Horseflow is a self-hosted push-to-talk dictation tool that records speech on a global hotkey, transcribes locally with Whisper, cleans the text with a local LLM via Ollama, and pastes the result into the focused application.
0
No data
0
0
MIT
2026-07-24
It offers fully local speech transcription and cleanup with no cloud dependency, giving users complete privacy and offline capability.
Horseflow is self-hosted push-to-talk dictation. It records while a global hotkey is held, transcribes locally with Whisper, cleans punctuation and false starts with a local LLM, then pastes the result into the focused application.
faster-whisperkeyboard hotkey -> local recording -> Horseflow API
|-> Whisper large-v3
`-> Ollama cleanup model
-> clipboard -> synthetic paste
The server and clients may run on the same machine or communicate over a private network such as Tailscale.
cd deploy
cp .env.example .env
nvidia-smi --query-gpu=uuid,name --format=csv
Edit .env. Set:
HORSEFLOW_GPU to the selected GPU UUIDHORSEFLOW_BIND_ADDRESS to 127.0.0.1docker compose up -d ollama
docker compose exec ollama ollama pull qwen3:8b
docker compose up -d --build api
curl http://127.0.0.1:8100/health
The API documentation is available at /docs.
Use Tailscale Serve to expose the localhost-bound API as a stable, tailnet-only HTTPS endpoint. Linux and macOS clients can use the same endpoint:
https://horseflow-server.example-tailnet.ts.net/dictate
Follow the complete cross-machine Tailscale setup. Do not bind Horseflow publicly or use Tailscale Funnel.
The Linux client exclusively grabs physical keyboards through evdev, forwards normal input through a virtual keyboard, and swallows Ctrl+Space. It records through PipeWire and pastes through ydotool.
pw-recordwl-copyydotoolnotify-send/dev/input/event* and write /dev/uinputOn many distributions, the user must belong to the input and
Self-hosted push-to-talk dictation with local Whisper and LLM cleanup
Similar projects matched by category, topics, and programming language.
Harness Engineering is a methodology for improving coding agent outputs by carefully crafting the environment around them—providing curated context, tools, and executable constraints that encode an organization’s nonfunctional requirements and cumulative lessons.
A 28.9 million parameter language model runs on an $8 ESP32-S3 microcontroller entirely on-device, generating simple stories at about 9.5 tokens per second.
mimic captures traffic from any iOS or web app and automatically generates a Python client library that lets you call the app's API like a regular library.