Container Sandbox

Container Sandbox — Isolated Execution Environment

Run potentially dangerous operations inside ephemeral Docker containers. The host filesystem is protected — containers get read-only project mounts with a writable overlay.

When to Use

• Destructive Ansible playbooks (--force, --limit testing)
• Untrusted scripts from external sources
• Operations that modify system state (rm -rf, package installs)
• Testing migrations or schema changes before applying to production
• Running code from unknown repos

Quick Start

Basic sandbox (Ubuntu + project files):

docker run --rm -it \
  -v "$(pwd):/project:ro" \
  -v /tmp/sandbox-work:/work \
  -w /work \
  ubuntu:24.04 \
  bash -c "cp -r /project/* /work/ && bash"

Python sandbox (with project deps):

docker run --rm -it \
  -v "$(pwd):/project:ro" \
  -v /tmp/sandbox-work:/work \
  -w /work \
  python:3.12-slim \
  bash -c "cp -r /project/* /work/ && pip install -r requirements.txt 2>/dev/null; bash"

GPU sandbox (for CUDA operations):

docker run --rm -it --gpus all \
  -v "$(pwd):/project:ro" \
  -v /tmp/sandbox-work:/work \
  -w /work \
  nvidia/cuda:13.1-devel-ubuntu24.04 \
  bash

Ansible dry-run sandbox:

docker run --rm -it \
  -v /opt/ai-project:/project:ro \
  -v ~/.ssh:/root/.ssh:ro \
  -w /work \
  --network host \
  willhallonline/ansible:latest \
  bash -c "cp -r /project/* /work/ && ansible-playbook -i inventory playbook.yml --check --diff"

Extracting Results

After running operations in the sandbox:

# Copy results out of the sandbox work directory
docker cp :/work/results/ ./sandbox-results/

# Or use the /tmp/sandbox-work mount (persists after --rm if pre-created)
ls /tmp/sandbox-work/

Safety Rules

1. Always mount project directories as :ro (read-only)
2. Use /tmp/sandbox-work or /work for writable operations
3. Use --rm to auto-cleanup containers after exit
4. Use --network none for maximum isolation (no internet access)
5. Never mount / or /home as writable into a sandbox
6. GPU sandboxes (--gpus all) still share the host GPU — use for testing, not isolation

Integration with Claude Code

When Claude needs to run something risky, instruct it to:

1. Build the Docker command with appropriate mounts
2. Execute the command via Bash tool
3. Capture stdout/stderr from the container
4. Extract any produced files from the work directory
5. Clean up: rm -rf /tmp/sandbox-work

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: scoobydont-666
• Source: scoobydont-666/shared-claude-skills
• License: MIT

Install and usage instructions live in the source repository linked above.