ai

Run Claude Code or Cursor Agent inside a sandboxed Docker container with your own shared commands, skills, and rules mounted in.

Why

Keep AI coding assistants isolated from the host while reusing a single set of prompts, slash commands, and configuration across both tools.

Requirements

• Docker

Setup

git clone https://github.com/michalkowol/ai.git
cd ai
./ai

You can clone the repo anywhere — the launcher resolves its own location. For convenience, symlink it onto your PATH:

ln -s "$(pwd)/ai" ~/.local/bin/ai

Usage

ai                   # Claude Code (default), current directory
ai --claude path/    # Claude Code in a specific directory
ai --cursor path/    # Cursor Agent
ai --bash path/      # Plain bash shell in the container
ai --java 21         # Build the Java image with a specific version (default: 25)
ai --node jod        # Build the Node.js image with a specific tag
ai --java 21 path/   # Combine with any tool/path

The given path is mounted as /workspace inside the container. Git worktrees are detected and their common git dir is mounted automatically.

The --java flag selects the eclipse-temurin:<version> base image and tags the built image as ai:java<version>. The --node flag selects the node:<tag> base image (e.g. jod for Node.js 22 LTS) and tags it as ai:node<tag>. The two flags are mutually exclusive; each variant is cached independently.

Layout

.
├── Dockerfile         # Configurable base (via --java / --node) + Claude + Cursor + Docker CLI
├── ai                 # Launcher script
├── claude/            # Per-tool config mounted into the container
├── cursor/
├── common/
│   ├── commands/      # Shared slash commands
│   ├── skills/        # Shared skills
│   └── rules/         # Shared coding rules
└── config/

License

MIT