Localflow

Talk, and get clean text where you type.

A dictation app that runs on your computer. Say it the way it comes out, and Localflow writes what you meant into whatever you were typing in.

DownloadSee how it works

Message/Sam

You say: “um I think we should uh ship it on Monday”. It types: “I think we should ship it on Monday.”

Listening

What happens when you stop talking

Three steps, usually done in about a second. Then it pastes into the window you were typing in, and nothing keeps running.

  1. 01

    It hears you

    The speech engine runs inside the app, on your GPU when there is one. Silence is skipped, and your vocabulary primes the spelling of names.

  2. 02

    It cleans up

    A small language model removes fillers and changes of mind, adds punctuation, and turns spoken counting into lists and long rambles into paragraphs.

  3. 03

    It checks its work

    Automatic checks look for concrete mistakes, like a misheard name or a dropped “maybe”, and repair only those. If cleanup can't run, you get the words as heard.

Made for talking to coding agents

In a terminal running Claude Code, Codex, opencode or Gemini CLI, a few phrases press the keys for you, and files you name become @path references. A command only counts when it's all you said.

“…and send it”
Enter
“stop”
Esc
“undo that”
EscEsc
“clear the chat”
/clear
“compact”
/compact
“the install file”
@install.py

It learns the words you fix

When it gets a name wrong and you correct it, it adds the version you kept to your vocabulary. Changes of mind, like Tuesday to Wednesday, are left alone.

You correct it

Ask Claude to review it.

Added to your vocabulary

words.txt
  • Claudeheard as “Claw”
  • Pine Scriptheard as “Py script”
  • Omarchyheard as “omar key”

Runs on your computer

Your voice never leaves your machine. Cleanup runs locally with Ollama, or on a small hosted model with your own API key.

On this computer

Nothing is sent anywhere. Cleanup runs on qwen3.5:4b through Ollama: about 0.4 seconds a sentence on a GPU.

Disk
3.3 GB for the cleanup model
Works offline
Yes

With your API key

Claude Haiku, or any provider with an OpenAI-compatible API. Each dictation's text is sent to that provider. The audio never is.

Cost
Under $0.0005 a dictation
Works offline
No
Between dictations
Nothing running on Linux. About 6 MB on macOS and Windows, waiting for the hotkey.
9 seconds of speech
Transcribed in 0.11 to 0.21 seconds on a GPU, under a second on a CPU.
Your history
The last 200 dictations, readable only by you. Turn it off in the settings.

Light on memory

The app holds almost nothing between dictations. The speech model loads when you stop talking and is dropped right after.

Memory through one dictation: Waiting, Nothing on Linux. About 6 MB on macOS and Windows, listening for the hotkey. While you talk, ~40 MB for the recording pill and the recorder. Transcribing, ~550 MB 100 MB of RAM and 450 MB of GPU memory, for 0.1 to 0.8 seconds. About 330 MB of RAM on a CPU. After it pastes, A few MB for three minutes, watching for fixes to learn. Then back to waiting.
Waiting

Nothingon Linux. About 6 MB on macOS and Windows, listening for the hotkey.

While you talk

~40 MBfor the recording pill and the recorder.

Transcribing

~550 MB100 MB of RAM and 450 MB of GPU memory, for 0.1 to 0.8 seconds. About 330 MB of RAM on a CPU.

After it pastes

A few MBfor three minutes, watching for fixes to learn. Then back to waiting.

Speech models, 9 s of speech

Measured on Linux, Ryzen 5 5600 and Radeon RX 9070 XT

RAMGPU memory
ModelRAMGPUTime
tiny.en CPU104 MB–0.18 s
base.en CPU164 MB–0.27 s
small.en CPU330 MB–0.78 s
small.en GPU, default~100 MB448 MB0.11–0.21 s
large-v3-turbo GPU118 MB846 MB0.26–0.51 s

The cleanup model

The biggest piece, and it lives in Ollama rather than the app. Ollama unloads it after an hour without a dictation, or two minutes on macOS.

qwen3.5:4b, while loaded
~3.8 GB
On disk
3.3 GB
With your API key
None

Download

Hold the hotkey and talk, or tap it, talk, and tap again. Esc cancels.

macOS

Apple Silicon (M1 or newer) or Intel

Hotkey
⌥⇧Space

Allow Accessibility and Microphone when macOS asks.

Download .dmg14 MB

Windows

x64 installer

Hotkey
CtrlShiftSpace

Starts when you sign in. Change settings from the Start menu.

Download .msi11 MB

Linux

x86-64, glibc 2.39 or newer

Hotkey
You choose

Bind a key to “localflow toggle”. The settings window shows how.

Download .AppImage26 MB

Version 0.1.2. The installers aren't signed yet, so your system asks once before opening them. Prefer to build it yourself? Build from source.

Questions

Something else? The documentation covers every setting.