Your loyal companion for safe shell command generation. Published on crates.io with core features working - advanced features in active development.
Get started with Caro based on your workflow and goals
Get started with the core CLI tool - install via cargo and start generating safe commands
Early access to MLX-optimized inference for blazing-fast local command generation
Install Caro as a Claude Skill, or use the upcoming MCP server for automatic delegation
Connect to remote inference servers or configure custom backends
Real-world examples of safe, POSIX-compliant commands generated by Caro
find . -type f -size +1G -exec ls -lh {} \; ✓ Safe find . -name '*.png' -type f -mtime -7 ✓ Safe du -h --max-depth=1 | sort -hr ✓ Safe ps aux --sort=-%mem | head -n 11 ✓ Safe top -b -n 1 | head -n 20 ✓ Safe lsof -i :8080 ✓ Safe git log --since='1 week ago' --oneline ✓ Safe git for-each-ref --sort=-committerdate refs/heads/ --format='%(committerdate:short) %(refname:short)' ✓ Safe git log --follow -p -- <filename> ✓ Safe grep -i 'error' /var/log/app.log | tail -n 50 ✓ Safe awk '{print $9}' /var/log/nginx/access.log | sort | uniq -c | sort -rn ✓ Safe journalctl -p err -n 100 ✓ Safe ping -c 4 example.com ✓ Safe ss -tlnp ✓ Safe wget -c https://example.com/file.tar.gz ✓ Safe grep -Eo '\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Z|a-z]{2,}\b' file.txt | sort -u ✓ Safe wc $(find . -name '*.txt') ✓ Safe sed -i 's/old_text/new_text/g' *.txt ⚠ Moderate Choose how you want to work with Caro - standalone, with Claude, or both
Use Caro directly from your terminal for instant command generation
# Generate a command
caro "find all Python files modified today"
# Review the command
find . -name "*.py" -type f -mtime 0
# Execute (after confirmation)
[y/N]? yIntegrate Caro as an MCP server for seamless Claude Desktop integration
// claude_desktop_config.json
{
"mcpServers": {
"caro": {
"command": "caro",
"args": ["--mcp"],
"env": {
"CARO_BACKEND": "mlx"
}
}
}
}Use Caro as a specialized skill within Claude's broader workflow
# Install Caro as a Claude skill
/plugin install wildcard/caro
# Use the skill
You: /caro-shell-helper "find large log files and compress them"
Caro: I'll generate a safe command for that:
find /var/log -name "*.log" -size +100M \
-exec gzip {} \;
Risk Level: Moderate
- Modifies files (compression)
- Targets system logs
Proceed? [y/N]Caro's multi-layered safety validation protects you from dangerous commands
LLM generates POSIX-compliant command from natural language
Check against 52 dangerous patterns and destructive operations
Evaluate potential impact and assign risk level (Safe → Critical)
Display command with risk level, require explicit confirmation
Execute command only after user approval, log for audit trail
Built-in pattern matching helps identify potentially destructive operations
rm -rf / mkfs.* :(){ :|:& };: chmod 777 / dd if=/dev/zero Every command is evaluated and color-coded based on potential impact
Read-only operations with no system impact
ls -lagrep pattern file.txtfind . -name '*.log' File modifications or network operations
sed -i 's/old/new/g' file.txtwget https://example.com/filecurl -X POST api.example.com System-level changes requiring caution
chmod +x script.shchown user:group filekill -9 <pid> Dangerous operations that are blocked by default
rm -rf /mkfs /dev/sdafork bombs Ensures commands work reliably across different systems and shells
Flexible inference options from ultra-fast local to scalable cloud deployments
Apple Silicon Native
High-performance local inference using llama.cpp with Metal GPU acceleration on Apple Silicon
Distributed Inference
Connect to Exo distributed clusters for running large models across multiple devices
High-Throughput Server
Connect to vLLM servers for production-grade inference with optimized performance
Local Model Management
Connect to Ollama for local LLM inference with easy model management
Desktop LLM Runner
Connect to LM Studio's local server for a GUI-based local model experience
Built-in Fallback
Cross-platform CPU-based inference using Candle (currently in development)
On Apple Silicon? Use llama.cpp with Metal for fast, private, offline inference.
Have multiple Macs? Set up an Exo cluster to run larger models across devices.
Running a server? vLLM offers high-throughput production inference (testing welcome!).
Prefer a GUI? LM Studio and Ollama support are planned for easier model management.
Get Caro running on your preferred platform with step-by-step instructions
macOS 12+ (Monterey, Ventura, Sonoma)
bash <(curl --proto '=https' --tlsv1.2 -sSfL https://setup.caro.sh) cargo install caro # Add alias to ~/.zshrc or ~/.bashrc: alias ai='caro' git clone https://github.com/wildcard/caro.git cd caro . "$HOME/.cargo/env" cargo build --release Ubuntu 20.04+, Debian 11+, Fedora 35+, Arch, RHEL 8+
bash <(curl --proto '=https' --tlsv1.2 -sSfL https://setup.caro.sh) cargo install caro alias ai='caro' # Add to ~/.bashrc git clone https://github.com/wildcard/caro.git cd caro cargo build --release Windows 10/11 (WSL2 recommended)
# In PowerShell (Admin): wsl --install # After restart, in WSL Ubuntu: bash <(curl --proto '=https' --tlsv1.2 -sSfL https://setup.caro.sh) cargo install caro alias ai='caro' FreeBSD 13+, OpenBSD 7+, NetBSD 9+
# Install Rust if needed pkg install rust cargo cargo install caro git clone https://github.com/wildcard/caro.git cd caro cargo build --release Once installed, Caro's configuration is consistent across all platforms:
caro config set backend [mlx|ollama|vllm|embedded] caro config set safety.level [strict|moderate|permissive] caro config show Caro is open source and built with the community. Contribute, report issues, or join the conversation.