Claude Code Local

Run Claude Code with local LLMs on Apple Silicon using Ollama and LiteLLM.

Overview

This project provides configuration files to run Claude Code CLI against local Ollama models via a LiteLLM proxy. All inference runs locally on your Mac using Metal GPU acceleration.

Key Benefits:

• Fully offline operation
• No API costs
• Data stays on your machine
• Metal GPU acceleration on Apple Silicon

Architecture

┌─────────────────────────────────────────────────────────┐
│                   Host (Apple Silicon Mac)              │
├─────────────────────────────────────────────────────────┤
│  Ollama (native)                                        │
│  └── Port 11434 ← Metal GPU acceleration                │
│                                                         │
│  Docker Network (claude-local-network)                  │
│  ├── LiteLLM Proxy :4000 ← API translation              │
│  └── DevContainer        ← Claude Code CLI              │
└─────────────────────────────────────────────────────────┘

Prerequisites

Quick Start

# 1. Pull a model
ollama pull qwen2.5-coder:7b

# 2. Start Ollama with Docker-accessible binding
./start-ollama.sh

# 3. Start LiteLLM proxy
docker compose up -d

# 4. Open in VS Code DevContainer, then run:
claude

Model Mapping

Claude Code model requests are translated to local Ollama models:

Tip: On 16GB Macs, edit litellm_config.yaml to map all models to the 7b variant.

Setup Guide

1. Pull Ollama Models

For full model support:

ollama pull qwen2.5-coder:14b
ollama pull qwen2.5-coder:7b

For limited RAM (7b only):

ollama pull qwen2.5-coder:7b

2. Start Ollama

Ollama must bind to 0.0.0.0 for Docker containers to connect.

# Stop existing Ollama service if running
pkill ollama

# Start with correct binding
./start-ollama.sh

# Or manually:
# export OLLAMA_HOST=0.0.0.0:11434
# ollama serve

Verify:

curl http://localhost:11434/api/tags

3. Start LiteLLM Proxy

docker compose up -d

Verify:

curl -H "Authorization: Bearer sk-local-dev-key" http://localhost:4000/health

4. Pre-warm Models (Optional)

Avoid cold-start timeout on first request:

./prewarm-models.sh

5. Open DevContainer

Note: Complete steps 2-3 first. The DevContainer requires the claude-local-network created by docker compose up.

1. Open this folder in VS Code
2. Press Cmd+Shift+P
3. Select "Dev Containers: Reopen in Container"
4. Run claude in the terminal

Project Structure

.
├── .devcontainer/
│   └── devcontainer.json    # DevContainer with Claude Code pre-installed
├── docker-compose.yaml      # LiteLLM container configuration
├── litellm_config.yaml      # Model aliases and routing
├── start-ollama.sh          # Start Ollama with correct binding
└── prewarm-models.sh        # Load models into RAM

Configuration

Environment Variables

The DevContainer automatically sets these for Claude Code:

Customization

Edit litellm_config.yaml to:

• Change model mappings (use different Ollama models)
• Adjust request timeouts
• Add cloud API fallback routing

Troubleshooting

License

MIT