LocalAIbundle

Fully local AI coding assistant for macOS (Apple Silicon). Installs and configures a complete private AI stack for code completion, chat, and codebase-aware answers — with zero cloud dependencies.

Your code never leaves your machine.

What it installs

Component	Purpose
Ollama	Local inference engine (Metal-accelerated)
Qwen2.5-Coder (small)	Fast tab completion (<200ms)
Qwen2.5-Coder (large)	Chat, explanations, refactoring
nomic-embed-text	Codebase indexing for context-aware answers
VS Code	Editor
Continue.dev	IDE integration (connects everything)

Quick start

Prerequisites

• macOS on Apple Silicon (M1/M2/M3/M4)
• 16GB RAM minimum (32GB recommended)
• ~20GB free disk space
• Internet connection (for initial model downloads only)

Install

git clone https://github.com/DevenDucommun/LocalAIbundle.git
cd LocalAIbundle
chmod +x install.sh

# Preview what will be installed (no changes made)
./install.sh --dry-run

# Run the full installation (~10-15 min, mostly model downloads)
./install.sh

Verify

# Check all components are running
./install.sh status

# Run inference speed tests
./install.sh test

Start coding

1. Open VS Code: code /path/to/your/project
2. Wait ~10 seconds for Continue.dev to initialize and index your project
3. Start typing — tab completion suggestions appear automatically
4. Press Cmd+L to open the AI chat panel
5. Type @codebase in chat to give the AI full project context

Model tiers

The installer auto-detects your RAM and selects the best models:

RAM	Completion	Chat	Quality
16GB	qwen2.5-coder:1.5b	qwen2.5-coder:7b	Good
24-32GB	qwen2.5-coder:3b	qwen2.5-coder:14b	Great
48GB+	qwen2.5-coder:7b	qwen2.5-coder:32b	Near GPT-4o

Usage

Action	Shortcut	Description
Tab complete	Just type	Suggestions appear inline as you code
Chat	Cmd+L	Ask questions, get explanations
Inline edit	Cmd+I	AI rewrites selected code
Codebase context	@codebase	Search entire project for relevant code
Open file context	@open	Reference currently open files
Terminal context	@terminal	Reference recent terminal output
Diff context	@diff	Reference uncommitted changes

Commands

./install.sh install    # Install and configure everything (default)
./install.sh status     # Check what's running and installed
./install.sh test       # Smoke test inference speed and model loading
./install.sh uninstall  # Remove all components cleanly
./install.sh --dry-run  # Preview install without making changes
./install.sh --help     # Show all commands

How it works

┌─────────────────────────────────────────────────────┐
│  VS Code + Continue.dev Extension                    │
│  (tab completion, chat panel, inline edit)           │
└────────────────────┬────────────────────────────────┘
                     │ OpenAI-compatible API
                     ▼
┌─────────────────────────────────────────────────────┐
│  Ollama (localhost:11434)                            │
│  ├─ qwen2.5-coder:3b   → autocomplete (fast)       │
│  ├─ qwen2.5-coder:14b  → chat/edit (quality)       │
│  └─ nomic-embed-text   → codebase indexing          │
│                                                     │
│  Metal GPU acceleration · Auto-start on login       │
└─────────────────────────────────────────────────────┘

All requests stay on localhost. No API keys needed. No cloud accounts.

After installation

Ollama starts automatically on login via a LaunchAgent. If you need to manage it manually:

# Stop Ollama
pkill ollama

# Start Ollama
ollama serve

# Update models to latest
ollama pull qwen2.5-coder:14b
ollama pull qwen2.5-coder:3b

# List installed models
ollama list

Troubleshooting

Tab completion not working?

• Check Ollama is running: curl http://localhost:11434
• Restart VS Code (Continue.dev connects on startup)
• Check Continue logs: Cmd+Shift+P → "Continue: View Logs"

Slow responses?

• Run ./install.sh test to check token speed
• Close memory-heavy apps (models need RAM)
• Smaller models = faster: edit ~/.continue/config.yaml to use 1.5b for completion

Models not loading?

• Check disk space: df -h /
• Re-pull: ollama pull qwen2.5-coder:14b

Privacy

• Ollama: no telemetry, no network calls, fully air-gappable
• Continue.dev: telemetry disabled in VS Code settings
• All inference runs on-device via Apple Metal GPU
• Models stored locally in ~/.ollama/
• Safe for proprietary/confidential codebases
• Can operate completely offline after initial install

Uninstall

./install.sh uninstall

Removes: Ollama app, all models (~11GB), LaunchAgent, Continue.dev extension and config. Does not remove VS Code itself.