Installation¶
Ripple is distributed as a Homebrew formula built from source. Because it depends on MLX's Metal shaders, the formula drives xcodebuild rather than swift build - that is the only build system that emits default.metallib alongside the binary. A plain brew install from a bottle is not available yet; you install from a local clone.
Requirements¶
| Requirement | Detail |
|---|---|
| macOS | 26+ (Tahoe) |
| Architecture | Apple Silicon (arm64) - required for MLX on-device inference |
| Xcode | 26+ - must be installed and active (xcode-select) |
| Homebrew | Any recent version |
hf CLI |
Optional, for pre-fetching models from Hugging Face |
Intel Macs are not supported
MLX targets Apple Silicon exclusively. Ripple can run remote models on Intel if you build it, but the on-device inference path is arm64-only. The Homebrew formula will refuse to install on x86_64.
Step 1 - Clone the repository¶
Step 2 - Install via Homebrew¶
Install the --HEAD formula from the local clone:
Homebrew calls xcodebuild under the hood. The build typically takes 3-5 minutes the first time while Xcode resolves Swift packages and compiles the Metal shaders. Subsequent installs (after brew reinstall) are faster because the SPM cache is warm.
If the build fails inside Homebrew's sandbox
On some machines Homebrew's build sandbox blocks xcodebuild's network access to the Swift Package Index. If that happens, set HOMEBREW_NO_SANDBOX=1:
This disables Homebrew's process sandbox for this install only; it is safe.
The formula installs two things into Homebrew's libexec directory:
- The
ripplebinary and*.bundleresource files (Metal shaders, model configs) - A launcher shim in
bin/ripplethat sets up the correct resource path before exec-ing the real binary
Both must stay co-located; do not move or symlink just the binary.
Step 3 - Pre-fetch a planner model¶
Ripple does not auto-download a model on first launch - you supply one explicitly. The recommended starting model is LiquidAI's LFM2.5 1.2B instruct in MLX bf16 format:
This saves the weights to ~/.cache/huggingface/hub/, which is where Ripple looks for local models. The download is roughly 1.2 GB.
If you do not have the hf CLI, install it with:
You can also pull models via Ripple itself after installation:
Larger models
The 1.2B model runs comfortably on 16 GB unified memory. If you have 32 GB or more, the 8B variant (LiquidAI/LFM2.5-8B-Instruct-MLX-bf16) handles longer context and more complex reasoning tasks. See Local MLX models for the full catalog and memory guidance.
Step 4 - Apple Notes Automation permission¶
Ripple's Apple Notes tools use the macOS Automation APIs. The first time you run a task that touches Notes, macOS will prompt for permission. If you want to grant it proactively:
- Open System Settings > Privacy & Security > Automation.
- Find ripple (or Terminal, depending on how you launched it).
- Enable the Notes checkbox.
You can skip this step and grant permission on demand when the prompt appears.
Step 5 - Verify the installation¶
You should see the top-level usage showing chat, run, mcp, and model subcommands. To confirm the model loads correctly, run a quick one-shot:
--yes tells Ripple to load the model without prompting. If you see a response, the Metal shaders and model weights are correctly wired up.
Keeping Ripple up to date¶
After pulling new commits:
Troubleshooting¶
xcodebuild: command not found
Xcode 26 must be installed from the App Store or the Apple Developer portal, not just the Command Line Tools package. After installing, run:
ripple crashes immediately after install
The most common cause is the default.metallib not being found. Confirm the shim is in place:
brew install --HEAD ./Formula/ripple.rb.
Model not found
Ripple looks in ~/.cache/huggingface/hub/. Run ripple model list to see what's cached, or ripple model pull <id> to download.
Next steps¶
- Quickstart - your first interactive session
- Local MLX models - browse the model catalog, memory requirements, and variant naming