andbo-sandboxlisted
Install: claude install-skill qosikz/andbo
# Andbo Sandbox
Run anything risky inside an isolated, policy-controlled container instead of the host.
## When to use
- Executing risky or untrusted commands (installers, curl-pipe-sh, unknown scripts).
- Validating generated code, new tools, skills, or subagents before trusting them.
- Testing changes without touching the real workspace.
- Vetting an MCP server before connecting it.
- Auditing past sandboxed runs.
## Core commands
```
andbo exec "<command>" --json # sandboxed command; exit code passes through; changed files + diff captured
andbo exec --dry-run "<command>" # preview the sandbox (mounts, policy, image) without executing
andbo run "<task>" --agent <name> # drive a coding agent in the sandbox: custom, claude, codex, gemini, goose, opencode
andbo mcp scan <path> --json # static-scan an MCP server; exit 2 = unsafe
andbo session list # audit recorded runs
andbo session show latest # inspect the most recent run; artifacts in .andbo/sessions/<id>/
```
Prefer `--json` when you will parse the result.
## Reading results
`--json` fields: `exit_code`, `stdout`, `stderr`, `changed_files`, `session_dir`. A non-zero `exec` exit code means the command failed inside the sandbox — report it, do not retry on the host.
## Safety rules
- Runs are policy-governed by `andbo.yaml`: network denied by default, secrets and `.env` excluded.
- NEVER pass `--unsafe`, `--yes-unsafe`, `--runtime local`, `--allow-host-home`, or `--