Universal voice-to-anything hotkey: capture once, paste anywhere, and keep every recording safe. VoxCompose adapts to your speech over time, strips filler, and optionally adds local LLM polish—perfect for AI prompt workflows and cross-app dictation.

brew tap cliffmin/tap
brew install voxcompose ollama

ollama serve &
ollama pull llama3.1

echo "i want to pushto github and committhis code" | voxcompose
# → "I want to push to GitHub and commit this code"

• Java 21
• macOS 11+ (primary target; Ollama required for LLM refinement)
• Optional (golden/accuracy suite): ffmpeg/ffprobe, whisper-cpp (whisper-cli), jq, bc

• Universal hotkey → any app: speak once, paste into ChatGPT/Claude/Cursor/email/docs.
• Never lose work: every capture is saved locally; failures don’t cost you recordings.
• Self-learning & adaptive: corrections improve with your speech patterns; fixes concatenations/tech terms.
• Duration-aware: <21s uses fast corrections-only; longer adds LLM refinement for clarity.
• Fast & local: ~140ms short-path; 100% on-device with Ollama (no API keys); privacy-first unless you point at a remote AI_AGENT_URL/OLLAMA_HOST.

Recent improvements (0.4.4):

• Short clips skip LLM (~90% faster for <21s)
• Upfront fixes for concatenations/tech terms (pushto, committhis, github/json, etc.)
• Optional caching for repeated prompts

Automatic corrections (examples):

• pushto → push to, committhis → commit this
• github → GitHub, json → JSON, nodejs → Node.js

echo "i want to pushto github" | voxcompose --duration 10

Env vars: AI_AGENT_MODEL, VOX_REFINE=0 (disable LLM), VOX_CACHE_ENABLED=1, OLLAMA_HOST (override endpoint).

Full guide: docs/voxcore-integration.md

brew install cliffmin/tap/voxcore
vim ~/.hammerspoon/ptt_config.lua
# LLM_REFINER = { ENABLED = true, CMD = { "voxcompose", "--duration" }, ... }

brew tap cliffmin/tap
brew install voxcompose ollama

Build from source:

brew install openjdk@21 ollama
git clone https://github.com/cliffmin/voxcompose.git
cd voxcompose && ./gradlew --no-daemon clean fatJar

• Unit tests: ./gradlew test
• Integration: ./tests/run_tests.sh
• Golden accuracy/perf (local-only, needs ffmpeg + whisper-cpp + Ollama running): tests/run_golden.sh (writes to tests/results/, gitignored)

brew update && brew upgrade voxcompose
voxcompose --version

Data is preserved (learned profile at ~/.config/voxcompose/learned_profile.json).