Build from source

Requirements: an Apple Silicon Mac with 32 GB+ of memory, macOS 15+ (developed on macOS 26), Homebrew, Xcode command line tools, and about 40 GB of free disk.

git clone https://github.com/gorunrunai/local-ai.git
cd local-ai
make setup      # Homebrew deps, Python env (uv), llama-swap, models, web search, test fixtures, web app
make backend    # API + web app + models on http://127.0.0.1:8000

For frontend work with hot reload, use make dev and open http://127.0.0.1:5173.

Useful targets

Command	What it does
make desktop-install	Build the Swift Mac app and install it with its LaunchAgent
make models-video	Optional video generation engines and weights (~52 GB)
make menubar-install	Menu-bar helper with global screenshot hotkeys
make remote / make remote-off	Phone access over Tailscale (same as the switch in Settings → Phone access)
make test	Python and frontend unit tests (no models needed)
make test-integration	Tests against the real local models
make test-e2e	Playwright browser tests, including accessibility audits
make bench / make bench-voice	Speed, memory and voice-latency benchmarks

Repository layout

inference/      model manager, llama-swap config, providers, speech/embedding services, video
media/          file-type detection and processors for images, OCR, audio, video, documents
orchestrator/   FastAPI app: prompt pipeline, agent loop, tools, policy, MCP, RAG, memory, voice
frontend/       React web app, artifact runtime, Capacitor iOS/Android projects, Playwright tests
desktop/        Swift Mac app (WKWebView) that starts and stops the backend
menubar/        Swift menu-bar helper (screenshot hotkeys)
videogen/       separate environment for the video engines (LTX-2.3, Wan 2.2)
config/         models.yaml, tools.yaml, mcp.json, searxng.yml, prompts/, styles/
scripts/        setup, downloads, benchmarks, backend launcher
tests/          unit, integration and fixtures
install.sh      the one-line installer