Deploy VibeCody on macOS
Works on MacBook (Air/Pro) and Mac Mini/Studio. Apple Silicon gets Metal GPU acceleration for fast local inference.
| Setup time: 2 minutes | Cost: Free | GPU: Metal (Apple Silicon) |
Quick Start
curl -fsSL https://raw.githubusercontent.com/TuringWorks/vibecody/main/install.sh | sh
vibecli
Or use the guided setup:
cd vibecody/deploy/macos
./setup.sh --always-on
Step-by-Step
1. Install VibeCLI
curl -fsSL https://raw.githubusercontent.com/TuringWorks/vibecody/main/install.sh | sh
2. Install Ollama (Optional)
brew install ollama
ollama serve &
ollama pull codellama:7b
Apple Silicon Macs run Ollama with Metal GPU acceleration — a MacBook Pro M3 can run 13B models at ~30 tokens/sec.
3. Start Chatting
vibecli
4. Enable Always-On (Optional)
Install VibeCody as a launchd service that starts at login:
./setup.sh --always-on
# Or manually:
vibecli --service install
vibecli --service start
Access at: http://localhost:7878
Mac Mini as a Server
A Mac Mini with Apple Silicon is the best desktop “server” — silent, efficient, and Metal GPU runs 13B models at full speed.
# Headless setup with remote access
./setup.sh --always-on
vibecli --serve --tailscale # Expose via Tailscale
Model Recommendations
| Mac | RAM | Recommended Model | Performance |
|---|---|---|---|
| MacBook Air M1 | 8 GB | codellama:7b | ~15 tok/s |
| MacBook Pro M3 | 18 GB | codellama:13b | ~30 tok/s |
| Mac Mini M2 Pro | 32 GB | codellama:34b | ~20 tok/s |
| Mac Studio M2 Ultra | 64 GB+ | llama3:70b | ~15 tok/s |
Uninstall
./uninstall.sh
# Or: vibecli --service stop && rm ~/.local/bin/vibecli
Troubleshooting
| Problem | Solution |
|---|---|
| “command not found” | Add ~/.local/bin to PATH: export PATH="$HOME/.local/bin:$PATH" |
| Ollama slow | Ensure Ollama is using Metal — check with ollama ps |
| Service won’t start | Check logs: cat /tmp/vibecody-stderr.log |
| Intel Mac | Works fine but no GPU acceleration — use cloud providers for large models |