Get started
Install from source
Build and run Ghira on a Mac from source
Ghira is pre-release. There is no signed download yet, so you build the Mac app yourself. This page takes you from a clean Mac to a running app.
What you need
- A Mac with Apple Silicon, running macOS 14.2 or later.
- Xcode, with its command line tools.
- Rust, installed with rustup. The repository pins the version it needs, and rustup installs it for you.
- Node.js 22 or later, with pnpm (
corepack enableturns it on). - CMake 3.26 or later.
- About 4 GB of free disk space for the models, plus room for the build.
Build and run
git clone https://github.com/nhtera/Ghira.git && cd Ghira
git submodule update --init # not --recursive
pnpm install && pnpm build
./tools/scripts/build-nemo.sh # speech engines (Metal)
pnpm --filter @ghi/desktop tauri dev --features nemoHere is what the commands do:
- Clone the repository. The speech engine lives in a git submodule. Run
git submodule update --initwithout--recursive: the engine nests large repositories Ghira does not use. - Build the interface.
pnpm install && pnpm buildbuilds the web interface the desktop app shows. - Build the speech engines.
./tools/scripts/build-nemo.shbuilds NeMo-Speech.cpp with Metal. It takes a while the first time. - Start the app.
pnpm --filter @ghi/desktop tauri dev --features nemobuilds and opens Ghira.
First launch
Ghira walks you through setup. The step that matters here is Speech models: the app downloads the models for your Mac's preset (about 3.4 GB on Light, about 4 GB on Balanced and Max) from Hugging Face, and checks each file before it uses it. You can keep going while they download. See Speech models for what is downloaded and how to install the files without a network.
Record your first meeting covers the rest of the setup.
Fetch models for development
If you work on the code, you can fetch the same pinned models into ./models
without the app:
./tools/scripts/fetch-models.sh # all the standard models
./tools/scripts/fetch-models.sh nemotron-3.5-asr # or name the ones you wantThe model ids are in crates/ghi-models/registry.toml. Models marked optional,
such as the Whisper models, are skipped unless you name them.
A development build is not a release build
tauri dev makes a debug build. A debug build keeps the encryption key for your
meetings in a file next to the data folder, not in the Keychain. Do not keep
real meetings in a debug build. To use the Keychain in a debug build, set
GHI_KEYSTORE=keychain before you start it.
To build an app bundle and a disk image for testing, run
tools/release/build-dmg.sh --adhoc. The result is signed ad hoc, so macOS
treats it as an unknown developer's app.
Troubleshooting
- A
cargobuild of the desktop app fails. Runpnpm buildfirst. The desktop crate needs the built interface. - A submodule folder is empty. Run
git submodule update --initfrom the repository root. build-nemo.shfails. Check that CMake is 3.26 or later (cmake --version) and that Xcode's command line tools are installed (xcode-select --install).- The app opens but there is no live transcript. The speech models are not
installed yet, or you started the app without
--features nemo. Ghira still records, and it processes the recording once the models are installed. - The models do not download. Check that Strict offline is off in Settings → Privacy. The download resumes where it stopped.
- macOS asks for the microphone or system audio. Allow both. The microphone hears you and the room. Screen & System Audio Recording hears the other people on a call. Without it, Ghira records your microphone only.
Limits
- Only macOS on Apple Silicon is supported. Windows is not shipped, and there is no Android app.
- There is no automatic update in a build from source. To update, pull the repository and build again.
Edit this page on GitHubLast updated