Overview & quick start

Lavox is a local-first meeting recorder and dictation tool for macOS that turns everything you say into a memory any AI can read. This page gets you from a fresh clone to a working memory.

The pieces

Part	What it is
Lavox Hub (`app/`)	The Mac app: the bar under your notch, meeting recording (mic + system audio, no bot), push-to-talk dictation, on-device Whisper.
Server (`server/`)	A local FastAPI process: meeting transcription, speaker identification, the memory store and its MCP server. Runs on `127.0.0.1:8040`.
Memory	One SQLite file in `~/Lavox/memory/`. Verbatim transcript chunks + a typed assertion index, searched together. See The memory.

Requirements

macOS 13+ on Apple Silicon
Rust and pnpm (for the app), Python 3.12 (for the server)
Xcode with the macOS 15 SDK (the system-audio helper uses the newer ScreenCaptureKit recording API)

Quick start

git clone https://github.com/lavox-app/lavox.git && cd lavox

# 1 · The server: transcription, memory, MCP
cd server
python3 -m venv .venv && .venv/bin/pip install -r requirements.txt
.venv/bin/uvicorn app:app --host 127.0.0.1 --port 8040   # keep running

# 2 · The app: recorder and the bar
cd ../app
pnpm install
./build-and-install.command   # builds, signs (ad-hoc by default), installs

# 3 · Wire the memory into Claude Code
claude mcp add lavox-memory -- "$PWD/../bin/lavox-mcp.sh"

Record a meeting or dictate something, then ask Claude to search lavox-memory for it. Decision extraction is optional and needs an LLM key; see Self-hosting.

Nothing leaves your machine by default. Transcription, embeddings and the memory all run locally; the only network calls are ones you configure yourself. The privacy policy says the same thing in GDPR language.