partial | ~partial | ~partial |
| Makes no network calls while you use it | ✓yes | ✕no | ~partial | ~partial | ~partial |
| Transcribe a file to SRT / VTT on-device | ✓yes | ✕no | ~partial | ✕no | ~partial |
| Pay once, no subscription | ✓yes | ✕no | ~partial | ✓yes | ~partial |
A “~” means partial, unstated, or not documented by the vendor. Everything else here is checkable. If we’ve got something wrong, tell us and we’ll fix it.
If you mostly talk to a coding agent, Edict is the one built for that.
Pay once. Keep it.
No subscription, no seats, no recurring charge. One license, yours.
- Everything works, nothing locked
- No account, no card, no email
- It just runs
Nothing to cancel. Buy a key only if it sticks.
- Apple Silicon (M1 or newer)
- macOS 15 Sequoia or later
- 500 MB download, speech models included
- ✓One-time purchase, yours to keep
- ✓Dictation, file transcription, and the MCP server, all included
- ✓Free updates across this version
- ✓14 days to get your money back, no questions
- ✓Works offline, no account
14-day money-back, no questions asked
A few common questions.
Is it actually private?
Yes. The speech models ship inside the app, so dictation and file transcription run entirely on your Mac from the first launch, offline included. Your audio is never uploaded and your transcript history stays local; the text you dictate goes only where you paste it. The one network call Edict can make is its update check, and it asks you first.
Can it transcribe a recording, not just live dictation?
Yes. Drop an audio or video file onto Edict and it transcribes on your Mac, labeling who said what. Transcription takes about 24 seconds for a 60-minute video. Export it as SRT or VTT subtitles, or plain text. The file never leaves the machine.
What’s the MCP thing?
Edict runs a small local server your coding agent can connect to. Through it, Claude Code or Cursor can read your transcript history, pull your last transcript, edit your dictionary, and hand it a recording to transcribe, all on your Mac. Same idea as the rest of the app: your voice and your agent, on your machine, nothing in the cloud.
How fast is it?
The model loads when the app does and stays warm, so dictation starts the instant you press the key. Transcription runs on the Neural Engine at roughly 150x real time: a spoken prompt lands about as soon as you stop talking, and a 60-minute file takes about 24 seconds.
How accurate is it?
Dictation and transcription run NVIDIA’s Parakeet v3, from the model family that has topped Hugging Face’s open ASR leaderboard, and your dictionary catches the words only your codebase uses. Speaker detection uses pyannote’s community model. Because Edict never rewords anything, what you said is what lands. Accuracy varies by language: on NVIDIA’s published Fleurs results the model runs from 3.00% word error rate in Italian to 24.03% in Slovenian, averaging 11.97% across the 25.
Can I dictate in a language other than English?
Yes. Edict dictates in 25 European languages, including French, German, Spanish, Italian, Portuguese, Dutch, Polish, and Swedish. There is no setting to change: NVIDIA’s model card says it “automatically detects the language of the audio and transcribes it without requiring additional prompting”, and that is what happens. Two honest caveats. Edict’s own interface is English only. And accuracy is uneven across the 25, strongest in Italian, Spanish, and Portuguese, weakest in Slovenian and Latvian, so try it in yours during the trial.
Does it clean up or reword what I say?
No, and that’s on purpose. You get your words as you said them, plus your own short list of spelling fixes. The agent you’re talking to already cleans things up, a cleanup model would cost you a beat of latency and would eventually correct an identifier you meant.
Why not just use the dictation built into macOS?
macOS dictation is built for prose, not prompts. There’s no word list for your jargon, no guard against typing into the wrong window, and nothing your agent can read back. Edict does one job: get your exact words into the agent you’re driving, as fast as you can say them.
What do I get when I buy it?
A license for Edict on your Apple Silicon Macs, free updates across this version, and 14 days to get a refund if it’s not for you. No subscription, no account.
What do I need to run it?
Apple Silicon (M1 or newer) and macOS 15 Sequoia or later. The model runs on the Neural Engine, so it needs an Apple Silicon Mac.
Still wondering something? Ask a real person.
Your fastest input device
is your voice.
Tap a key, say what you want, and get back to reading the diff. It runs on your Mac, types where you're working, and your voice never leaves the machine.
Apple Silicon (M1 or newer) · macOS 15 Sequoia or later · 500 MB download · one-time purchase