Features

Everything it does, and nothing it sends.

Wrenscribe is one shortcut, two local models and a lot of small decisions about what should happen between the moment you start talking and the moment the text appears.

Dictation

The part you touch fifty times a day, so it has to be quick and get out of the way.

  • Hold to talk, or tap to toggle

    One shortcut does both. Hold it while you speak, or tap it once to go hands-free and tap again to finish. Escape cancels.

  • Transcribed while you talk

    Parakeet works through your recording at each pause, so a long dictation is mostly done by the time you let go.

  • The pill

    A small indicator at the bottom of the screen shows live microphone levels, a countdown near the time limit, and a clear message if a paste does not go through.

  • Sounds you can hear across the room

    Short cues for start, stop, the warning before the time limit and errors, so you know what happened without looking.

  • Five-minute recordings

    The default limit, adjustable. Thirty seconds before the end you get a warning sound, and reaching the limit transcribes normally.

  • Quiet Discord while you dictate

    Optional: Wrenscribe mutes Discord's microphone while you record, so your friends do not hear you drafting an email.

Cleanup

A small language model on your computer edits the transcript before it is pasted. It edits; it does not rewrite.

  • Fillers out

    Um, uh, and "like" or "you know" when they are filler: gone, along with accidentally repeated words.

  • Corrections applied

    Say "Tuesday, no, Wednesday" and only Wednesday makes it to the page. Same for "scratch that".

  • Lists that are lists

    "First this, second that" comes out as a numbered list where the app takes one.

  • Spoken formatting

    Say "new line" or "new paragraph" and get one. Numbers come out the way you would type them (6 PM, $10, 20%), and "slash" and "dot" become characters inside paths.

  • Room for long dictations

    The time the cleanup is allowed grows with the length of what you said, so a long ramble is never cut off half-edited.

Context

The same words read differently in a terminal, a chat and a group message.

  • Styles by app

    Coding agents get plain prompts with paths kept exact. Shells get the literal command. Chat apps get lowercase. Everything else gets normal sentences.

  • Claude Code, recognised

    Wrenscribe tells Claude Code apart from the rest of the terminal by its window title, so prompts and shell commands are formatted differently.

  • Profiles

    Give any app its own style, built-in or written by you. Email in your mail app, terse prompts in your editor.

  • Pasting that works in terminals

    Ctrl+Shift+V where a terminal needs it, Ctrl+V everywhere else, and your clipboard is restored afterwards.

Your words

Names and jargon are where speech recognition usually falls down.

  • Dictionary

    Words spelled your way, with the ways they get misheard. Sound-alike matching catches the near misses you did not list.

  • Built-in vocabulary

    Product and tech names that measurably come out wrong, corrected out of the box. Every entry was tested against recordings before it went in.

  • Snippets

    A spoken trigger that pastes saved text: a template, an address, the prompt you keep retyping.

  • Command mode

    Select text, hold the command shortcut, say what to do with it. The local model rewrites the selection in place.

  • Learns from your correctionsPlanned

    Suggest dictionary words from the fixes you make by hand.

History

Every dictation is kept on disk, so nothing you said is lost to a failed paste.

  • Raw and final, side by side

    History keeps what the speech model heard and what the cleanup produced, with the app and style it used.

  • Copy, paste again, re-format

    Send an old dictation somewhere new, or run it through a different style.

  • Report a problem

    Mark a dictation that came out wrong and record what you expected. It is saved with the dictation, ready to replay.

  • Replay

    Run any saved dictation through the pipeline again from the command line, with or without the cleanup model.

Waitlist

Try it on your own voice.

Early access opens on Linux first. Leave your email and you will hear from us once, when there is a build to install.

Early access opens on Linux first. No spam, unsubscribe any time.